JP7708355B2 - Machining condition learning device, machining device with learning device, and machining system with learning device - Google Patents

Description

The present invention relates to a learning device for processing conditions, a processing device with a learning device, and a processing system with a learning device, and more specifically to a learning device for joining conditions for two or more parts, a joining device with a learning device, and a joining system with a learning device, a learning device for fusing conditions for one or more parts, a fusing device with a learning device, and a fusing system with a learning device.

Conventionally, methods for joining two or more plastic or metal parts include (1) high-frequency welding, in which two or more parts are welded by dielectric heating by applying a high-frequency current to them, (2) non-contact hot plate joining, in which a hot plate is brought close to the surfaces of two or more parts without contact, (3) vibration welding, in which two or more parts are pressed together and welded by applying low-frequency vibrations of about 100 Hz, (4) laser joining, in which two or more parts are joined by irradiating them with a laser beam, (5) ultrasonic welding, in which two or more plastic parts are welded, (6) ultrasonic metal joining, in which two or more metal sheets are solid-state joined, and (7) ultrasonic joining, in which a metal electrode is joined to the surface of a glass plate.

Methods for cutting one or more parts of plastics, food, rubber sheets, etc. include (a) a laser cutting method in which a laser beam is applied to cut the parts, and (b) an ultrasonic cutting method in which an ultrasonically vibrating cutter is pressed against one or more parts to cut them.

In the above (1) to (7), (a) and (b), vibrational energy or thermal energy is applied to the parts to heat them up, melt them, or form solid-state bonds, resulting in the joining or cutting of the parts to create products. Therefore, by changing the amount of energy applied to the parts and the time for which the energy is supplied, it is possible to find the joining or cutting conditions.

However, the joining and fusing conditions vary depending on the material, shape, and dimensions of the parts to be joined or fusing. Conventionally, when joining or fusing new parts, new joining and fusing conditions that meet the joining and fusing specifications had to be found through experiments each time.

On the other hand, processing equipment such as joining devices and fusing devices must be versatile. These processing devices can process a wide range of objects, and the processing content is wide-ranging and diverse. For this reason, one state is selected from the multiple states of energy that can be supplied, and the energy supply time is determined.

The following describes conventional methods of determining processing conditions such as joining conditions and cutting conditions, using the above-mentioned (1) conventional high-frequency welding method and the high-frequency welding device using this as representative examples. Note that many aspects of the above-mentioned (2) to (7) conventional joining methods and (a) and (b) conventional cutting methods, including the method of determining processing conditions such as the desired joining conditions and cutting conditions by changing the magnitude of the applied energy and the supply time, are the same as or similar to (1), so a description will be omitted.

In high-frequency welding, when a strong high-frequency electric field is applied to plastic materials, films, sheets, etc., the polarity of the electrodes changes continuously at the molecular level, causing collisions, vibrations, and friction at the molecular level inside the material, which then self-heats and fuses and welds the plastic materials, films, sheets, etc.

The most common method of controlling the high-frequency current is to set the time that the high-frequency current is supplied to the workpiece (hereinafter referred to as the "welding time") to a specified value, but if a uniform welding time is used to control the welding time for workpieces with different dimensions and shapes, excessive welding (excessive welding), sparks, or insufficient welding may occur.

Figure 42 shows the relationship between welding time (S) and current (A) when high-frequency welding is controlled only by the conventional welding time to weld plastic materials that have different optimal welding times. In Figure 42, the X-axis shows welding time (S) and the Y-axis shows current (A).

Figure 42 shows an example of three types of plastic materials (material A, material B, and material C) that have different optimal welding times, and shows that when the materials are welded for a uniform welding time, that is, until the welding end time T2, (1) material A will be over-welded, and (2) material C will be under-welded. Note that in Figure 42, material B is shown as an example of normal welding.

In Figure 42, high-frequency current is applied uniformly until the end of welding (T2), but for material A (graph with dashed dotted line in Figure 42), welding of the workpieces is already completed from the start of welding (T0) to the end of welding (T1), which is before the end of welding (T2). From the start of welding (T1) to the end of welding (T2), an unnecessary high-frequency current is applied to the workpieces. As a result, the plastic material (material A) is over-welded. This is not only a waste of energy, but also damages the workpieces.

Conversely, with material C (dotted line graph in Figure 42), welding time should normally be up to T3, but at the end of welding at T2, welding is not yet complete, but the high-frequency current is stopped, resulting in insufficient welding.

Therefore, another conventional control method is known in which the change in the distance between two dies sandwiching the workpieces with different optimal welding times is detected by the current value supplied to the high-frequency welding processing section of the high-frequency welding device, and the amount of high-frequency energy supplied to the high-frequency welding processing section is controlled by detecting the current value (see, for example, Patent Document 1).

In this other conventional control method, welding begins when the workpiece comes into contact with both dies, and as the welding progresses and the distance between the dies becomes shorter, the current value supplied to the high-frequency welding processing unit increases, and the welding completion point is the inflection point where the current value reaches a maximum and then converges. In other words, the welding completion point is when the current value passes the current value at the welding start point (L reference value), reaches a maximum, and then reaches the current value (H reference value) which becomes the inflection point.

Figure 43 shows the relationship between welding time (S) and current (A) when plastic materials, A, B, and C, are welded using another conventional high-frequency welding control method. As with Figure 42, the X-axis shows welding time (S) and the Y-axis shows current (A).

In the example shown in Figure 43, the welding start point for three types of workpieces A, B, and C is the same TS, but the welding completion points are different, TF', TF'', and TF''', respectively, and optimal welding times TA, TB, and TC are assigned to each. As a result, compared to the conventional example shown in Figure 42, (1) over-welding is reduced in the case of workpieces with a short optimal welding time, and (2) under-welding is reduced in the case of workpieces with a long optimal welding time.

In Figure 43, the H reference value, which is the current value at which welding is completed, is measured and checked for workpieces of predetermined dimensions and shapes in advance, and then set in the welding device. During operation, the current transition is detected, and when the current exceeds the H reference value once, reaches a maximum, and then reaches the H reference value again, the high-frequency output is controlled to stop, and the workpieces are welded.

However, the task of measuring and checking the welding conditions for workpieces of specified dimensions and shapes in advance was time-consuming, even for experienced workers, requiring repeated trial and error. Each time a new workpiece was to be welded, the welding conditions had to be measured and checked in advance.

Recently, artificial intelligence technology (hereafter referred to as AI technology) has been attracting attention. AI technology learns from data accumulated in the past. It is expected that the use of AI technology will eliminate the need to find new joining and cutting conditions every time new joining or cutting is performed.

However, in the conventional method of joining two or more parts or the conventional method of fusing one or more parts, no practical and specific AI technology is disclosed or suggested. A learning device that uses AI technology to learn the conditions for arc welding of metals is known, but the document does not disclose the update equation Q(s,a)←Q(s,a)+α(r+γmaxQ(Snext,a)-Q(s,a)) of the action value function Q(a,s) shown in FIG. 44, which is known as Q-learning.
It only conceptually discloses applying the above to one state of arc welding and arc welding, but does not disclose any actual or specific technical details (see Patent Document 2).

As shown in Figure 45, conventional AI technologies include reinforcement learning, Q-learning, statistical learning (neural networks), deep learning, and deep reinforcement learning, which combines deep learning and reinforcement learning, and research is progressing (see publicly known literature 1).

In reinforcement learning, in state S in a certain environment, by trial and error with action a, the system learns which action a will maximize the action value as a reward. A simple example will be explained using FIG. 46. In FIG. 46, when four actions (at1), (at2), (at3), and (at4) are taken in state (St), Q(St, at1), Q(St, at2), Q(St, at3), and Q(St, at4) are obtained as Q values (action values) that evaluate the results of the actions as rewards. The idea behind reinforcement learning is to select the action (a) that will maximize the reward by calculating backwards from the size of the reward. In FIG. 46, Q(St, at3) is the maximum value among the four rewards, so the maximum reward (Q value) can be obtained by using action (at3).

In Q-learning, the action value function Q(a,s) already shown in Figure 44 is used as the update equation, and by continuing online learning to take a new action a in a state S in a certain environment and obtain a reward, the learning experience increases and the action value function Q(a,s) is updated, and an action that can obtain a large reward is quickly reached (see non-patent documents 1 and 2).

In deep learning, a multi-layered neural network like that shown in Figure 47(b) is created by stacking multiple perceptrons like those shown in Figure 47(a). When an action is input to the input layer of the multi-layered neural network, each input is weighted, learning takes place in the hidden layer, and a strategy is output from the output layer that is the maximum reward multiplied by a probability (the probability of being adopted).

In deep reinforcement learning, as shown in Figure 48, when an action is input to the input layer of a multi-layered neural network in a certain state, each input is weighted, and a policy multiplying the action by a probability is output from the output layer through the hidden layer. After that, the probability of actions that lead to high rewards is increased, and learning is repeated in the multi-layered neural network until an action that can earn a large reward is reached. Then, the policy multiplying the action by a probability changes, and when the environment or state changes, these are taken into account in learning.

However, conventional AI technology has not disclosed (1) a practical and specific AI technology using a learning device for processing methods such as a method for joining two or more parts or a method for fusing one or more parts. Also, (2) an AI technology has not been disclosed that checks the state of energy applied to a workpiece when processing begins in a versatile processing device capable of processing in multiple states, and selects specific processing conditions as actions that will provide the maximum reward in that state, i.e., actions that satisfy the processing specifications. (3) An AI technology has not been disclosed that monitors the state of energy supply during processing by a processing device, determines whether the processing state is normal or abnormal using a learning device, and takes appropriate action if an abnormality is detected.

JP 2002-370283 A JP 2017-30014 A

"Learn an Introduction to Artificial Intelligence with Illustrations" by Tadahiro Taniguchi, Kodansha, December 2020 "Reinforcement Learning and Deep Learning: Simulation using C Language" by Tomohiro Kodaka, Ohmsha, October 2017

In conventional processing methods such as joining and fusing, particularly in methods for joining two or more parts or fusing one or more parts, when joining or fusing a new part, new joining conditions or fusing conditions that meet the joining or fusing specifications were found through experiments each time. Since new parts do not have the same material, dimensions, shape, etc. as the parts to be joined or fusing, it was rare to use the joining or fusing conditions found in the past as they were. Each time, workers were required to conduct experiments to find the desired joining or fusing conditions, which was a burden.

The first object of the present invention is to provide a learning device, a processing device with a learning device, and a processing system with a learning device that learn the processing conditions when processing parts. More specifically, the object is to provide a learning device, a joining device with a learning device, and a joining system with a learning device that learn the joining conditions when joining two or more parts, and a learning device, a fusing device with a learning device, and a fusing system with a learning device that learn the fusing conditions for one or more parts.

Conventional general-purpose processing equipment has a wide range of processing objects and processing contents, so the processing conditions change frequently. By connecting a learning device that learns the processing conditions, it is convenient to be guaranteed that processing is always supported under processing conditions that meet the processing specifications.

The second object of the present invention is to provide a learning device, a processing device with a learning device, and a processing system with a learning device that constantly supports processing under processing conditions that satisfy the processing specifications by monitoring the processing specifications even during processing after processing has started under processing conditions that satisfy the processing specifications. For example, even if the processing specifications have been updated from the previous processing operation or have been changed during processing, the object is to perform processing under processing conditions that satisfy the latest and current processing specifications.

Furthermore, even if the materials, shapes, and dimensions of the parts being joined, cut, or otherwise processed are the same, variations within the tolerances will cause subtle changes in the amount of energy supplied to the parts. Therefore, even if processing conditions such as joining and cutting conditions that are thought to be optimal and determined in advance through experiments are used, if work is done under only one processing condition, variations within the tolerances of the parts being processed will cause variations in the welding force and the aesthetics of the cut surface. This poses the issue of inconsistent quality after processing.

In addition, conventional joining devices would join parts that were out of tolerance even when they were mistakenly set as the parts to be welded, which could result in defective products being produced by joining parts that were out of tolerance.

The third objective of the present invention is to determine the state in which energy supply begins even if the parts to be joined, cut, or otherwise processed are roughly the same in material, shape, and dimensions, and even if there is variation within the tolerances, and to read out processing conditions such as joining conditions and cutting conditions from a learning device to determine the energy supply time that satisfies the joining, cutting, or other processing specifications for that state. And, when a part that is out of tolerance is set, joining is not performed, thereby preventing the production of defective products in which parts that are out of tolerance are joined.

Furthermore, it would be convenient if the learning device itself could indicate the need for updated learning when it is unable to find joining conditions, fusing conditions, and other processing conditions that satisfy the processing specifications from the learning device that has learned the processing conditions.

The fourth objective of the present invention is to determine whether processing conditions such as joining conditions and fusing conditions that satisfy the processing specifications can be found from the learned learning device, and to indicate that updated learning is necessary when processing conditions such as joining conditions and fusing conditions that satisfy the processing specifications cannot be found.

The fifth object of the present invention is to provide a processing device with a learning device that, when it is determined that no processing conditions, such as joining conditions or fusing conditions, that satisfy the processing specifications can be found from the learned learning device, updates and learns processing conditions, such as joining conditions or fusing conditions, that satisfy the processing specifications, and can perform processing such as joining two or more parts or fusing one or more parts under the updated and learned processing conditions, such as joining conditions or fusing conditions.

The sixth object of the present invention is to have a versatile joining or fusing device that checks the state of energy applied when starting joining, fusing, or other processing, reads out the joining conditions, fusing conditions, and other processing conditions from a learning device to set the energy supply time that satisfies the joining, fusing, or other processing specifications for that state, and then performs joining, fusing, or other processing.

The seventh objective of the present invention is to monitor the energy state of the processing device during processing, and when the energy state changes to an abnormal state, to indicate this abnormal state and stop processing, thereby ensuring the safety of the processing device.

In order to achieve the above object, the learning device of the present invention has at least (1) a "processing result measurement means" for measuring the processing result, (2) a "processing result evaluation means" for evaluating the processing result as a reward, (3) a "learning storage means" for storing the processing result and the reward obtained by evaluating the processing result, (4) a "processing condition learning means" for organizing the learning results evaluated based on the processing result so that the processing conditions can be read out, (5) a "processing specification monitoring means" for monitoring the processing specifications, and (6) a "processing condition reading means" for reading out the processing conditions from the learning storage means, and the processing result is measured in advance by the processing result measurement means of (1). The processing conditions are measured, the processing result is evaluated as a reward by the processing result evaluation means (2), the processing result and the reward are stored in the learning storage means (3), the learning results evaluated based on the processing result are organized by the processing condition learning means (4) so that the processing conditions can be read out, and when the processing specifications are input to the processing means connected to the learning device, the processing specifications are monitored by the processing specification monitoring means (5), and the processing conditions that satisfy the latest and current processing specifications are read out from the learning storage means by the processing condition reading means (6), and the read out processing conditions are output to the processing means, forming a processing condition learning device.

As a result, when processing specifications are input to the processing specification input means, processing conditions such as joining conditions or fusing conditions that satisfy the processing specifications are read from the learning storage means and output to the processing means. The processing means stores the processing conditions in the "main storage means" and performs processing using the stored processing conditions.

In addition, once processing has begun under conditions that satisfy the processing specifications, the processing specifications are monitored even during processing, providing constant support to ensure that processing is carried out under conditions that satisfy the processing specifications.

In addition, when there are variations in the material, shape, or dimensions of the parts being machined, the system is configured to learn the machining conditions that correspond to those variations and read out the machining conditions that satisfy the machining specifications.

Furthermore, it is configured to determine whether machining conditions that satisfy the machining specifications are found from the learning storage means, and if machining conditions that satisfy the machining specifications are not found, to indicate the need for updated learning.

When it is determined that machining conditions that satisfy the machining specifications cannot be found from the learning storage means, the system is configured to update and learn machining conditions that satisfy the machining specifications, and to read out the updated and learned machining conditions.

Furthermore, a "state monitoring means" is provided that monitors the state of energy applied to the workpiece by the processing device connected to the learning device, and is configured to learn the behavior and reward in the energy state when the processing device starts processing, confirm the state when the processing device starts processing, and read out the processing conditions that satisfy the processing specifications in that state.

The status monitoring means monitors the state of energy applied to the workpiece by the processing device after processing has started, and if the state of the processing device becomes abnormal compared to the state assumed by the learning device, this is indicated and processing by the processing device is stopped.

In addition, a learning device has been constructed in which the processing conditions to be learned are the joining conditions for two or more parts.

The learning device also configures the machining conditions to be learned as the melting conditions for one or more parts.

The processing device with a learning device, which is connected to any one of the above learning devices via a communication line, is provided with at least (i) a "processing specification input/output means" for inputting processing specifications, (ii) a "main memory means", and (iii) a "main control means". The processing conditions read from the learning memory means of the connected learning device are stored in the main memory means, and the main control means is used to perform processing work using the stored processing conditions, thereby constituting the processing device with a learning device.

The processing means constituting the processing system is provided with at least (a) a "processing specification input/output means" for inputting processing specifications, (b) a "main memory means", and (c) a "main control means". The processing conditions read from the learning memory means of the connected learning device are stored in the main memory means, and the main control means is used to perform processing work using the stored processing conditions, forming a processing system with a learning device.

In the present invention, a method and a learning device for learning the processing conditions for joining two or more parts, melting one or more parts, and processing can be performed, such as joining two or more parts or melting one or more parts.

The present invention constantly supports processing under processing conditions that satisfy the processing specifications by monitoring the processing specifications even during processing, from the time processing begins under the processing conditions that satisfy the processing specifications. Therefore, even if the processing specifications have been updated from the previous processing operation or changed during processing, for example, processing can be performed under processing conditions that satisfy the latest and current processing specifications.

The present invention also uses a method and device for learning processing conditions such as joining conditions and fusing conditions that satisfy processing specifications when the parts to be joined are approximately the same in material, shape, and size, but have variations within the tolerances, and can perform processing such as joining two or more parts or fusing one or more parts. Also, when a part outside the tolerances is set as the object to be welded, joining is not performed, thereby preventing the production of defective products in which parts outside the tolerances are joined.

Furthermore, the present invention can determine whether joining conditions or fusing conditions that satisfy the processing specifications can be found from the learned learning device, and if joining conditions or fusing conditions that satisfy the processing specifications cannot be found, it can indicate the need for updated learning.

And when the present invention cannot find processing conditions such as joining conditions or fusing conditions that satisfy the processing specifications from the learned learning device, it can update and learn processing conditions such as joining conditions or fusing conditions that satisfy the processing specifications, and perform processing such as joining two or more parts or fusing one or more parts under the updated and learned processing conditions such as joining conditions or fusing conditions.

The present invention also enables versatile processing equipment such as joining and cutting devices to check the state of energy applied when starting processing such as joining or cutting, and to perform processing operations such as joining and cutting under processing conditions that satisfy the processing specifications in that state.

Furthermore, the present invention monitors the energy state of the processing device during processing, and if the energy state changes to an abnormal state, it notifies the user of this abnormal state and stops processing, thereby ensuring the safety of the processing device.

1 is an external perspective view of a high-frequency welding device with a learning device according to a first embodiment of the present invention; 1 is a schematic configuration diagram of a high-frequency welding device with a learning device according to a first embodiment of the present invention; 4 is a diagram showing the relationship between the transition of the anode current and the welding force of the high-frequency welding device with a learning device according to the first embodiment of the present invention. FIG. FIG. 4 is a diagram showing the relationship between the transition of the anode current and the welding force of the high-frequency welding device with a learning device according to the first embodiment of the present invention. 1A to 1D are diagrams individually showing the rewards (Q(A, TA)) obtained when several actions (TA) are taken in one state (A) in a high-frequency welding device with a learning device according to a first embodiment of the present invention. FIG. 1A is a diagram showing, in the form of a tree, the relationship between states (S, A, D), actions (TA) (TD), and rewards (Q values) (Q(A, TA) (Q(D, TD)) learned by the high-frequency welding apparatus with a learning device according to the first embodiment of the present invention; FIG. 1B is a diagram showing, with arrows, the relationship between states, actions, and rewards learned by the high-frequency welding apparatus with a learning device according to the first embodiment of the present invention. 1A to 1D are perspective views showing an image of the procedure for measuring welding strength using a learning device according to a first embodiment of the present invention. (a) A diagram showing the contents of the data stored in the learning memory means of the first embodiment of the present invention. (b) A diagram organizing the memory contents during learning. (c) A diagram showing the learning results after learning. FIG. 4 is a flowchart showing a learning procedure according to the first embodiment of the present invention. FIG. 2 is a flow diagram showing a welding procedure according to the first embodiment of the present invention. 5A, 5B, and 5C are diagrams showing how welding conditions are identified from learning results stored in a learning storage means by the learning device according to the first embodiment of the present invention. FIG. 13 is a diagram showing a range (TA2 to TA3) of welding times specified as welding conditions when a lower limit welding force and an upper limit welding force are input as welding specifications according to the first embodiment of the present invention. FIG. 4 is a diagram showing the relationship between the transition of the anode current and the welding force of the high-frequency welding device when update learning is performed in the first embodiment of the present invention. FIG. 13 is a diagram showing typical thicknesses when components of a workpiece are stacked according to a second embodiment of the present invention. FIG. 11 is a diagram showing the transition of an anode current in a high-frequency welding device with a learning device according to a second embodiment of the present invention. FIG. 11 is a diagram showing the relationship between the transition of the anode current and the welding force of the high-frequency welding device with a learning device according to the second embodiment of the present invention. FIG. 11 is a diagram showing the relationship between the transition of the anode current and the welding force of the high-frequency welding device with a learning device according to the second embodiment of the present invention. FIG. 13 is a tree diagram showing the relationship between states (S, A, B, C), actions (TA, TB, TC), and rewards (Q(A, TA), (Q(B, TB), (Q(C, TC)) learned by a learning device according to a second embodiment of the present invention. FIG. 11 is a diagram showing the contents of the learning results stored in a learning storage means according to the second embodiment of the present invention. 6A, 6B, and 6C are diagrams showing how welding conditions are identified from learning results stored in a learning storage means by a learning device according to a second embodiment of the present invention. FIG. 11 is a diagram showing a transition of an anode current in a high-frequency welding device according to a third embodiment of the present invention. FIG. 11 is a flow chart showing a welding procedure of a high-frequency welding device with a learning device according to a third embodiment of the present invention. FIG. 13 is a diagram showing a neural network according to a fourth embodiment of the present invention. FIG. 13 is a diagram showing a configuration in which a plurality of processing devices and a learning device are wirelessly connected according to a fifth embodiment of the present invention. 13A and 13B are tree diagrams when the learning device according to the sixth embodiment of the present invention is combined with a joining device and a fusing device to form a joining device with a learning device and a fusing device with a learning device, respectively. FIG. 13 is an external view of a non-contact hot plate welding apparatus 302 with a learning device according to a sixth embodiment of the present invention. FIG. 13 is an external view of a vibration welding apparatus 303 with a learning device according to a sixth embodiment of the present invention. FIG. 13 is an external view of a laser welding apparatus 304 with a learning device according to a sixth embodiment of the present invention. FIG. 13 is an external view of an ultrasonic welding apparatus 305 with a learning device according to a sixth embodiment of the present invention. 13A, 13B, and 13C are cross-sectional views illustrating the welding process of a spin welder with a learning device according to a sixth embodiment of the present invention. FIG. 13 is a diagram showing the change in temperature of the workpieces welded by an impulse welder with a learning device according to a sixth embodiment of the present invention. FIG. 13 is an external perspective view of an ultrasonic fusing apparatus 402 with a learning device according to a seventh embodiment of the present invention. FIG. 13 is a diagram showing the relationship between the workpiece and the appearance and pressing force in the ultrasonic fusing apparatus 402 with a learning device according to the seventh embodiment of the present invention. FIG. 13 is a schematic configuration diagram of an ultrasonic fusing apparatus 402 with a learning device according to a seventh embodiment of the present invention. 13 is a diagram showing the transition of current and the fusing depth of the ultrasonic fusing device 402 with a learning device according to the seventh embodiment of the present invention. FIG. 13A to 13D are diagrams individually showing the cutting depth L obtained when several cutting times (TG) are taken in one state (G) using an ultrasonic cutting apparatus with a learning device according to a seventh embodiment of the present invention. 13A to 13D are perspective views showing an image of the procedure for measuring melting depth using a learning device according to a seventh embodiment of the present invention. FIG. 13 is a flowchart showing a learning procedure according to a seventh embodiment of the present invention. (a) A diagram showing the contents of the data stored in the learning memory means of the seventh embodiment of the present invention. (b) A diagram showing the contents stored during learning. (c) A diagram showing the learning results after learning. FIG. 13 is a flow chart showing a melt-cutting procedure according to a seventh embodiment of the present invention. 13A and 13B are diagrams showing how a learning device according to a seventh embodiment of the present invention identifies a melting condition from learning results stored in a melting learning storage means. FIG. 13 is a diagram showing a transition of current during welding in a conventional welding control method. FIG. 13 is a diagram showing a transition of current during welding in another conventional welding control method. FIG. 1 is an explanatory diagram of an action value function of conventional AI technology. A diagram showing part of the system of conventional AI technology. A tree diagram showing the relationship between state (St), action (at), and reward (Q(St, at)) in an example of reinforcement learning in conventional AI technology. (a) A diagram showing a perceptron, a conventional AI technology. (b) A diagram showing a neural network, a conventional AI technology. A diagram explaining conventional deep reinforcement learning.

Regarding the forms for implementing the processing condition learning device, processing device with learning device, and processing system with learning device of the present invention, in the first to sixth embodiments, a learning device for joining conditions of two or more parts, a joining device with learning device, and a joining system with learning device will be described, and then in the seventh embodiment, a form for implementing a learning device for cutting conditions of one or more parts, a cutting device with learning device, and a cutting system with learning device will be described.

First Embodiment
In the first embodiment, a high-frequency welding method using a learning device that learns the joining conditions for two or more parts, and a high-frequency welding device 301 with a learning device will be described. Fig. 1 shows an external view of the high-frequency welding device 301 with a learning device. Fig. 2 shows the internal configuration of the high-frequency welding device 301 with a learning device. In the high-frequency welding device 301 with a learning device, the learning device 100 is disposed to the right of the high-frequency welding means 1, and the two are connected by a communication line 120 for integral control.

In FIG. 1, the high-frequency welding means 1 has a table 70, an upper die 2 and a lower die 3 that vertically sandwich the workpiece 4, and is provided with a high-frequency current supply means 6 that passes a high-frequency current through the workpiece 4 sandwiched between the upper die 2 and the lower die 3, a status output means 13 that outputs to the learning device 100 the status in which the high-frequency current is being supplied by the high-frequency current supply means 6, a main memory means 11 that stores information necessary for controlling the welding conditions, etc., a welding specification input/output means 12 that inputs and outputs welding specifications, a main control means 10 that controls the entire high-frequency welding means 1, an air cylinder 5, and a lever 71 that operates the air cylinder 5.

In the high-frequency welding means 1, when the lever 71 is pressed down, the force of the air cylinder 5 causes the upper die 2 to press against the top surface of the workpiece 4 placed on the lower die 3 on the table 70. Then, the high-frequency current supply means 6 applies high-frequency current to the workpiece 4 sandwiched between the upper die 2 and the lower die 3, causing the workpiece 4 to generate dielectric heat and weld.

As shown in FIG. 2, the learning device 100 is provided with a thickness measuring means 101, an appearance inspection means 102, a welding force measuring means 103, etc. as a welding result measuring means 130. It is also provided with a welding result evaluating means 104 that evaluates the value measured by the welding result measuring means 130 as a reward, a learning storage means 105 that stores the welding result and the reward for evaluating the welding result, a welding condition learning means 106 that organizes the learning results evaluated based on the welding result and enables the welding conditions to be read out, an I/O means 107 that inputs and outputs the learning plan of the AI technology, specifically, the software of the learning program, a welding condition reading means 108 that reads out the welding conditions that satisfy the welding specifications from the learning storage means 105, a specification monitoring means 109A, and a status monitoring means 109B.

The learning device 100 is divided into two parts, a lower part 100a and an upper part 100b, as shown by the two-dot chain line frame in Figure 2. The lower part 100a includes a welding result measuring means 130, such as a thickness measuring means 101, an appearance inspection means 102, and a welding strength measuring means 103. The upper part 100b includes a welding result evaluating means 104, a learning memory means 105, a welding condition learning means 106, an I/O means 107, a welding condition reading means 108, a specification monitoring means 109A, and a status monitoring means 109B. The communication lines 120 connecting each means in Figure 2, indicated by black circles ●, are connected by connectors 101a, 102a, 103a, 110a, and 110b. The connectors 101a, 102a, 103a, 110a, and 110b are detachable, and the configuration of the welding result measuring means 130 can be rearranged as desired.

In addition, the high-frequency welding means 1 connected to the learning device 100 by the connector 110b can be connected to another processing device to configure it as a processing device with another learning device. As will be described later in FIG. 24, instead of the connector 101a, etc., wireless transmission/reception means 601 to 609 can be used to configure the learning device and the processing means to be spatially separated.

A learning plan is input to the learning device 100 via the I/O means 107. For the learning plan to be input, software for reinforcement learning, Q-learning, or other learning methods, which are AI technologies, is selected as necessary, and the software to be used is stored in the learning storage means 105. In the first embodiment, as an example for understanding the invention, a simple reinforcement learning method software is stored.

In the learning device 100, the welding result measured by the welding result measuring means 130 is evaluated as a reward by the welding result evaluation means 104, and the information on the evaluation of the welding result is stored in the learning storage means 105. The welding condition learning means 106 learns the welding conditions from the stored information in the learning storage means 105, and generates the learning results as a learned model. Then, the welding condition reading means 108 is made capable of reading out the welding conditions that satisfy the welding specifications from the learned model in the learning storage means 105.

The specification monitoring means 109A monitors the latest and current welding specifications, even if the welding specifications have been updated since the previous welding operation or have been changed during welding, and reads out welding conditions that satisfy the latest and current welding specifications from the trained model in the learning storage means 105.

The status monitoring means 109B obtains information on the status of the high-frequency current that the high-frequency welding means 1 has started to flow from the status output means 13 of the high-frequency welding means 1, and transmits this information to the welding condition reading means 108. The welding condition reading means 108 reads out the welding conditions that satisfy the welding specifications in the state of the high-frequency current that has started to flow from the learned model in the learning storage means 105. The status monitoring means 109B also monitors whether the state of the high-frequency current during high-frequency welding is normal or abnormal.

Figure 3 shows the relationship between the transition of the anode current of the high-frequency welding device 301 with a learning device according to the first embodiment and the welding force after welding. The left vertical axis (left Y axis) of Figure 3 shows the value of the high-frequency current flowing through the workpiece 4 sandwiched between the upper die 2 and the lower die 3, i.e., the anode current value (A), and the right vertical axis (right Y axis) shows the welding force after welding (F), in other words, the force required to peel off the workpiece (peeling force). When the high-frequency current starts to flow from the high-frequency current supply means 6, the anode current begins to flow through the workpiece 4 sandwiched between the upper die 2 and the lower die 3. As the anode current flows through the workpiece 4, the current value gradually increases. In other words, the current rises little by little. Even if the anode current flows through the workpiece 4, the heat generation is insufficient until a certain measurement reference current value (It) is reached, so the workpiece 4 does not melt or fuse, and no welding occurs. After that, if the anode current continues to flow, the workpiece 4 melts and fuses, and welding begins. Then, as the welding time progresses from TA1 to TA2, TA3, and TA4, the welding force at which the workpiece 4 is welded increases respectively to FA1, FA2, FA3, and FA4.

As shown in Figure 3, the high-frequency current, that is, the current curve that supplies the high-frequency current from when the anode current rises until it reaches the target current value (IWA), can be considered as one state, "State A." In "State A," welding is performed along the "State A" curve of the current with low energy, or so-called power.

Figure 4 shows the current curve "State D" when the high-frequency current target current value (IWD) is made larger than that in Figure 3. In "State D", welding is performed along the current curve of "State D" where the energy, or power, is large. In "State D", the rise time (tD) is shorter than in "State A", and the welding force of the workpiece 4 becomes larger, as shown by FD1, FD2, FD3, and FD4.

In the present invention, for example, for "state A," as shown in Figures 5(a) to (d), welding is performed for each welding time (TA1, TA2, TA3, TA4), the welding forces after welding (FA1, FA2, FA3, FA4) are individually measured using the welding result measurement means 130, the welding results are evaluated as rewards by the welding result evaluation means 104, and stored in the learning storage means 105. Then, the welding conditions are learned from the welding results evaluated as rewards by the welding condition learning means 106, and the learning results are generated as a learned model. Then, when the welding specifications are input, the welding times can be read out from the learned model as welding conditions that satisfy the welding specifications.

When welding force FA1 is input as the welding specification to the welding specification input/output means 12 of the high-frequency welding means 1, the learning device 100 calculates backwards from the welding force FA1 of the learned model and reads out the welding time TA1 as the action. When welding force FA2 is input as the welding specification, it calculates backwards from the welding force FA2 of the learned model and reads out the welding time TA2 as the action. When welding force FA3 is input as the welding specification, it calculates backwards from the welding force FA3 of the learned model and reads out the welding time TA3 as the action. The same applies when welding force FA4 is input as the welding specification.

When the welding specifications are entered as values between the four welding strengths FA1, FA2, FA3, and FA4, the relationship between the welding time and welding strength is mathematically formulated using regression analysis, a method often used in AI technology, to calculate and read out the welding time corresponding to the input welding strength. Also, by increasing the amount of data to be learned, the system can read out a welding time that satisfies even detailed welding specifications.

To help understand the invention, Figure 6 is shown as a simple example of reinforcement learning. Figure 6(a) shows a tree in which the curves of the anode current flowing in the high-frequency welding device 301 with a learning device according to the first embodiment are each considered to be a "state," the welding time is considered to be an "action," and the obtained welding force is considered to be a Q value (action value) as a "reward."

In Figure 6 (a), the state of the high-frequency welding device is "State A" where the energy, or power, is low, and "State D" where the power is high. In "State A," if the welding times TA1, TA2, TA3, and TA4 are taken as "actions," then the Q values Q(A, TA1), Q(A, TA2), Q(A, TA3), and Q(A, TA4) are obtained as "rewards." Here, the Q value is shown as Q(A, TA1). The A in parentheses indicates that the state is "State A." TA1 indicates that the action is "welding time TA1."

Similarly, in "State D", if the "actions" are welding times TD1, TD2, TD3, and TD4, then the "rewards" obtained are Q values Q(D, TD1), Q(D, TD2), Q(D, TD3), and Q(D, TD4).

If you measure the welding time as an "action" and the Q value as a "reward" for each of "State A" and "State D" in Figure 6(a), then when you want to obtain a certain "reward" in a certain "state," you can find the "action" by calculating backwards from the "reward," as shown in Figure 6(b).

Figures 7(a) to (d) show an image of measuring the welding forces (FA1, FA2, FA3, FA4) when welding four samples with randomly set welding times (TA1, TA2, TA3, TA4) using the high-frequency welding device 301 with learning device according to the first embodiment of the present invention.

In Fig. 7(a) to (d), an upper die 2 with a ring-shaped end surface is pressed against the upper surface of two overlapping elongated plastic sheets 4a and 4b such as polyvinyl chloride sheets, and the sheets are welded into rings at random welding times. For example, in Fig. 7(a), the welding times are TA1, TA2, TA3, and TA4 from the right to form four ring-shaped welded parts. In Fig. 7(b), the welding times are TA4, TA3, TA2, and TA1 from the right, in Fig. 7(c), the welding times are TA2, TA4, TA1, and TA3 from the right, and in Fig. 7(d), the welding times are TA3, TA1, TA4, and TA2 from the right to weld four ring-shaped welded parts. After welding, the upper plastic sheet 4a is lifted vertically and peeled off from the lower plastic sheet 4b to measure the welding strength. The welding result measurement data is stored in the learning storage means 105.

The learning device 100 stores the measurement results as shown in Figures 7(a) to (d) in the learning storage means 105, evaluates them as rewards in the welding result evaluation means 104, learns the relationship between the welding time, which is an action, and rewards in the welding condition learning means 106, summarizes the learning results in a table, and treats the learning results as a learned model. Then, when the welding specifications are input, the welding conditions that satisfy the welding specifications can be read out by the welding condition reading means 108.

Figures 8(a) to (c) show the memory contents stored in the learning memory means 105. In Figure 8, the welding time and welding force are stored as one pair. In Figures 8(a) and (b), each pair is shown with a trapezoid, circle, or square frame to make it easier to understand.

In Figure 8 (a), the measured data is stored in the order of measurement. For example, data 1 shows that, from the right, the pair of welding force FA11, Q value, and Q(A, TA11), the pair of welding force FA21, Q value, and Q(A, TA21), the pair of welding force FA31, Q value, and Q(A, TA31), and the pair of welding force FA41, Q value, and Q(A, TA41) are stored. Data 2 to Data 4 show that each pair is randomly arranged in the order of measurement.

In FIG. 8(b), the welding condition learning means 106 rearranges the data stored in the learning storage means 105 from a random form into a form in which the magnitude of the welding force is arranged in order from smallest to largest. For example, in data 1, the order is reversed compared to FIG. 8(a), and from the left, there are a pair of welding force FA11 and Q value (A, TA11), a pair of welding force FA21 and Q value (A, TA21), a pair of welding force FA31 and Q value (A, TA31), and a pair of welding force FA41 and Q value (A, TA41). In data 2 to data 4, each pair is arranged in the same order as data 1.

In FIG. 8(c), the welding condition learning means 106 calculates the average welding force for each welding time for the data organized in FIG. 8(b), and creates a table from which the welding time can be read out by calculating backwards from the welding force. In the present invention, the contents of FIG. 8(c), which are the learning results learned from data 1 to data 4, are used as the learned model. Then, the welding time can be read out as the welding condition that satisfies the welding force of the welding specifications.

Figure 9 shows the details of the learning procedure in the learning device 100 as a flow diagram. In the flow diagram in Figure 9, the learning plan is input by the I/O means 107 and stored in the learning storage means 105, and then learning is started (step ST1). The target current value is set so that the input welding specifications are satisfied. Since a learning device that has not learned at all is ignorant, it is recommended to input a target current value that is expected to satisfy the welding specifications as an initial value and include it in the welding specifications (step ST2). The anode current begins to flow (step ST3). Here, a question is asked whether the state should be confirmed during welding (step ST4). This is because when the high-frequency welding means 1 has only a single energy supply state, there is only one energy state, so there is no need to confirm the state during welding. In that case, the next step ST5 is skipped (NO in step ST4). When the high-frequency welding means 1 is a versatile processing device and one state can be selected from multiple energy supply states, step ST5 is performed. In step ST5, the rise time until the anode current reaches the measurement reference current value is measured and stored in the learning storage means 105. Then, the anode current is passed for each welding time to perform welding (step ST6). The welding force is measured as a reward for the welding time, which is the action, and corresponding data of the welding time and welding force is created and stored in the learning storage means 105 (step ST7). The welding condition learning means 106 learns the welding conditions from the information stored in the learning storage means 105, generates the learning results as a learned model, and stores the welding conditions in the learning storage means 105 in a form that can be read from the learned model (step ST8).

Then, a question is asked whether a predetermined number of pieces of data on welding time and corresponding welding force have been obtained for one state (i.e., one current curve) (step ST9). If the predetermined number of pieces of data have not been obtained, the process returns to step ST3 (NO in step ST9). If the predetermined number of pieces of data have been obtained (YES in step ST9), a question is asked whether the target current value should be updated and another state should be learned (step ST10). If the target current value is to be updated based on the input learning plan, the process returns to step ST2 (YES in step ST10). If the target current value is not to be updated (NO in step ST10), a question is asked whether learning has been performed as planned (step ST11). If learning has not been performed as planned, the process returns to step ST2 (NO in step ST11), and if learning has been completed as planned (YES in step ST11), the process proceeds to step ST12 and learning ends.

In addition, steps ST8 and ST9 may be interchanged to acquire a predetermined number of data, and then the learning process may be performed and the learning results may be stored in the learning storage means 105.

When updating learning, the process goes from step ST34 to step ST1, and the following steps are performed after step ST1. By performing the steps in the flow diagram of FIG. 9, data on the welding force for each welding time, which is each action, is obtained for one or more states, the welding force data is stored as a reward, the stored data is learned, and the learning results are made into a form that can be read out as welding conditions, and are stored in the learning storage means 105.

Figure 10 shows the procedure of the welding operation as a flow diagram. In Figure 10, the welding specifications are input and the welding operation is started (step ST20). The workpiece is set (step ST21). The question "Confirm the state?" is asked (step ST22). As explained in the flow diagram of Figure 9, when the high-frequency welding means 1 has only a single energy supply state, there is only one energy state, so there is no need to confirm the state during welding. In that case, the next steps ST23 and ST24 are skipped (NO in step ST22). When the high-frequency welding means 1 can select one state from a plurality of energy supply states, the next steps ST23 and ST24 are performed. In step ST23, the rise time of the anode current is measured. Then, the state of the energy that the high-frequency welding means 1 has started to supply is confirmed (step ST24). Then, in the state of the energy that has started to be supplied, are there any welding conditions in the learned model of the learning storage means 105 that satisfy the latest and current welding specifications confirmed by the specification monitoring means 109A? (Step ST25). If the welding conditions are found, they are read from the learning storage means 105 (Step ST26), stored in the main storage means 11, and the high-frequency welding means 1 performs welding (Step ST27). The status monitoring means 109B monitors whether the status during welding is abnormal (Step ST28). If the status is normal (YES in Step ST28), it is confirmed that the welding time for the end of the welding operation has been reached (Step ST29), and the welding operation is terminated (Step ST31). If the status is abnormal in Step ST28, a message indicating the abnormality is displayed and welding is stopped (Step ST30). Then, the welding operation is terminated (Step ST31).

In addition, if the welding conditions are not found in the trained model in the learning storage means 105 in step ST25 (NO in step ST25), a message "Update learning is necessary" is displayed (step ST32). Then, a question is asked whether update learning should be performed (step ST33). If update learning is to be performed, the flow jumps from step ST34 to the flow chart in FIG. 9, and update learning is performed in the following procedure from step ST1 in FIG. 9.

The contents of the learning storage means 105 when the welding conditions are read out in step ST26 are shown in Figure 11. Note that Figure 11 is simplified to allow the reader to get an idea of the invention.

The case of Figure 11 (a) will be explained. First, when the welding force is input as "FA2" as the welding specification (step ST20). When the rise time is measured (step ST23), the rise time is "TA", and therefore the state is confirmed to be "state A" (step ST24). Since the welding force that satisfies the welding specification in "state A" is "FA2" in the table, the welding time "TA2" paired with the welding force "FA2" is read out as the action to obtain the welding force "FA2" of the welding specification, that is, the welding time of the welding conditions.

Similarly, in the case of FIG. 11(b), when the welding force is input as "FD3" as the welding specification (step ST20), the rise time is measured (step ST23), and since the rise time is "TD", the state is confirmed to be "state D" (step ST24). Since the welding force that satisfies the welding specification in "state D" is "FD3" in the table, the welding time "TD3" paired with the welding force "FD3" is read out as the action to obtain the welding force "FD3" of the welding specification, that is, the welding time of the welding conditions.

In the case of FIG. 11(c), when the welding force is input as the welding specification with an upper limit "FA3" and a lower limit "FA2" (step ST20), the rise time is measured (step ST23), and since the rise time is "TA", the state is confirmed to be "State A" (step ST24). Since the welding forces that satisfy the welding specification in "State A" are "FA3" and "FA2" in the table, the welding time "TA3" paired with the welding force "FA3" and the welding time "TA2" paired with the welding force "FA2" are read out as the action to obtain the welding forces "FA3" and "FA2" of the welding specification, that is, the upper limit welding time "TA3" and the lower limit welding time "TA2" of the welding conditions. Once the upper limit welding time "TA3" and the lower limit welding time "TA2" are determined, the welding time within the range between the upper limit welding time "TA3" and the lower limit welding time "TA2" will be used as the welding conditions. Figure 12 shows the range of welding times (TA2 to TA3) that are specified as welding conditions when the upper limit welding force and the lower limit welding force are input as the welding specifications for the case of Figure 11 (c).

Update learning will be explained with reference to FIG. 13. For example, when welding forces FA1 to FA4 and FD1 to FD4 have already been learned, if the welding force entered in the welding specifications is smaller than FA1 or larger than FD4, welding conditions that satisfy the specifications cannot be read from the learned model. If the welding force entered as the welding specifications is not in the learning storage means 105 at step ST25 in FIG. 10 (NO at step ST25), "Update learning required" is displayed at step ST32. Then, when update learning is to be performed (YES at step ST33), the process jumps from step ST34 in FIG. 10 to step ST1 in FIG. 9.

In step ST1 of FIG. 9, since the learning plan has already been input, update learning begins and the process proceeds to step ST2. At this time, the learning device 100 sets the target current value in the direction that will obtain a welding force that satisfies the welding specifications in light of the welding specifications (step ST2). Then, steps ST3 and onward are carried out according to the flow diagram of FIG. 9, and a learning operation is performed to determine the welding conditions that will obtain a welding force that satisfies the processing specifications.

Figure 13 shows an example of updating and learning welding conditions that obtain welding forces smaller than FA1, as shown by FH1 to FH4. From the results of the updating and learning in Figure 13, the welding times are obtained as welding conditions that obtain welding forces FH1 to FH4 that are smaller than FA1.

After the update learning, the process returns to step ST20 in the flow chart of FIG. 10, and when the welding work is resumed, in step ST25, it is determined that the welding conditions exist (YES in step ST25), and the updated and learned welding conditions are read from the learned model in the learning storage means 105 (step ST26). Then, new welding conditions that did not exist before the update learning, for example, the welding time TH1 corresponding to FH1, are read and welding is performed (step ST27), and the welding work is completed (step ST31).

In this way, in the first embodiment of the present invention, the joining conditions for joining two or more parts are learned using a learning device, and when welding specifications are input, the welding conditions that satisfy the input welding specifications are read from the learned model of the learning device, and the two or more parts are joined under the read welding conditions.

This solves the first problem of the present invention, namely, to use a method and a learning device to learn the processing conditions for joining two or more parts, melting one or more parts, or processing the parts by joining two or more parts or melting one or more parts.

And the second objective of the present invention is realized, that is, to constantly support machining under machining conditions that satisfy the machining specifications by monitoring the machining specifications even during machining once machining has begun under those conditions.

The fourth problem of the present invention is also solved, namely, to determine whether joining conditions or fusing conditions that satisfy the processing specifications can be found from the learned learning device, and when joining conditions or fusing conditions that satisfy the processing specifications cannot be found, to indicate the need for updated learning.

The fifth problem of the present invention is solved by, when processing conditions such as joining conditions or fusing conditions that satisfy the processing specifications cannot be found from the learned learning device, updating and learning processing conditions such as joining conditions or fusing conditions that satisfy the processing specifications, and performing processing such as joining two or more parts or fusing one or more parts under the updated and learned processing conditions such as joining conditions or fusing conditions.

The sixth problem of the present invention is solved, namely, that a versatile processing device such as a joining device or a fusing device checks the state of energy applied when starting a process such as joining or fusing, and performs a processing operation such as joining or fusing under processing conditions that satisfy the processing specifications in that state.

Furthermore, the seventh problem of the present invention is solved, namely, by monitoring the energy state of the processing device during processing, and when the energy state changes to an abnormal state, a notification of this abnormal state is displayed and processing is stopped.

The solution to the third problem of the present invention will be explained in the second and third embodiments.

Second Embodiment
In a second embodiment of the present invention, a high-frequency welding apparatus with a learning device for welding synthetic resin sheets such as polyvinyl chloride sheets will be described.

Generally, when parts 4a and 4b are welded together to create a product with a constant thickness as the weld target 4, a tolerance is set for the thickness of each of the parts 4a and 4b. Ideally, if the parts 4a and 4b are within the tolerance, a product that meets the welding specifications will be produced when welded under a single welding condition.

However, when the thicknesses of parts 4a and 4b are the maximum thickness combination within the tolerance, or the minimum thickness combination within the tolerance, the difference in thickness causes the anode current to flow differently.

To help understand the invention, Figure 14 shows an example of the thickness when parts 4a and 4b of the workpiece 4 are combined. In Figure 14, if the minimum value of the thickness within the tolerance of parts 4a and 4b is D1 and the maximum value is D2, then the intermediate combination of the minimum and maximum values is Da (= D1 + D2). The combination of the thinnest minimum values is Db (= 2 x D1). The combination of the thickest maximum values is Dc (= 2 x D2).

Figure 15 shows the progression of the anode current in a high-frequency welding device with a learning device according to the second embodiment of the present invention. When the thickness tolerance of the workpiece is the median value Da (= D1 + D2), the anode current flows in "state A." When the thickness tolerance is the minimum value Db (= D1 + D1), the anode current flows in "state B." And when the thickness tolerance is the maximum value Dc (= D2 + D2), the anode current flows in "state C."

In Figure 15, "State B," in which the thickness of the workpiece is the minimum value Db within the tolerance, rises earlier than "State A," in which the thickness is the median value Da within the tolerance. Also, "State C," in which the thickness is the maximum value Dc within the tolerance, rises later than "State A."

Figure 16 compares the progression of the anode current in "state A" and "state B" and the welding forces after welding. In Figure 16, in "state A", the lines connecting the anode current flowing through parts 4a and 4b and the welding forces FA1, FA2, FA3, and FA4 are shown with solid lines. In "state B", the lines connecting the anode current flowing through parts 4a and 4b and the welding forces FB1, FB2, FB3, and FB4 after welding are shown with dashed dotted lines. Looking at Figure 16, it can be seen that the welding forces FB1, FB2, FB3, and FB4 in "state B" increase faster and are shifted to the left than the welding forces in "state A".

Figure 17 compares the transition of the anode current in "State A" and "State C" and the relationship with the welding force after welding. In Figure 17, in "State A", the lines connecting the anode current flowing through parts 4a and 4b and the welding forces FA1, FA2, FA3, and FA4 are shown with solid lines. In addition, in "State C", the lines connecting the anode current flowing through parts 4a and 4b and the welding forces FC1, FC2, FC3, and FC4 after welding are shown with dashed lines. Looking at Figure 17, it can be seen that the welding forces FC1, FC2, FC3, and FC4 in "State C" increase more slowly than the welding forces in "State A", and are shifted to the right.

If the welding time for condition A is used for conditions B and C as is, there will be excessive or insufficient welding, as in the conventional example shown in Figure 42, and the welding specifications will not be met.

In the second embodiment of the present invention, the welding strength after welding for "State A," "State B," and "State C" of the tree shown in Figure 18 is measured as a Q value when welding is performed for each of multiple welding times, and is learned by a learning device.

The tree in Figure 18 shows that there are three states in a high-frequency welding device with a learning device: "State A" in which an anode current is passed through the median value Da (= D1 + D2) within the tolerance of the workpiece, "State B" in which an anode current is passed through the minimum value Db (= D1 + D1) within the tolerance of the workpiece, and "State C" in which an anode current is passed through the maximum value Dc (= D2 + D2) within the tolerance of the workpiece. Figure 18 also shows that each of the three states is connected to multiple actions and a Q value that serves as a reward for each action.

The learning procedure is the same as that shown in the flow diagram of FIG. 9 in the first embodiment. A detailed explanation of the learning procedure will be omitted to avoid redundancy, but when learning is performed according to the flow diagram of FIG. 9, the learning result shown in FIG. 19 is obtained as a learned model. In FIG. 19, the material and rise time of parts 4a and 4b, the target welding current value, and pairs of welding force and welding time are stored in the learning storage means 105.

The welding procedure is the same as that shown in the flow chart of FIG. 10 of the first embodiment. Although the overlapping parts of the welding procedure will not be described, welding is performed according to the flow chart of FIG. 10. In the main procedure, in step ST23 of FIG. 10, the rise time of the anode current is measured, and in step ST24, it is confirmed whether the state is "state A," "state B," or "state C." If "state A" is confirmed, in step ST25, it is asked whether there are welding conditions that satisfy the welding specifications in state A. Similarly, when "state B" is confirmed based on the rise time of the anode current in step ST24, it is asked in step ST25 whether there are welding conditions that satisfy the welding specifications in state B. When "state C" is confirmed based on the rise time of the anode current in step ST24, it is asked in step ST25 whether there are welding conditions that satisfy the welding specifications in state C. Then, if there are welding conditions in the learned model stored in the learning storage means 105 of the learning device, in step ST26, they are read from the welding storage means 105 and stored in the main storage means 11 of the high-frequency welding means 1, and the parts 4a and 4b are welded using the welding conditions stored in the main storage means 11 to create the product, which is the welded object 4.

As a result, even if the thickness of the workpiece fluctuates within the tolerance, the state in which the anode current flows is confirmed from the rise time, and the welding time that satisfies the welding specifications in that state is read out as the welding condition, and welding is performed using the high-frequency welding means 1.

Using Figure 20, we will explain how the welding time is read out as a welding condition that satisfies the welding specifications. In Figure 20(a), when "state A" is confirmed based on the rise time (TA), the welding time (TA3) is read out as an action, i.e., a welding condition, from the learned model in the learning storage means 105 by calculating backwards from the welding force (FA3) of the welding specifications.

In FIG. 20(b), when "State B" is confirmed based on the rise time (TB), the welding time (TB3) is read out as an action, i.e., welding condition, from the learned model in the learning storage means 105 by calculating backwards from the welding force (FB3) of the welding specifications.

In FIG. 20(c), when "State C" is confirmed based on the rise time (TC), the welding time (TC3) is read out as an action, i.e., welding condition, from the learned model in the learning storage means 105 by calculating backwards from the welding force (FC3) of the welding specifications.

In the present invention, the rise time when the anode current starts to flow is measured to confirm the state of the anode current that has started to flow in the high frequency welding device. Then, the welding time as the welding condition that produces a welding force that satisfies the welding specifications in the confirmed state is found from the learned model in the learning storage means 105 of the learning device 100. If found, the welding condition is read out from the learning storage means 105 of the learning device 100, stored in the main storage means of the high frequency welding means 1, and welding is performed by the main control means.

As a result, even if the thickness of the workpiece fluctuates within the tolerance, welding conditions that satisfy the welding specifications in each state can be read out from the welding conditions learned by the learning device, and welding can be performed.

In the second embodiment, it was explained that the present invention, which addresses the third problem, when the parts to be welded are substantially the same in material, shape, and dimensions but have variations within the tolerances, checks the state in which the anode current begins to flow from the rise time when the anode current begins to be applied, learns the welding time that satisfies the welding specifications in the confirmed state as the welding condition, and reads it out from the learning storage means for welding.

The second embodiment of the present invention is based on the first embodiment, and so, as already explained in the first embodiment, it naturally solves the first, second, fourth to seventh problems of the present invention.

Third Embodiment
In the third embodiment, the latter half of the third object of the present invention will be described, namely, "when a part having a thickness outside the tolerance is set, no joining is performed, thereby preventing the production of defective products in which parts outside the tolerance are joined."

The configuration of the high-frequency welding device 301 with learning device in the third embodiment is basically the same as in the first and second embodiments, so a description will be omitted.

Figure 21 shows the transition of the anode current of the high-frequency welding device with a learning device according to the third embodiment. In the third embodiment, the anode current flowing through the parts 4a and 4b sandwiched between the upper die 2 and the lower die 3 flows in five states, from the left in Figure 21, "State B", "State E", "State A", "State F", and "State C". In each state, the welding force after welding when welding is performed for a plurality of welding times is measured, stored in the learning storage means 105 of the learning device 100, and learning is performed so that the welding conditions that satisfy the welding specifications can be read out, which is the same as in the first and second embodiments already described.

In the third embodiment, the rise time ranges (tB-tE), (tE-tA), (tA-tF), and (tF-tC) are set for the range of the rise time from when the anode current starts to flow until it reaches a predetermined current value (measurement reference current value). The state of the anode current is determined based on which rise time range the rise time (t) during welding operation falls within. The welding conditions that satisfy the welding specifications from the determined state are read from the trained model in the learning storage means 105. Then, the parts 4a and 4b sandwiched between the upper die 2 and the lower die 3 are welded under the read welding conditions.

Figure 22 shows a flow diagram of the welding procedure of the third embodiment. In Figure 22, the welding specifications are entered and the welding operation is started (step ST20). The workpieces to be welded are set (step ST21). In the third embodiment, the condition must be checked, so the rise time (t) of the anode current is measured (step ST23). For this reason, step ST22 in Figure 10 is not present in Figure 22. The condition is checked based on which rise time range the rise time (t) falls within (step ST24).

In the third embodiment, detailed steps are provided in step ST24. In FIG. 22, the detailed steps of step ST24 are shown within a dashed line frame.

The detailed procedure of step ST24 is to determine whether the rise time (t) is equal to or greater than (tB) and is equal to or smaller than (tE) (step ST40). If the answer is YES (YES in step ST40), it is determined that the state is E (step ST41). Then, the welding time is read out as a welding condition that satisfies the welding specification in state E (step ST26), and welding is performed (step ST27). If the answer is NO in step ST40, it is determined whether the rise time (t) is greater than (tE) and is equal to or smaller than (tA) (step ST42). If the answer is YES (YES in step ST42), it is determined that the state is A (step ST43). Then, the welding time is read out as a welding condition that satisfies the welding specification in state A (step ST26), and welding is performed (step ST27). If the answer is NO in step ST42, it is determined whether the rise time (t) is greater than (tA) and is equal to or smaller than (tF) (step ST44). If the answer is YES (YES in step ST44), it is determined that the state is F (step ST45). Then, the welding time is read out as the welding condition that satisfies the welding specification in state F (step ST26), and welding is performed (step ST27). If the answer is NO in step ST44, it is determined whether the rise time (t) is greater than (tF) and is equal to or less than (tC) (step ST46). If the answer is YES (YES in step ST46), it is determined that the state is C (step ST47). Then, the welding time is read out as the welding condition that satisfies the welding specification in state C (step ST26), and welding is performed (step ST27).

The procedure for completing the welding work from step ST27 after step ST26 to step ST31 is the same as that shown in FIG. 10 described in the first embodiment, so the explanation will be omitted.

If the answer is NO in step ST46, that is, if the state is not "state E", "state A", "state F", or "state C", the thickness of the overlapped parts 4a and 4b is not within the tolerance range, that is, it is outside the tolerance range. In this case, proceed to step ST48 (NO in step ST46). Then, determine whether the rise time (t) is smaller than (tB) (step ST48). If the answer is YES (YES in step ST48), it is notified that the rise time is too short (step ST49). If the answer is NO in step ST48, it is determined whether the rise time (t) is larger than (tC) (step ST50). If the answer is YES (YES in step ST50), it is notified that the rise time is too long (step ST51). If the answer is NO in step ST50, proceed to step ST32. The procedure from step ST32 to the end of the welding work (step ST31) is the same as that in FIG. 10 described in the first embodiment, so the description will be omitted.

By carrying out the procedure shown in Figure 22, when the anode current starts to flow, the rise time of the anode current is measured to confirm the state in which the high-frequency welding device is attempting to weld. Then, the welding conditions that satisfy the welding specifications in the confirmed state are read out.

By carrying out the above procedure, welding occurs only when the rise time of the anode current is within a certain range, i.e., when the thicknesses of both parts 4a and 4b are within the tolerance range, and does not occur when it is outside the certain range, i.e., when the thicknesses of both parts 4a and 4b are outside the tolerance range. In the third embodiment, when the thicknesses of the parts to be welded are determined by a tolerance, only parts to be welded that are within the tolerance are welded to produce a good product, and parts to be welded that are outside the tolerance are not welded. It has been explained that no defective products are produced.

As described in the second embodiment, the first half of the third objective of the present invention is to provide a learning device that, when the materials, shapes, and dimensions of the parts to be joined are approximately the same but there is variation within the tolerances, determines the state of energy supply according to the variation, and learns the energy supply time that maximizes the reward under that state as the joining condition or fusing condition.

In addition, since the third embodiment of the present invention is already based on the first embodiment, it solves the first, second, fourth to seventh problems of the present invention as explained in the first embodiment.

(Fourth embodiment)
In the fourth embodiment of the present invention, a learning method using a neural network will be described. In the learning method using a neural network, a multi-layered neural network is created by stacking multiple perceptrons, and when an action is input to the input layer of the multi-layered neural network, each input is weighted, learning is performed in the hidden layer, and a policy obtained by multiplying the maximum reward by a probability (probability of being adopted) is output from the output layer.

In the fourth embodiment, as shown in Figure 23, learning is performed by adding "material" to "thickness" and "welding force." When "thickness," "welding force," and "material" are input to the input side, "anode current state" and "welding time" are output as welding conditions to the output side.

In the fourth embodiment of the present invention, a learning plan using the multi-layered neural network shown in FIG. 23 is input from the I/O means of the learning device and stored in the learning storage means 105 in advance, and learning is performed in the welding condition learning means 107 taking into consideration "thickness," "welding force," and "material" in a composite manner, and the "anode current state" and "welding time" are output to generate a learned model. Then, welding conditions that satisfy the welding specifications are read from the learned model in the learning storage means 105. After the welding conditions that satisfy the welding specifications are read, they are stored in the main storage means 11 of the high-frequency welding means 1, and the workpiece 4 is welded by the main control means 10.

In the above, the first to fourth embodiments have described a learning device for joining conditions for two or more parts, a joining device using the same, and a joining system using the same.

Fifth embodiment
In the fifth embodiment of the present invention, a processing condition learning device, a processing device using the same, and a processing system using the same will be described. Note that in this specification, processing will be described as a higher-level concept including joining, fusing, and the like.

Figure 24 shows a schematic diagram of the fifth embodiment of the present invention. In Figure 24, three processing devices, a first processing device 501, a second processing device 502, and a third processing device 503, are arranged, and transmitting/receiving means 601, 602, and 603 are attached to the three processing devices, respectively. A transmitting/receiving means 604 that communicates with these transmitting/receiving means is attached to the learning device 200, and a system is shown in which one learning device 200 operates the three processing devices as a processing device with a learning device.

The transmission/reception means 605 of the learning device 200 communicates with the transmission/reception means 606, 607, 608, and 609 to which the thickness measurement means 201, the appearance inspection means 202, the welding strength measurement means 203, and the cutting depth measurement means 204 are respectively attached, thereby acquiring and learning from the processing result measurement data.

For example, one learning device 200 can be installed in one factory, and multiple processing devices 501, 502, 503, etc. can be installed in other factories on the same premises so that they can communicate with the learning device 200 individually, allowing the learning device 200 to be shared. The amount of learning data increases, and the system can be operated as a joint system in which learning progresses.

Instead of the first processing device 501, the second processing device 502, and the third processing device 503, the system may be configured in which the first welding device 371, the second welding device 372, and the third welding device 373 are respectively equipped with the transmitting/receiving means 601, 602, and 603. Also, the system may be configured in which the first fusing device 471, the second fusing device 472, and the third fusing device 473 are respectively equipped with the transmitting/receiving means 601, 602, and 603.

It can also be used as an IoT processing system by using the Internet as a means of communication.

(Sixth embodiment)
25 shows a system diagram of a joining device 300 equipped with a learning device 150 of the present invention, and a fusing device 400 equipped with a learning device 150. The learning device for a fusing method for one or more components, the fusing device using the same, and the fusing system using the same will be described later as a seventh embodiment of the present invention.

Examples of joining devices 350 include "high frequency welding devices," "non-contact hot plate welding devices," "vibration welding devices," "laser welding devices," "ultrasonic welding devices," "ultrasonic metal joining devices," and "ultrasonic joining devices." These joining devices apply vibrational or thermal energy to the parts to heat, melt, or solid-state bond the parts and join them to create products. Therefore, they have in common the fact that they can find joining conditions that meet processing specifications by changing the amount of energy applied to the parts and the time for which the energy is supplied.

Therefore, by attaching a learning device to these joining devices, they can be used to perform the joining method using the learning device of the present invention, respectively as "high frequency welding device with learning device 301", "non-contact hot plate welding device with learning device 302", "vibration welding device with learning device 303", "laser welding device with learning device 304", "ultrasonic welding device with learning device 305", "ultrasonic metal joining device with learning device 306", and "ultrasonic joining device with learning device 307".

Then, by attaching the learning device 150 to the fusing device 450, the fusing method can be performed using the learning device of the present invention as a "laser fusing device with learning device 401" and a "ultrasonic fusing device with learning device 402", respectively.

Although not shown in Figure 25, this can also be applied to a "spin welder" or "impulse welder." These will be described later using Figures 30 and 31.

Figure 26 shows an external view of the non-contact hot plate welding device 302 with a learning device of the present invention. In the non-contact hot plate welding device 352, the welding conditions are determined by the amount of thermal energy used to heat the hot plates 352a and 352b to a specified temperature and the time for which the thermal energy is applied.

For example, welding can be performed by learning the surface temperature at which the hot plates 352a and 352b are maintained as the "state" of the device, the heating time during which they are brought close to the workpiece 4W1 and heated as the "action," and the welding power or aesthetic level as the "reward."

The learning procedure and welding procedure are the same as those in the first embodiment. Therefore, by applying the joining method of the present invention already described in the first to fifth embodiments, the learning device 150 can be attached to the "non-contact hot plate welding device 302 with learning device" to perform welding using the learning device of the present invention.

Figure 27 shows an external view of the vibration welding device 303 with learning device of the present invention. The vibration welding device 353 is a joining device that applies low-frequency vibration of about 100 Hz to one part 4W2a with an upper die 353a, rubs the surface of the other part 4W2b fixed to a lower die 353b, generates frictional heat, and welds the contact surfaces of the parts 4W2a and 4W2b with the frictional heat. The welding conditions are determined by the frictional force to be generated and the time for which the frictional force is applied.

For example, welding can be performed by learning the magnitude of the frictional force applied to the workpiece parts 4W2a and 4W2b as the "state," the heating time for rubbing and heating the workpiece parts 4W2a and 4W2b, i.e., the time for applying the frictional force, as the "action," and the welding force or aesthetic level as the "reward."

The learning procedure and the joining/welding procedure are the same as in the first embodiment. Therefore, by applying the joining method of the present invention already described in the first to fifth embodiments, the learning device 150 can be attached to the "vibration welding device 303 with learning device" to perform welding using the learning device of the present invention.

Figure 28 shows an external view of the laser welding device 304 with learning device of the present invention. In the laser welding device 354, the welding conditions are determined by the amount of energy of the laser beam 354a and the time for which the energy is applied.

For example, welding can be performed by learning the magnitude of the laser light output as the "state," the laser light irradiation time for irradiating the pair of workpieces 4W3 with the laser light as the "action," and the welding force or aesthetic level as the "reward."

The learning procedure and joining/welding procedure are the same as in the first embodiment. Therefore, by applying the joining method of the present invention already described in the first to fifth embodiments, the learning device 150 can be attached to the "laser welding device with learning device 304" to perform welding using the learning device of the present invention.

Figure 29 shows an external view of an ultrasonic welding device 305 with a learning device. In the ultrasonic welding device 355, a tool horn 355a that is ultrasonically vibrating in the vertical direction is pressed against a pair of workpieces 4W4 to weld them.

The welding conditions are determined by the magnitude of the ultrasonic vibration energy and the time the energy is applied.

For example, welding can be performed by learning the frequency and amplitude of ultrasonic vibrations as the "state," the time it takes to apply ultrasonic vibrations as the "action," and the welding force or aesthetic level as the "reward."

The learning procedure and the joining/welding procedure are the same as in the first embodiment. Therefore, by applying the joining method of the present invention already described in the first to fifth embodiments, the learning device 150 can be attached to the "ultrasonic welding device 305 with learning device" to perform welding using the learning device of the present invention.

Although the external appearance of the device is not shown, the ultrasonic metal welding device 306 is a welding device that applies ultrasonic vibrations parallel to the contact surfaces of the metals to solid-phase bond the contact surfaces of the metals. The welding conditions are determined by the magnitude of the energy of the ultrasonic vibrations and the time for which the energy is applied.

The learning procedure and the joining/welding procedure are the same as in the first embodiment. Therefore, by applying the joining method of the present invention already described in the first to fifth embodiments, the learning device 150 can be attached to form an "ultrasonic metal joining device 306 with learning device" to perform joining using the learning device of the present invention.

Similarly, although the external appearance of the device is not shown, the ultrasonic bonding device 307 is a bonding device that places a metal sheet on the surface of glass and applies ultrasonic vibrations parallel to the contact surfaces of the two pieces of glass to solid-phase bond the metal sheet to the surface of the glass. The welding conditions are determined by the magnitude of the energy of the ultrasonic vibrations and the time for which the energy is applied.

The learning procedure and the joining/welding procedure are the same as in the first embodiment. Therefore, by applying the joining method of the present invention already described in the first to fifth embodiments, the learning device 150 can be attached to the "ultrasonic joining device 307 with learning device" to perform joining using the learning device of the present invention.

Figures 30(a) to (c) are cross-sectional views showing the welding process of a spin welder. As shown in Figure 30(a), the spin welder is a welding device that holds two cylindrical plastic parts 4W6a and 4W6b in contact with each other with their end faces facing each other, rotates the upper part 4W6a at high speed while keeping the lower part 4W6b fixed as shown in Figure 30(b), and welds the contacting surfaces together with frictional heat as shown in Figure 30(c). The upper part 4W6a is rotated at high speed to impart kinetic energy, which generates frictional heat at the contact surface with the lower part 4W6b, melting and joining the parts together. For this reason, the welding conditions are determined by the magnitude of the rotational kinetic energy imparted to the upper part 4W6a and the time for which the rotational kinetic energy is imparted. In the present invention, the welding conditions are learned by a learning device and then welded.

The learning procedure and the joining/welding procedure are the same as in the first embodiment. Therefore, the joining method of the present invention already described in the first to fifth embodiments can be applied to join two parts as a spin welder with a learning device.

Figure 31 shows the change in temperature of the workpieces when a large current is passed through them with an impulse welder. An impulse welder is a welding device that instantaneously passes a large current through two or more parts to generate heat and cool them, melting and joining the parts together. As shown in Figure 31, the welding conditions are determined by the amount of heat generated by passing a large current to heat and cool the parts, and the heating and cooling times. In the present invention, these welding conditions are learned by a learning device and then welded.

The learning procedure and joining/welding procedure are the same as in the first embodiment. Therefore, the joining method of the present invention already described in the first to fifth embodiments can be applied to join two parts as an impulse welder with a learning device.

Seventh embodiment
As a seventh embodiment of the present invention, a learning device for a method of fusing one or more components, a fusing device with a learning device, and a fusing system with a learning device will be described.

The cutting device 450 includes "laser cutting device" and "ultrasonic cutting device". These cutting devices cut one or more parts by irradiating the workpiece with a laser beam or pressing an ultrasonically vibrating cutter against it. The cutting conditions for these cutting devices are determined by the amount of energy of the laser beam or ultrasonic vibration, and the time for which the energy is applied. In the tree in Figure 25, "laser cutting device with learning device 401" and "ultrasonic cutting device with learning device 402" are shown in the bottom right.

The "laser cutting device with learning device 401" is not shown, but has a similar appearance to the laser welding device in FIG. 28. A laser beam is irradiated from the laser cutting means of the "laser cutting device with learning device 401" to the object to be cut, and cutting is performed. The procedure for learning the cutting conditions and the procedure for cutting are almost the same as in the first embodiment. Therefore, by applying the method of the present invention already described in the first to fifth embodiments and attaching the learning device 150, cutting can be performed using the learning device of the present invention as the "laser cutting device with learning device 401" and the "ultrasonic cutting device with learning device 402", respectively.

The ultrasonic fusing device 402 with learning device will be described in detail with reference to Figs. 32 to 41. Fig. 32 shows an external perspective view of the ultrasonic fusing device 402 with learning device for fusing one or more parts of the present invention. The ultrasonic fusing means 452 shown in Fig. 32 has an ultrasonic cutter 412 attached to the lower end of an ultrasonic transducer 411 that ultrasonically vibrates up and down, and the ultrasonic cutter 412, which is ultrasonically vibrated by an air cylinder 414 attached to a support 410, is pressed against the workpiece 4W5 placed on a table 413 to fuse it.

As shown in Figure 33, the objects to be cut 4W5 include relatively soft foods such as sponge cake, cake with strawberries, and freshly pounded rice cakes, as well as relatively hard foods such as noshi-mochi (rolled rice cakes), three-colored diamond-shaped rice cakes, rubber sheets, and old tires. Figure 33 shows the type of object to be cut, whether or not it needs to look good after cutting, and the amount of pressure applied. Some objects need to look good after cutting, while others do not. It also shows that some objects only need a small amount of pressure to cut, while others require a large amount of pressure.

In Figure 33, food is required to maintain its shape even when cut with a heat source. On the other hand, rubber sheets and old tires only need to be cut, and there are few problems if the surface condition changes or the shape changes.

The ultrasonic fusing device with learning device 402 determines the fusing conditions based on the magnitude of ultrasonic energy and the fusing time according to the fusing specifications of the object to be fusing. This is different from "fusing" and the "welding" of the first embodiment, but they have in common the point that the magnitude of ultrasonic energy and the energy supply time are controlled. Therefore, the contents of the present invention described in the first to fifth embodiments can be used as a fusing device with learning device.

In the following explanation, we will omit the parts that overlap with the explanations of the first to fifth embodiments already explained, but will focus on the main points.

Figure 34 is a schematic diagram of an ultrasonic fusing device 402 with a learning device according to a seventh embodiment of the present invention. In terms of content, it is similar to the schematic diagram in Figure 2. Instead of the high-frequency welding means 1, an ultrasonic fusing means 452 is connected to the learning device 150 via a communication line 170 and is controlled integrally.

The learning device 150 has different names from the learning device 100 described in the first embodiment, but the functions are the same or similar. That is, the learning device 150 is provided with a thickness measurement means 151, an appearance inspection means 152, a cutting depth measurement means 153, etc. as a cutting result measurement means 180. It also has a cutting result evaluation means 154 that evaluates the value measured by the cutting result measurement means 180 as a reward, a cutting learning storage means 155 that stores information on the evaluation of the cutting result, a cutting condition learning means 156 that learns the cutting conditions from the information stored in the cutting learning storage means 155, an I/O means 157 that inputs and outputs the learning plan of AI technology, a cutting condition reading means 158 that reads out the cutting conditions that satisfy the cutting specifications from the cutting learning storage means 155, a specification monitoring means 159A, and a state monitoring means 159B.

The specification monitoring means 159A monitors the latest and current cutting specifications, even if the processing specifications have been updated since the previous cutting operation or have been changed during cutting, and reads out the cutting conditions that satisfy the latest and current cutting specifications from the learned model in the cutting learning storage means 155.

The status monitoring means 159B obtains information on the state of the current that the ultrasonic cutting means 452 has started to flow from the status output means 463 of the ultrasonic cutting means 452, and transmits this information to the fusing condition reading means 158. The fusing condition reading means 158 reads out the fusing conditions that satisfy the fusing specifications in the state of the current that has started to flow from the learned model in the fusing learning storage means 155. The status monitoring means 159B also monitors whether the state of the current during ultrasonic fusing is normal or abnormal.

Figure 35 shows the transition of the current value and the value of the fusion depth when ultrasonic energy is applied by the fusing device with learning device according to the seventh embodiment of the present invention. The X-axis is the time axis of the fusing operation, and the left Y-axis is the value of the current flowing through the ultrasonic driving means. The right Y-axis is the value of the fusion depth. When the tip of the cutting edge of the ultrasonic cutter is pressed against the object to be cut and the current starts to flow, fusing begins after a relatively short rise time.

Figure 36 shows the cutting time for several actions (TG) in one state (G) in a cutting device according to the seventh embodiment of the present invention, the reward when cutting is performed within that cutting time, and the cutting depth as a Q value (Q(G,TG)).

Here, the learning device 150 performs melting for each melting time (TG1, TG2, TG3, TG4) for "state G" as shown in Figures 36 (a) to (d), measures the melting depth values after melting (LG1, LG2, LG3, LG4) individually using the melting depth measuring means 153 of the melting result measuring means 180, evaluates the melting results as rewards using the melting result evaluation means 154, and stores them in the melting learning storage means 155. Then, the melting condition learning means 156 learns the melting conditions from the melting results evaluated as rewards, so that when the melting specifications are input, the melting time can be read out as the melting conditions that satisfy the melting specifications.

For example, if cutting depth LG1 is input as the cutting specifications to the cutting specifications input/output means 462, which inputs the cutting specifications of the ultrasonic cutting means 452, the learning device 150 calculates backwards from the cutting depth LG1 and reads out the cutting time TG1 as the action. If cutting depth LG2 is input as the cutting specifications, it calculates backwards from the cutting depth LG2 and reads out the cutting time TG2 as the action. If cutting depth LG3 is input as the cutting specifications, it calculates backwards from the cutting depth LG3 and reads out the cutting time TG3 as the action. The same applies when cutting depth LG4 is input as the cutting specifications.

When a numerical value between the above four cutting depths LG1, LG2, LG3, and LG4 is input as the cutting specification, the cutting time corresponding to the input cutting depth is calculated and read out by formulating the relationship between the cutting time and the cutting depth using regression analysis, which is often used in AI technology. Also, by increasing the amount of data to be learned, a cutting time that satisfies the specifications can be read out even for detailed cutting specifications. This is the same as with the welding device of the first embodiment.

Figure 37 shows an image of the procedure for measuring the cutting depth of four samples when the cutting times (TG1, TG2, TG3, TG4) are set randomly using an ultrasonic cutting device 402 with a learning device according to the seventh embodiment of the present invention.

In Figure 37, a rubber sheet is used as the object to be cut 4W5, and an ultrasonic cutter 412 with a straight edge (not shown) is pressed against the upper surface of the rubber sheet, and the rubber sheet is cut at random cutting times. Figure 37 shows cases where the cutting depth L is shallow and the sheet is not cut, cases where the cutting depth L is deep, and cases where the sheet is completely cut.

In FIG. 37(a), the cutting times are TG1, TG2, TG3, and TG4 from the right; in FIG. 37(b), the cutting times are TG4, TG3, TG2, and TG1 from the right; in FIG. 37(c), the cutting times are TG2, TG4, TG1, and TG3 from the right; and in FIG. 37(d), the cutting times are TG3, TG1, TG4, and TG2 from the right. The cutting depth L is measured using the value of the descent amount of the ultrasonic cutter 411 measured by the cutting depth measuring means 153. The cutting time that satisfies the cutting specifications is stored in the cutting learning storage means 155 as the cutting condition that satisfies the cutting specifications.

Figure 38 shows a flow diagram of the learning procedure according to the seventh embodiment of the present invention. In the flow diagram of Figure 38, the learning procedure of reinforcement learning is input as a learning plan by the I/O means 157, stored in the fusing learning storage means 155, and then learning is started (step ST51). Based on the fusing specifications, a target current value is set so as to satisfy the fusing specifications. Since a learning device that has not learned at all is ignorant, it is advisable to input a target current value that is expected to satisfy the fusing specifications as an initial value and include it in the welding specifications (step ST52). A current is passed through the ultrasonic driving means (step ST53). Here, a question is asked whether the state is to be confirmed at the time of fusing (step ST54). This is because when the ultrasonic fusing means 452 has only a single energy supply state, there is only one energy state, so there is no need to confirm the state at the time of fusing. In that case, the next step ST55 is skipped (NO in step ST54). When the ultrasonic fusing means 452 is a versatile processing device and one state can be selected from a plurality of energy supply states, step ST55 is performed. In step ST55, the rise time until the reference current value is reached is measured and stored in the melting learning storage means 155. Then, a current is passed for each melting time to melt the object (step ST56). The melting depth L is measured as a reward for the behavior, and data corresponding to the melting time is created and stored in the melting learning storage means 155 (step ST57). The melting condition learning means 156 learns the melting conditions from the information stored in the melting learning storage means 155, and stores the learning results in a form that can be read out as melting conditions in the melting learning storage means 155 as a learned model (step ST58).

Then, a question is asked whether a predetermined number of pieces of data corresponding to the melting time have been obtained for one state (i.e., one current curve) (step ST59). If the predetermined number of pieces of data have not been obtained, the process returns to step ST53 (NO in step ST59). If the predetermined number of pieces of data have been obtained (YES in step ST59), a question is asked whether the target current value should be updated and other states should be learned (step ST60). If the target current value is to be updated, the process returns to step ST52 (YES in step ST60). If other states are not to be learned (NO in step ST60), a question is asked whether learning has been performed as planned (step ST61). If learning has not been performed as planned, the process returns to step ST52 (NO in step ST61), and if learning has been completed as planned (YES in step ST61), the process proceeds to step ST62 and learning is terminated. Note that steps ST58 and ST59 may be interchanged, and a predetermined number of pieces of data may be obtained before learning and the learning results may be stored in the melting learning storage means 155.

When updating learning, the process goes from step ST84 to step ST51, and the following steps are performed after step ST51. By performing the steps in the flow diagram of FIG. 38, data on the cutting depth L for each cutting time, which is each action, is obtained for the state in which the ultrasonic cutter 412 is vibrating, the data is stored as a reward, the stored data is learned, and the learning results are made into a form that can be read out as cutting conditions, and are stored in the cutting learning storage means 155.

As a result, if the cutting depth L is input in the cutting specifications, the cutting time can be read out from the cutting learning storage means 155 as the cutting condition that satisfies the cutting specifications.

When performing the fusing operation, the fusing time as the fusing condition that satisfies the fusing specifications is read from the learned model of the learning device, stored in the main memory means of the ultrasonic fusing means 452, and then fusing is performed.

Figure 39 shows the contents of the learning results stored in the storage means according to the seventh embodiment of the present invention. In Figure 39, the melting time and melting depth L are stored as one pair. In addition, each pair is shown with a trapezoid, circle, or square frame to make it easier to understand.

In FIG. 39(a), the measured data is stored in the order in which it was measured. For example, data 1 shows that, from the right, a pair of melting depth LG11 and Q value (G, TG11), a pair of melting depth LG21 and Q value (G, TG21), a pair of melting depth LG31 and Q value (G, TG31), and a pair of melting depth LG41 and Q value (G, TG41) are stored. In data 2 to data 4, each pair is arranged randomly.

In FIG. 39(b), the data stored in the cutting learning storage means 155 is organized by cutting depth L by the cutting condition learning means 156. For example, in data 1, the order is reversed compared to FIG. 39(a), and from the left, there is a pair of cutting depth LG11 and Q value (G, TG11), a pair of cutting depth LG21 and Q value (G, TG21), a pair of cutting depth LG31 and Q value (G, TG31), and a pair of cutting depth LG41 and Q value (G, TG41). For data 2 to data 4, each pair is arranged in the same order as data 1.

In FIG. 39(c), the cutting condition learning means 156 calculates the average cutting depth L for each cutting time for the data organized in FIG. 39(b), and creates a table from which the cutting time can be read out by calculating backwards from the cutting depth L. This is then used as the learned model. Since data 1 to data 4 have been organized, when cutting depth L is input as the cutting specifications, the cutting time can be read out from FIG. 39(c) as the cutting condition that satisfies the cutting specifications.

In the fusing operation, the fusing operation is performed according to the fusing operation procedure shown in the flow diagram of FIG. 40. In FIG. 40, the fusing specifications are entered to start the fusing operation (step ST70). The object to be fusing is set (step ST71). A question is asked, "Do you want to confirm the state?" (step ST72). As explained in the flow diagram of FIG. 38, when the ultrasonic fusing means 452 has only a single energy supply state, there is only one energy state, so there is no need to confirm the state at the time of fusing. In that case, the next steps ST73 and ST74 are skipped (NO in step ST72). When the ultrasonic fusing means 452 can select one state from a plurality of energy supply states, the next steps ST73 and ST74 are performed. In step ST73, the rise time is measured. Then, the state of the energy that the ultrasonic fusing means 452 has started to supply is confirmed (step ST74). In the confirmed state, a question is asked, "Are there any fusing conditions that satisfy the latest and current fusing specifications?" (step ST75). If the cutting conditions are found, they are read from the cutting learning storage means 155 (step ST76) and stored in the main storage means 461. Cutting is performed by the ultrasonic cutting means 452 (step ST77). The status monitoring means 159B monitors whether the status during cutting is abnormal (step ST78). If the status is normal (YES in step ST78), it is confirmed that the cutting time that ends the cutting operation has been reached (step ST79), and the cutting operation is terminated (step ST81). If the status is abnormal in step ST78, an abnormality is displayed and the cutting is stopped (step ST80). Then, the cutting operation is terminated (step ST81).

In addition, if no fusing conditions that satisfy the fusing specifications are found in the fusing learning storage means 155 in step ST75 (NO in step ST75), a message "Update learning is required" is displayed (step ST82). Then, a question is asked whether update learning should be performed (step ST83). If update learning is to be performed, the flow jumps from step ST84 to the flow chart in FIG. 38, and update learning is performed in the following procedure from step ST50 in FIG. 38.

Figure 41 shows the contents of the cutting learning storage means 155 when the cutting conditions are read out in step ST76. In the case of Figure 41 (a), "LG2" is input as the cutting depth L of the cutting specifications, and the cutting operation is started (step ST70). When the rise time is measured (step ST73), the rise time is "TG", and the state is confirmed to be "state G" (step ST74). Since the cutting depth L that satisfies the cutting specifications in "state G" is "LG2" in the table, the cutting time "TG2" paired with the cutting depth "LG2" is the action to obtain the cutting depth "FD3" of the cutting specifications, in other words, the cutting time "TG2" is read out as the cutting condition.

Similarly, in the case of FIG. 41(b), "LH3" is input as the cutting depth L of the cutting specifications, and the cutting operation is started (step ST70). When the rise time is measured (step ST73), the rise time is "TH", and therefore the state is confirmed to be "state H" (step ST74). Since the cutting depth L that satisfies the cutting specifications in "state H" is "LH3" in the table, the cutting time "TH3" paired with the cutting depth "LH3" is read as the action to obtain the cutting depth "FD3" of the cutting specifications, that is, the cutting time as the cutting condition is "TH3". The update learning is the same as that described in the first embodiment, so a description thereof will be omitted.

As explained above, in the present invention, processing conditions such as joining conditions or cutting conditions are learned by a learning means using AI technology, and welding conditions that realize processing specifications such as welding specifications and cutting specifications that satisfy the processing specifications are extracted from the learning results, and the workpiece is welded, cut, etc. under the extracted welding conditions, cutting conditions, etc.

By using the processing method learning device of the present invention and the processing device using the same, the present invention makes it possible to use the method and learning device to learn processing conditions such as joining conditions and cutting conditions when joining two or more parts or processing one or more parts such as fusing, to join two or more parts or to fusing one or more parts.

In addition, even if the processing specifications have been updated since the previous processing operation or have been changed during processing, the present invention can monitor the latest and current processing specifications with the specification monitoring means, and can read out processing conditions such as joining conditions or cutting conditions that satisfy the latest and current processing specifications from the learning storage means to perform processing operations such as joining and cutting. Therefore, it is possible to switch models after learning to handle cutting operations with complex cutting specifications, such as when the hardness of the object to be cut changes midway, such as in the case of old tires containing steel wire, or when hard fruit is contained in a soft sponge cake, such as strawberry cake.

The present invention uses a method and a learning device for learning joining and fusing conditions to stably join two or more parts or fusing one or more parts when the parts to be joined are approximately the same in material, shape, and dimensions but have variations within the tolerances.

Furthermore, the present invention can provide a method and device for determining whether joining conditions or fusing conditions that satisfy the processing specifications can be found from the learning storage means, and when joining conditions or fusing conditions that satisfy the processing specifications cannot be found, indicating the need for updated learning.

When the present invention determines that no joining conditions or fusing conditions that satisfy the processing specifications can be found from the learning storage means, it updates and learns joining conditions or fusing conditions that satisfy the processing specifications, and can join two or more parts or fusing one or more parts under the updated and learned joining and fusing conditions.

In addition, when a single joining device or a single fusing device has multiple energy level conditions for applying to the workpiece, the present invention can confirm the state of energy that actually begins to be supplied, read from the learning device the joining conditions or fusing conditions that satisfy the processing specifications in the confirmed state, and perform the joining or fusing operation.

Furthermore, the present invention monitors the state of the processing device after processing has begun, and if the state of the processing device changes and becomes abnormal, this is indicated and processing by the processing device can be stopped.

In addition, the present invention is configured to check the state of energy when it actually starts to be supplied, read from the learning device the processing conditions that are actions that satisfy the processing specifications in the checked state, and perform processing, and the learning device monitors whether the state of the processing device after processing has started is normal or abnormal. This is useful for a processing method with a learning device, a processing device using the same, and a processing system using the same.

In the explanation of the present invention, a simple example of reinforcement learning was shown as an AI technology to help understand the invention. AI technology will continue to evolve in the future. Therefore, in the learning device 100 of the present invention, by using the I/O means 107 to use any of complex reinforcement learning, Q-learning, statistical learning (neural network), deep learning, and deep reinforcement learning that combines deep learning and reinforcement learning as a learning plan (learning algorithm, learning software), it is possible to provide a learning device with more advanced processing conditions, a processing device using this, and a processing system using this.

The present invention can be widely applied to high-frequency welding devices for polyvinyl chloride sheets and the like, as well as to other learning devices for joining methods of two or more parts, joining devices using the same, joining systems using the same, learning devices for fusing methods of one or more parts, fusing devices using the same, and fusing systems using the same.

1 High frequency welding means 2 Upper die 3 Lower die 4 Workpieces 4a, 4b Part 5 Air cylinder 6 High frequency current supply means 10 Main control means 11 Main memory means 12 Welding specification input/output means 13 Status output means 100 Learning device 101 Thickness measurement means 102 Appearance inspection means 103 Welding strength measurement means 104 Welding result evaluation means 105 Learning memory means 106 Welding condition learning means 107 I/O means 108 Welding condition reading means 109A Specification monitoring means 109B Status monitoring means 120 Communication line 130 Welding result measurement means

Claims (9)

A learning device that learns machining conditions,
At least, (1) a "processing result measuring means" for measuring the processing result;
(2) a "processing result evaluation means" for evaluating the processing result as a reward;
(3) a "learning storage means" for storing the processing result and a reward for evaluating the processing result;
(4) a "machining condition learning means" for organizing the learning results evaluated based on the machining results and enabling the machining conditions to be read out;
(5) a "processing specification monitoring means" for monitoring the processing specifications;
(6) a "machining condition reading means" for reading out the machining conditions from the learning storage means;
having
In advance, the processing result is measured by the processing result measuring means of (1), the processing result is evaluated as a reward by the processing result evaluation means of (2), the processing result and the reward are stored in the learning storage means of (3), and the learning result evaluated based on the processing result is organized by the processing condition learning means of (4), so that the processing conditions can be read out;
When the processing specifications are input to the processing means connected to the learning device,
The machining specifications are monitored by the machining specification monitoring means of (5), the machining conditions that satisfy the latest and current machining specifications are read from the learning storage means by the machining condition reading means of (6), and the read machining conditions are output to the machining means,
The machining condition learning means of (4) is configured to read out the machining conditions from the output of a trained model that has been trained to output machining conditions that satisfy at least a desired machining specification when the desired machining specification is input based on the information stored in the learning storage means of (3),
The processing conditions in (4) are processing conditions in a high-frequency welding device that applies a high-frequency current to two or more parts to weld them by dielectric heating, and include a supply time of the high-frequency current supplied to the workpieces,
The processing specifications of (5) include welding strength,
A device for learning processing conditions. 2. The learning device according to claim 1, wherein the processing conditions (4) further include an "anode current state" that indicates a transition of an anode current flowing through the high-frequency welding device. The learning device according to claim 1 is configured to learn machining conditions according to the variations in material, shape, and dimensions of the parts to be machined, and to read out machining conditions that satisfy the machining specifications. The learning device according to claim 1 is configured to determine whether machining conditions that satisfy the machining specifications are found from the learning storage means, and if machining conditions that satisfy the machining specifications are not found, to notify the user that update learning is necessary. The learning device according to claim 1 is configured to, when it is determined that no machining conditions satisfying the machining specifications are found in the learning storage means, update and learn machining conditions satisfying the machining specifications, and read out the machining conditions from the learning storage means after the update and learning. The learning device according to claim 1 further includes a "state monitoring means" for monitoring the state of energy applied to the workpiece by the processing device connected to the learning device,
The learning device of claim 1 is configured to learn the behavior and rewards in the energy state when the processing device starts processing, check the state when the processing device starts processing, and read out processing conditions that satisfy the processing specifications in that state. The learning device of claim 6, wherein the state monitoring means monitors the state of energy applied to the workpiece by the processing device after processing has begun, and when the state of the processing device becomes abnormal compared to the state assumed by the learning device, this is indicated and processing by the processing device is stopped. A processing device with a learning device, which is connected to any one of the learning devices according to claims 1 to 7 by a communication line and controlled as an integrated unit,
At least, (a) a "processing specification input/output means" for inputting processing specifications;
(b) a "main memory means";
(c) Having a "master control means";
storing the machining conditions read from the learning storage means of the connected learning device in the main storage means;
The main control means is configured to perform a machining operation using the stored machining conditions,
The processing device is a processing device with a learning device, the processing device being a "high frequency welding device." A processing system with a learning device, to which any one of the learning devices according to claims 1 to 7 is connected,
The processing means constituting the processing system include:
At least (i) a "processing specification input/output means" for inputting processing specifications;
(b) a "main memory means";
(c) Having a "master control means";
The machining conditions read from the learning storage means of the connected learning device are stored in the main storage means, and a machining operation using the stored machining conditions is performed by the main control means.
A processing system with a learning device, in which the processing device used in the processing system is a "high frequency welding device."