TWI809595B - A Hierarchical Parameter System for Predicting Machine Abnormalities - Google Patents

Abstract

The present invention provides a hierarchical parameter system for predicting machine abnormalities. The system includes: a server connected to a host in the factory; an integrated module connected to the server and a learning module connected to the integrated module; and The prediction module connected to the learning module uses the server to collect various parameters of the machine in operation, and uses the integration module to sort the parameters of different time periods in the past into sequences of different unit times according to the time category, making It is the "machine data", and marks the "past failure and abnormal record" of the machine in operation, and then integrates it into a one-dimensional vector. Simulate the mapping relationship between "machine data" and "past failure abnormal records" to judge the possibility of machine failure in operation, and apply the "machine data" judged to be a failure as "prediction parameters" to the machine in operation , and can correctly provide an abnormal warning forecast for the operation status of the machine, so that the factory personnel can arrange maintenance or adjust the machine in the production line in advance, avoid temporary shutdowns and reduce factory losses, and reduce factory employees and production Make better arrangements for production line machines and production materials, and increase the overall efficiency of the factory.

Description

A Hierarchical Parameter System for Predicting Machine Abnormalities

The present invention relates to a machine abnormality prediction system, in particular to a hierarchical parameter system for predicting machine abnormalities that can properly input parameters of different time scales to obtain more accurate prediction results.

With the digitalization and evolution of factory machines and the development of machine learning algorithms, it is more important to construct parameters to predict the health status of machines (normal, abnormal, and failure) for maintenance planning. The parameter prediction of machine abnormalities in traditional factories can only be predicted using the latest specific period of time, for example: the data of the past 60 seconds, so it is impossible to detect the parameter sequence characteristics (patten) of different scales, that is, the data of the past few days can be used to predict There is often a big difference between making predictions and using the data of the past few months or even years, and with the use of time, the performance difference of the machine will tend to increase day by day, so using fixed parameter values cannot effectively reflect the actual situation. State, in other words, the parameter prediction of machine abnormalities in traditional factories cannot be predicted in the medium and long term, and when making parameter predictions, if the data of different time scales such as the last month, the last year, and the last 5 to 10 years are combined When the time scale is too large, it often causes many difficulties in prediction and incorrect prediction results.

According to supervised learning in machine learning is a method to solve this problem, the application of this method usually requires the labeling of a large amount of data, and then use the method of neural network classifier, according to the method of maximum likelihood (Maximum Likelihood) The neural network classifier is trained to learn the mapping relationship (mapping) between machine data and fault exceptions. However, this type of method usually requires a lot of manpower to judge and mark whether the machine data is abnormal at the beginning. In addition, as mentioned above, the aging of the machine will also cause the value to be different. Therefore, how to Constructing machine parameter data at different time periods to predict the abnormal state of the machine is indeed a subject to be broken through.

The inventor of this case originally engaged in related work experience in the factory for many years, combined factory experience, network and machine learning design and learned from mistakes, and created the present invention.

The purpose of the present invention is to provide a hierarchical parameter system for predicting machine abnormality. The system collects various parameters of the machine in operation, and sorts the parameters of different time periods in the past into sequences of different unit time according to the time category. As "machine data", and mark (label) the "past failure record" of the machine in operation, and then integrate it into a one-dimensional vector. After concatenation, the said "machine Simulate the mapping relationship (mapping) between "data" and "abnormal record of past faults" to judge whether the machine in operation is faulty, and apply the "machine data" judged to be faulty as "prediction parameters" to the machine in operation , and can correctly provide an abnormal early warning forecast for the operation status of the machine, so that factory personnel can arrange maintenance or adjust the machine in the production line in advance, avoiding temporary shutdowns and reducing factory losses.

In order to achieve the above purpose, the present invention provides a hierarchical parameter system for predicting machine abnormalities. The system includes: a server, which is connected to the main machine in the factory or the machinery and equipment of the factory production line, and is used to collect and store the operating machines in the factory. According to the time category, the parameters of different time periods in the past are sorted into sequences of different unit time, which are used as "machine data", and the "past failure and abnormal records" of the running machine are marked (label); an integrated model The group is connected to the server and includes a time series encoder (Time Series Encoder) and a decoder (Decoder). The time series encoder constructs a variable-length time from the "machine data" in a one-dimensional vector Sequence (hereinafter referred to as variable time series), extracting the variable time series into a fixed-length one-dimensional vector (hereinafter referred to as fixed vector); the decoder converts the fixed vector into a variable-length target signal through calculation Sequence (hereinafter referred to as the target signal sequence) output; and a learning module connected to the integrated module, including a neural network classifier (DNN classifier), which trains the target signal sequence to obtain the machine Abnormal target probability; and a prediction module connected with the learning module, using the minimization model to optimize the target probability and the "past fault abnormal record" to obtain the prediction parameters of machine abnormality.

According to the above, the time series encoder is given a time series with a length of time from the "machine data" as the original data, such as: the time series Xsec in seconds in the past 60 seconds, and minutes in the past 30 minutes The time series Xmin in unit, and the time series Xhr in units of 30 minutes in the past 12 hours, and then use the parameters of temperature, humidity, pressure or factory machine as the original data parameter value of each dimension of the one-dimensional vector, and output mutable time series of the 1D vector.

According to the above, the decoder converts the fixed vector into a variable-length target signal sequence through the following three-step algorithm:

1. Calculate attention weights

2. Standardized processing to calculate the attention output value

3. Combining the attention score and the hidden state value to calculate the context state C

According to the above, wherein the learning module trains the target signal sequence with an algorithm, the algorithm uses a backpropagation algorithm, and uses a stochastic gradient descent method to solve the weight update in the backpropagation algorithm, and the inverse The propagation algorithm is as follows:

(wherein, η is the learning rate, and C is the cost function)

The C cost function uses cross-entropy; defined as (where dj represents the target probability of the output unit j, and pj represents the probability output of the unit j after the activation function is applied), and the activation function uses a rectified linear unit (ReLU) as the activation function of the neuron.

According to the above, the forecasting module optimizes the model through the cross entropy loss of the target probability and the "abnormal past fault record".

Please refer to FIG. 1, which is a structural diagram of the integrated module of the hierarchical parameter system for predicting machine abnormality in the present invention. As shown in the figure, the hierarchical parameter system 1 for predicting machine abnormality in the present invention includes: a server 2, an integrated module Group 3, learning module 4, and forecasting module 5, wherein the server 2 is connected to the host computer 6 in the factory, and is used to collect and store various parameters of the machine in operation in the factory, and to record the past data according to the time category. The parameters of the time period are organized into sequences of different unit times, which are used as "machine data", and the "past failure and abnormal records" of the machines in operation are marked.

The integration module 3 is connected to the server 2. The integration module 3 includes a time series encoder 31 (Time Series Encoder) and a decoder 32 (Decoder) as shown in FIG. 2. In a time series [x1,...,xt] with a given length t, each item xi is raw data, and each dimension is a parameter value (e.g., temperature, humidity, pressure or Factory machine parameters), the goal is to output a one-dimensional vector, as this one-dimensional vector represents the time series, convert xi to hi, and then use the attention (attention; see later) mechanism to calculate the weight (weight) αi, and use αi to carry out the weighted sum of hi to obtain the average vector S (serires emebdding s) of the time series embedding.

In other words, the time series encoder 31 constructs a variable-length time series (hereinafter referred to as variable time series) from the "machine data" with a one-dimensional vector, extracts the variable time series and converts it into a fixed-length A one-dimensional vector (hereinafter referred to as a fixed vector); the decoder converts the fixed vector into a variable-length target signal sequence (hereinafter referred to as a target signal sequence) output through calculation. In other words, the time series encoder 31 is from the Given a time series of time length t in the above "machine data" as the original data, such as: the time series Xsec of the past 60 seconds in seconds, the time series Xmin of the past 30 minutes in minutes, and the past 12 The hour is the time series Xhr with 30 minutes as the unit (refer to FIG. 4 ), and then the temperature, humidity or pressure is used as the raw data parameter value of each dimension of the one-dimensional vector, and the variable time series of the one-dimensional vector is output.

The decoder 32 converts the fixed vector into a variable-length target signal sequence through the following 3-step calculation formula: 1. Calculate the attention weight; 2. Calculate the attention output value through standardization processing; 3. Combine the attention score and The hidden state value calculates the context state C, and the relative calculation formula is as follows:

The learning module 4 is connected with the integration module 3, as shown in Fig. 4 , it includes a neural network classifier 41 (DNN classifier), which trains the target signal sequence to obtain the target probability of machine abnormality. The learning module 4 trains the target signal sequence with an algorithm, the algorithm adopts a backpropagation algorithm, and solves the weight update in the backpropagation algorithm with a stochastic gradient descent method, and the backpropagation algorithm is as follows:

(wherein, η is the learning rate, and C is the cost function)

The C cost function uses cross-entropy; defined as (where dj represents the target probability of the output unit j, and pj represents the probability output of the unit j after the activation function is applied), and the activation function uses a rectified linear unit (ReLU) as the activation function of the neuron. The prediction module 5 is connected with the learning module 4, and uses the minimized model prediction result (y ^pred ) to optimize the model of the target probability and the "past failure abnormal record", and obtains the prediction parameters of machine abnormality , that is, optimize the model through the cross entropy loss of the target probability and the "past fault anomaly record".

If so, please look back at FIG. 1 and observe it in conjunction with FIG. 4. The hierarchical parameter system for predicting machine abnormalities of the present invention uses the server 2 to collect various parameters of the machine in operation, and uses the time series of the integration module 3 The encoder 31 organizes the parameters of different time periods in the past into sequences of different unit times according to the time category (as described in the previous paragraph), as "machine data", and labels the "past failure and abnormal records" of the machine in operation , and then integrated into a one-dimensional vector and concatenated, through the learning module 4 and the prediction module 5 to maximize the likelihood (Maximum Likelihood), the relationship between the "machine data" and the "past failure record" The mapping relationship (mapping) is simulated to determine whether the machine in operation is faulty, and the "machine data" judged to be faulty is applied to the machine in operation as a "prediction parameter", so that the correct information on the machine can be provided. The operation status provides abnormal early warning predictions, allowing factory personnel to arrange maintenance or adjust production line machines in advance, avoiding temporary shutdowns and reducing factory losses.

In summary, the hierarchical parameter system for predicting machine abnormalities in the present invention can indeed achieve the purpose of the invention and meet the requirements of the patent. However, the above-mentioned ones are only preferred embodiments of the present invention. Various modifications and changes should still be included within the scope of this patent application.

1: Hierarchical parameter system for predicting machine abnormalities 2: Server 3: Integrated modules 4: Learning modules 5: Prediction module 6: Host in the factory 31: Time Series Encoder 32: Decoder xi: original data hi: time series αi: weight

FIG. 1 is a structure diagram of a hierarchical parameter system for predicting machine abnormality according to the present invention. FIG. 2 is a structural diagram of an integrated module of a hierarchical parameter system for predicting machine abnormality according to the present invention. FIG. 3 is a structure diagram of the learning module of the hierarchical parameter system for predicting machine abnormality according to the present invention. FIG. 4 is a structural diagram of the prediction module of the hierarchical parameter system for predicting machine abnormality according to the present invention.

1: Hierarchical parameter system for predicting machine abnormality 2: Server 3: Integrated modules 4: Learning modules 5: Prediction module 6: Host in the factory

Claims (4)

A hierarchical parameter system for predicting machine abnormality, the system includes: a server, which is connected to the main machine in the factory or the factory production line equipment, used to collect and store various parameters of the machine in operation in the factory, and according to the time The category organizes the parameters of different time periods in the past into sequences of different unit times, as "machine data", and labels the "past failure and abnormal records" of the running machine; an integrated module is connected to the servo The device includes a time series encoder (Time Series Encoder) and a decoder (Decoder). The time series encoder constructs a variable-length time series (hereinafter referred to as variable time series) with a one-dimensional vector in the "machine data". sequence), extracting the variable time sequence into a fixed-length one-dimensional vector (hereinafter referred to as a fixed vector); the decoder converts the fixed vector into a variable-length target signal sequence (hereinafter referred to as a target signal sequence) through calculation ) output; and a learning module connected to the integration module, including a neural network classifier (DNN classifier), which trains the target signal sequence to obtain the target probability of machine abnormality; and a The prediction module is connected with the learning module, and uses the minimization model to optimize the target probability and the "abnormal record of past faults" to obtain the prediction parameters of machine abnormalities; wherein: the time series encoder A time series with a given time length from the "machine data" is used as the original data, such as: the time series Xsec in seconds in the past 60 seconds, the time series Xmin in minutes in the past 30 minutes, and the past The 12-hour time series Xhr with a unit of 30 minutes is used, and then the temperature, humidity or pressure is used as the raw data parameter value of each dimension of the one-dimensional vector, and the variable time series of the one-dimensional vector is output. A hierarchical parameter system for predicting machine abnormality as in claim item 1, wherein the decoder converts the fixed vector into a variable-length target signal sequence through the following three-step calculation formula: 1. Calculating attention weights; 2. Calculate the attention output value through standardized processing; 3. Calculate the context state C by combining the attention score and the hidden state value, as follows:

Attention weights

Attention output y ₂ = f ( c _t , h _t )=tanh( W _c [ c _t : h _t ]) Final output A hierarchical parameter system for predicting machine abnormality as in claim item 1, wherein the learning module uses an algorithm to train the target signal sequence, the algorithm uses a backpropagation algorithm, and weights in the backpropagation algorithm The update is solved by the stochastic gradient descent method, and the backpropagation algorithm is as follows:

(wherein, n is the learning rate, and C is the cost function) The C cost function uses cross entropy; defined as

(where dj represents the target probability of the output unit j, and pj represents the probability output of the unit j after the activation function is applied), and the activation function uses a rectified linear unit (ReLU) as the activation function of the neuron. A hierarchical parameter system for predicting machine abnormality as in Claim 1, wherein the prediction module optimizes the model through the cross entropy loss of the target probability and the "past fault abnormal record".

Images