TWI840303B - Training method and system of model used for predicting characteristics of workpiece - Google Patents

Abstract

A training method and a system of model used for predicting characteristics of workpiece are provided. The training method comprises an image capture step, a conversion step, a formation step, a model building step, and an evaluation step. By recording various data of a workpiece during the manufacturing process and utilizing image processing to extract multiple feature values from the workpiece's images, these are integrated into a production record. This production record is used to predict the final characteristics of the workpiece, thereby enhancing the effectiveness of quality management for the workpiece.

Description

Method and system for training a model for predicting workpiece characteristics

The present invention relates to a training method and system, and more particularly to a training method and system for a model for predicting workpiece characteristics.

Selective Laser Melting uses high-power lasers to focus on the surface of metal powders through optics, causing the metal powders to melt instantly and then solidify instantly to form various special shapes. It is a key technology in the metal layer manufacturing process and also plays an important role in process yield control.

However, in the process of forming various special shapes of metal powder, the surface of the metal powder is in a molten state, and the center collides on the surface of the substrate in a supercooled liquid state, causing the molten metal powder to splash from the melt pool to the surrounding area, producing a thin diffuse splash-like part of the splash point overlap, which in turn affects the quality of the finished product. In the traditional process, if you want to know the quality of the finished product, you need to wait until the finished product is printed and then measure it with special instruments to know the quality of the finished product. This method is not only time-consuming and consumable, but also, if destructive testing is required, it will also cause damage to the finished product.

Therefore, in order to overcome the shortcomings and deficiencies in the prior art, the present invention is necessary to provide an improved training method for a model for predicting workpiece characteristics to solve the problems existing in the above-mentioned conventional technology.

The main purpose of the present invention is to provide a training method and system for a model for predicting workpiece characteristics, by recording various data of the workpiece during the manufacturing process and using image processing to obtain multiple feature values from the image of the workpiece to integrate them into a product history, so as to predict the characteristics of the finished product of the workpiece, thereby improving the quality management effect of the workpiece.

To achieve the above-mentioned purpose, the present invention provides a training method for a model for predicting workpiece characteristics. The training method includes an image capture step, an eigenvalue conversion step, a history formation step, a model building step, and an evaluation step. In the image capture step, a plurality of layer-by-layer images of the workpiece are obtained by using a photosensitive coupling element in a printing process of forming a workpiece through a machine. In the eigenvalue conversion step, a conversion module is used to perform calculations through a gray-level co-occurrence matrix to obtain a plurality of eigenvalues of the layer-by-layer images. The gray-level co-occurrence matrix is formulated as follows:

in is the gray-level symbiosis matrix, is the grayscale value of the reference pixel, is the grayscale value of the adjacent pixels, each layer of image The size is , is a column of the gray-level symbiosis matrix. is the column shift, is the row of the gray-level symbiosis matrix, is row displacement (please confirm whether the parameter name is correct); in the history forming step, an integration module is used to integrate the feature values and multiple process parameters of the workpiece in the printing process to form a product history of the workpiece; in the model building step, a machine learning module is used to process the product history of the workpiece to generate a training model; in the evaluation step, an evaluation module is used to evaluate whether the training model is an optimal model through a model evaluation index. If so, a result is reported, if not, the above steps are re-executed.

In an embodiment of the present invention, in the eigenvalue conversion step, an image grayscale value of the grayscale co-occurrence matrix is set to 256, the relative position of the pixels of the grayscale co-occurrence matrix is omnidirectional, and the distance between the pixels of the grayscale co-occurrence matrix is 1.

In one embodiment of the present invention, in the eigenvalue conversion step, the center of the gray level co-occurrence matrix is taken as the center of the circle, and the omnidirectional direction, the center of the circle direction, the center of the circle Direction and center of the circle direction.

In one embodiment of the present invention, in the eigenvalue conversion step, the eigenvalues of the gray-level symbiosis matrix include: calculating the gray-level symbiosis matrix Direction and direction of an image gray level average; calculate the gray level co-occurrence matrix Direction and direction; calculate the gray-level co-occurrence matrix of the image Direction and a gray level standard deviation of an image in a direction; a uniformity of a texture feature value of the gray level symbiosis matrix; a randomness and variability of the texture feature value of the gray level symbiosis matrix; a contrast of the image of the gray level symbiosis matrix; a fineness and coarseness of the texture feature value of the gray level symbiosis matrix; a variation intensity of a local part of the image of the gray level symbiosis matrix; and a homogeneity of the texture feature value of the gray level symbiosis matrix.

In one embodiment of the present invention, before the model building step, the method further includes a data preprocessing step, in which the extreme values of the product history of the workpiece are removed through the machine learning module.

In one embodiment of the present invention, in the data preprocessing step, the machine learning module uses a mutual information to perform feature selection on the product history of the workpiece to obtain the optimal feature subset, wherein the equation of the mutual information is:

in and is a random variable, and Respectively indicate and The marginal probability mass function of express and The joint probability mass function of .

In one embodiment of the present invention, in the data preprocessing step, a hyperparameter range of the machine learning module is set.

In one embodiment of the present invention, in the image capturing step, the workpieces of the layer-by-layer images include a plurality of tensile test bars and a plurality of annular test pieces, the tensile test bars are used to detect a tensile strength, and the annular test pieces are used to detect a magnetic permeability and an iron loss.

In one embodiment of the present invention, in the model building step, the machine learning module uses a gradient boosting decision tree for calculation.

To achieve the above-mentioned purpose, the present invention provides a training system for a model for predicting workpiece characteristics, the training system includes a photosensitive coupling element, a conversion module, an integration module, a machine learning module and an evaluation module, wherein the photosensitive coupling element is configured to obtain multiple layer-by-layer images of a workpiece in a printing process of forming a workpiece through a machine; the conversion module is configured to calculate through a gray-level co-occurrence matrix to obtain multiple eigenvalues of the layer-by-layer images, and the equation of the gray-level co-occurrence matrix is:

in is the gray-level symbiosis matrix, is the grayscale value of the reference pixel, is the grayscale value of the adjacent pixels, each layer of image The size is , is a column of the gray-level symbiosis matrix. is the column shift, is the row of the gray-level symbiosis matrix, The integration module is configured to integrate the feature values and multiple process parameters of the workpiece in the printing process to form a product history of the workpiece; the machine learning module is configured to process the product history of the workpiece to generate a training model; the evaluation module is configured to evaluate whether the training model is an optimal model through a model evaluation indicator and report a result.

As described above, the training method and system of the model for predicting workpiece characteristics of the present invention focuses on predicting the quality of the workpiece in the manufacturing process, mainly by recording various data of the workpiece in the manufacturing process and using image processing to obtain multiple feature values from the image of the workpiece to integrate into the product history, and then using machine learning to use these feature values to predict the characteristics of the finished product of the workpiece. In this way, the material properties of the workpiece can be predicted in the process stage, so that unqualified workpieces can be replaced early, saving time and cost, and reducing the use of metal powder. It can also avoid the use of destructive testing and maintain the integrity of the finished product.

In order to make the above and other purposes, features, and advantages of the present invention more clearly understood, the following will specifically cite the present invention and provide a detailed description with reference to the attached drawings. Furthermore, the directional terms mentioned in the present invention, such as up, down, top, bottom, front, back, left, right, inside, outside, side, periphery, center, horizontal, transverse, vertical, longitudinal, axial, radial, topmost or bottommost, etc., are only for reference to the directions of the attached drawings. Therefore, the directional terms used are used to explain and understand the present invention, but not to limit the present invention.

Please refer to FIG. 1 , which is a training system for a model for predicting workpiece characteristics according to an embodiment of the present invention. The training system includes a photosensitive coupling element 2, a conversion module 3, an integration module 4, a machine learning module 5 and an evaluation module 6, wherein a metal lamination manufacturing machine is used to form at least one workpiece. The present invention will describe in detail the detailed structure, assembly relationship and operation principle of each component below.

1 and 3, the photosensitive coupling element 2 is mounted on the machine, and the photosensitive coupling element 2 is configured to obtain a plurality of layer-by-layer images 101 of the workpiece during a printing process of the workpiece formed by the machine, wherein the layer-by-layer images 101 are images of each layer of the workpiece printed during the printing process. In this embodiment, the workpieces of the layer-by-layer images 101 include a plurality of tensile test bars 102 and a plurality of ring-shaped test pieces 103, wherein the tensile test bars 102 are used to detect a tensile strength, and the ring-shaped test pieces 103 are used to detect a magnetic permeability and a iron loss.

1 and 3 , the conversion module 3 is electrically connected to the photosensitive coupling element 2 , and the conversion module 3 is configured to calculate through a grayscale symbiosis matrix to obtain multiple eigenvalues of the layer-by-layer images 101 . The grayscale symbiosis matrix is:

in is the gray-level symbiosis matrix, is the grayscale value of the reference pixel, is the grayscale value of the adjacent pixels, each layer of image The size is , is a column of the gray-level symbiosis matrix. is the column-wise displacement between pixels, is the row of the gray-level symbiosis matrix, is the row-wise displacement between pixels.

Please refer to Figures 1 and 3. The integration module 4 is electrically connected to the conversion module 3, and the integration module 4 is configured to receive the feature values of the layer-by-layer images 101, and integrate the feature values and multiple process parameters of the workpiece in the printing process (such as oxygen concentration, laser power, laser scanning speed, line spacing and energy density) to form a product history of the workpiece.

Please refer to FIG. 1 , the machine learning module 5 is electrically connected to the integration module 4 , and the machine learning module 5 is configured to receive the product history of the workpiece and process the product history of the workpiece to generate a training model.

Continuing with reference to FIG. 1 , the evaluation module 6 is electrically connected to the machine learning module 5 . The evaluation module 6 is configured to evaluate whether the training model is an optimal model through a model evaluation indicator and report a result.

According to the above-mentioned design, the training system of the model for predicting the characteristics of the workpiece of the present invention forms the product history of the workpiece through the integrated module 4 to record and describe various data in the process of layer-by-layer manufacturing of the workpiece, wherein the various data are composed of a plurality of process parameters and the characteristic values of the layer-by-layer images 101 in the process. The process parameters are mainly input into the parameter settings of the machine, and the characteristic values are captured by installing the photosensitive coupling element 2 in the machine to capture the layer-by-layer images 101 in the process, and a series of image characteristic values are generated through image processing to describe the texture characteristics of the workpiece. Then, the machine learning module 5 is used to predict the material properties of the workpiece according to its product history, so that the material properties of the workpiece can be known during the process stage of multilayer manufacturing, thereby achieving the effect of improving the quality management of the workpiece.

As described above, the training system of the model for predicting workpiece characteristics of the present invention focuses on improving the prediction quality of the workpiece during the manufacturing process (such as tensile strength, magnetic permeability and iron loss), mainly by recording various data of the workpiece during the manufacturing process and using image processing to obtain multiple feature values from the image of the workpiece to integrate into the product history, and then using machine learning to use these feature values to predict the characteristics of the finished product of the workpiece. In this way, the material properties of the workpiece can be predicted at the process stage, so that unqualified workpieces can be replaced early, saving time and cost, and reducing the use of metal powder. It can also avoid the use of destructive testing and maintain the integrity of the finished product.

Please refer to FIG. 2 in conjunction with FIG. 1 and FIG. 3, which is a training method for a model for predicting workpiece characteristics according to an embodiment of the present invention. The training method is operated according to the training system of the above embodiment, wherein the training method includes an image capture step S201, a feature value conversion step S202, a history formation step S203, a data preprocessing step S204, a model building step S205 and an evaluation step S206. The present invention will explain in detail the relationship between each step and its operating principle below.

Continuing to refer to FIG. 2 in conjunction with FIG. 1 and FIG. 3, in the image capturing step S201, in a printing process of forming a workpiece by a machine, a plurality of layer-by-layer images 101 of the workpiece are obtained by using a photosensitive coupling element 2. In this embodiment, the workpiece of the layer-by-layer images 101 includes a plurality of tensile test bars 102 and a plurality of annular test pieces 103, wherein the tensile test bars 102 are used to detect a tensile strength, and the annular test pieces 103 are used to detect a magnetic permeability and a iron loss.

Continuing to refer to FIG. 2 and FIG. 1 and FIG. 2 , in the eigenvalue conversion step S202, a conversion module 3 is used to calculate through a gray-level co-occurrence matrix to obtain multiple eigenvalues of the layer-by-layer images 101. The gray-level co-occurrence matrix is:

in is the gray-level symbiosis matrix, is the grayscale value of the reference pixel, is the grayscale value of the adjacent pixels, each layer of image The size is , is a column of the gray-level symbiosis matrix. is the column-wise displacement between pixels, is the row of the gray-level symbiosis matrix, is the row-wise displacement between pixels.

In this embodiment, an image gray value of the gray-level symbiosis matrix is set to 256, or 0 to 255, the relative position of the pixels of the gray-level symbiosis matrix is omnidirectional, and the distance between the pixels of the gray-level symbiosis matrix is 1. For example, with the center of the gray-level symbiosis matrix as the center of the circle, the omnidirectional includes the center of the circle. direction( ), the center of the circle direction( ), the center of the circle direction( ) and the center of the circle direction( ).

Specifically, the eigenvalues of the gray-level co-occurrence matrix include:

(1) Calculate the gray-level symbiosis matrix The average gray level of a horizontal image in the direction and Vertical image gray level average , is the grayscale size used in calculating the grayscale co-occurrence matrix, where the average grayscale value of the horizontal image is And the vertical image gray level average The equation for is:

(2) Calculate the gray-level symbiosis matrix The grayscale variance of a horizontal image in the direction and A vertical image grayscale variance in the direction , where the grayscale variance of the horizontal image is and the vertical image grayscale variance The equation for is:

(3) Calculate the gray-level symbiosis matrix The standard deviation of the gray level of a horizontal image in the direction and The vertical image grayscale standard deviation , where the grayscale standard deviation of the horizontal image is And the vertical image grayscale standard deviation The equation for is:

(4) The energy of the texture eigenvalue of the gray-level symbiosis matrix , i.e. homogeneous patterns, where the energy The equation for is:

(5) Entropy of the texture eigenvalue of the gray-level symbiosis matrix , i.e., randomness and variability, where the entropy The equation for is:

(6) The contrast of the image in the gray-level co-occurrence matrix , where the contrast The equation for is:

(7) Autocorrelation of the texture eigenvalues of the gray-level co-occurrence matrix , i.e., the degree of fineness and coarseness, where the autocorrelation The equation for is:

(8) The local correlation of the gray-level co-occurrence matrix and heterogeneity , i.e. the intensity of change, the correlation and heterogeneity The equation for is:

(9) Homogeneity of the texture eigenvalues of the gray-level co-occurrence matrix , where the homogeneity The equation for is:

Continuing to refer to FIG. 2 and FIG. 1 , in the resume forming step S203, an integration module 4 is used to integrate the feature values and multiple process parameters of the workpiece during the printing process to form a product resume of the workpiece, wherein the format of the product resume is as follows, all the process parameters are integrated into one product resume. and texture features Place the input matrix, and each set of process parameters and texture features will correspond to a finished product feature .

Continuing to refer to FIG. 2 and FIG. 1 , in the data preprocessing step S204, the extreme values of the product history of the workpiece are removed by a machine learning module 5 to retain the most relevant feature values. In this embodiment, the machine learning module 5 uses a mutual information to perform feature selection on the product history of the workpiece to obtain the best feature subset, that is, the most relevant first k feature values, where k ranges from 5 to 18. Finally, only the most relevant feature values are retained, and the rest are removed. The equation for the mutual information is:

in and is a random variable, and Respectively indicate and The marginal probability mass function of express and The joint probability mass function of .

In addition, a hyperparameter range of the machine learning module 5 is set. In this embodiment, the machine learning module 5 uses random search to find the best hyperparameter. Random search is a hyperparameter optimization method and also a method to replace grid search. Unlike grid search, random search does not search for all possible hyperparameter combinations. Instead, it randomly selects some combinations from the range of hyperparameter values for model training and evaluation, and finally selects the hyperparameter combination with the best performance as the best combination. The search range of random search is: learning rate: [0.1, 0.15, 0.2, 0.25, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5]; maximum decision tree depth: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]; number of nodes: [100, 500, 1000]; and number of eigenvalues: 5~18.

2 and in conjunction with FIG. 1 , in the model building step S205 , the machine learning module 5 is used to process the product history of the workpiece to generate a training model; in this embodiment, the machine learning module 5 uses a gradient boosting decision tree for calculation.

Furthermore, by utilizing the feature selection, the first k feature values that are most relevant to the finished product characteristics are selected from the product history data of the workpiece as the training data set for the training model. The model is then trained using the XGBoost model, which can predict the tensile strength of the tensile test bars 102 and the magnetic permeability and iron loss of the annular test pieces 103, respectively.

2 and in conjunction with FIG. 1 , in the evaluation step S206 , an evaluation module 6 is used to evaluate whether the training model is an optimal model through a model evaluation index. If so, a result is reported; if not, the above steps are executed again.

Furthermore, the evaluation module 6 evaluates the training model through the determination coefficient ( ), mean square error (MSE) and mean absolute error (MAE) are used as the basis for judging the accuracy of the training model. Finally, it can be found that for the finished product characteristics of the workpiece, the results obtained by the model for predicting the workpiece characteristics generated by the training method of the present invention are very accurate, and the error of the finished product characteristics of the workpiece can be within about 10%.

According to the above content, the training method of the model for predicting the characteristics of the workpiece of the present invention forms the product history of the workpiece through the integrated module 4 to record and describe various data in the process of layer-by-layer manufacturing of the workpiece, wherein the various data are composed of a plurality of process parameters and the characteristic values of the layer-by-layer images 101 in the process. The process parameters are mainly input into the parameter settings of the machine, and the characteristic values are captured by installing the photosensitive coupling element 2 in the machine to capture the layer-by-layer images 101 in the process, and a series of image characteristic values are generated through image processing to describe the texture characteristics of the workpiece. Then, the machine learning module 5 is used to predict the material properties of the workpiece according to its product history, so that the material properties of the workpiece can be known during the process stage of multilayer manufacturing, thereby achieving the effect of improving the quality management of the workpiece.

As described above, the training method of the model for predicting workpiece characteristics of the present invention focuses on improving the prediction quality of the workpiece during the manufacturing process (such as tensile strength, magnetic permeability and iron loss), mainly by recording various data of the workpiece during the manufacturing process and using image processing to obtain multiple feature values from the image of the workpiece to integrate into the product history, and then using machine learning to use these feature values to predict the characteristics of the finished product of the workpiece. In this way, the material properties of the workpiece can be predicted in the process stage, so that unqualified workpieces can be replaced early, saving time and cost, and reducing the use of metal powder. It can also avoid the use of destructive testing and maintain the integrity of the finished product.

Although some aspects have been described in the context of a system, it should be understood that the aspects also represent a description of a corresponding method, and therefore, a block or structural element of the system should also be understood as a corresponding method step or a feature of a method step. By analogy, aspects that have been described in the context of a method step or as a method step also represent a description of a corresponding block or detail or feature of a corresponding device. Some or all of the method steps can be performed when using a hardware device such as a microprocessor, a programmable computer or an electronic circuit. In some embodiments, some or several of the most important method steps can be performed by such a device.

Depending on the specific implementation requirements, the embodiments of the present invention can be implemented in hardware or in software. The implementation can be implemented when using a digital storage medium, such as a floppy disk drive, DVD, Blu-ray disk drive, CD, ROM, PROM, EPROM, EEPROM or FLASH memory, a hard disk or any other magnetic or optical storage device, which has an electronically readable control signal stored therein, which can cooperate with a programmable computer system, or collaborate, so that the corresponding method is executed. This is why the digital storage medium can be computer readable. Therefore, some embodiments according to the present invention include a data carrier, which includes an electronically readable control signal that can cooperate with a programmable computer system to execute any method described herein. Generally, embodiments of the present invention can be implemented as a computer program product with program code, and when the computer program product is run on a computer, the program code can effectively execute any method. For example, the program code can also be stored on a machine-readable carrier. Other embodiments include a computer program for executing any method described herein, which is stored on a machine-readable carrier. In other words, one embodiment of the method of the present invention is a computer program having a program code for executing any method described herein, when the computer program is run on a computer. Therefore, another embodiment of the method of the present invention is a data carrier (or digital storage medium or computer-readable medium) in which a computer program for executing any method described herein is recorded. The data carrier, digital storage medium or recording medium is generally tangible or non-volatile. Therefore, another embodiment of the method of the present invention is a data stream or a signal sequence representing a computer program for executing any method described herein. The data stream or signal sequence can be configured to be transmitted, for example, via a data communication link, such as via a network. Further embodiments include a processing unit, such as a computer or a programmable logic device, configured to or suitable for executing any method described herein. Further embodiments include a computer on which a computer program for executing any method described herein is installed.

Another embodiment according to the present invention includes a device or system configured to transmit a computer program for performing at least one of the methods described herein to a receiver. For example, the transmission can be electrical or optical. For example, the receiver can be a computer, a mobile device, a storage device, or a similar device. For example, the device or system can include a file server for transmitting the computer program to the receiver. In some embodiments, a programmable logic device (e.g., a field programmable gate array, FPGA) can be used to perform some or all functions of the methods described herein. In some embodiments, the field programmable gate array can cooperate with a microprocessor to perform any method described herein. Generally, in some embodiments, these methods are performed by any hardware device. The hardware device can be any general hardware such as a computer processor (CPU), or method-specific hardware such as an ASIC.

Although the present invention has been disclosed by way of embodiments, they are not intended to limit the present invention. Any person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of the attached patent application.

101: Layer-by-layer image 102: Tensile test rod 103: Ring specimen 2: Photosensitive coupling element 3: Conversion module 4: Integration module 5: Machine learning module 6: Evaluation module S201: Image acquisition step S202: Eigenvalue conversion step S203: Resume formation step S204: Data preprocessing step S205: Model building step S206: Evaluation step

FIG1 is a schematic diagram of a training system for a model for predicting workpiece characteristics according to an embodiment of the present invention. FIG2 is a flow chart of a training method for a model for predicting workpiece characteristics according to an embodiment of the present invention. FIG3 is a schematic diagram of layer-by-layer images of a training method for a model for predicting workpiece characteristics according to an embodiment of the present invention.

S201: Image capture step

S202: Eigenvalue conversion step

S203: Resume formation step

S204: Data preprocessing step

S205: Model building step

S206: Evaluation step

Claims (10)

A training method for a model for predicting workpiece characteristics includes: an image capture step, in which a plurality of layer-by-layer images of a workpiece are obtained by using a photosensitive coupling element during a printing process of forming a workpiece by a machine; an eigenvalue conversion step, in which a conversion module is used to calculate through a gray-level co-occurrence matrix to obtain a plurality of eigenvalues of the layer-by-layer images, wherein the gray-level co-occurrence matrix is: in is the gray-level symbiosis matrix, is the grayscale value of the reference pixel, is the grayscale value of the adjacent pixels, each layer of image The size is , is a column of the gray-level symbiosis matrix. is the column shift, is the row of the gray-level symbiosis matrix, for row displacement; a history forming step, utilizing an integration module to integrate the feature values and a plurality of process parameters of the workpiece in the printing process to form a product history of the workpiece; a model building step, utilizing a machine learning module to process the product history of the workpiece to generate a training model; and an evaluation step, utilizing an evaluation module to evaluate whether the training model is an optimal model through a model evaluation indicator, and if so, reporting a result, and if not, re-executing the above steps. In the training method of the model for predicting workpiece characteristics as described in claim 1, in the eigenvalue conversion step, an image grayscale value of the grayscale co-occurrence matrix is set to 256, the relative position of the pixels of the grayscale co-occurrence matrix is omnidirectional, and the distance between the pixels of the grayscale co-occurrence matrix is 1. In the training method for predicting workpiece characteristics as described in claim 2, in the eigenvalue conversion step, the center of the gray-level co-occurrence matrix is taken as the center of the circle, and the omnidirectional direction, the center of the circle direction, the center of the circle Direction and center of the circle direction. In the training method for predicting workpiece characteristics of claim 1, in the eigenvalue conversion step, the eigenvalues of the gray-level symbiosis matrix include: calculating the gray-level symbiosis matrix Direction and The gray level average of an image in the direction; Calculate the gray level co-occurrence matrix Direction and The gray-level variance of an image in the direction of Direction and a grayscale standard deviation of an image in a direction; a uniformity of the texture eigenvalues of the grayscale symbiosis matrix; a randomness and variability of the texture eigenvalues of the grayscale symbiosis matrix; a contrast of the image of the grayscale symbiosis matrix; a fineness and coarseness of the texture eigenvalues of the grayscale symbiosis matrix; a variation intensity of a local part of the image of the grayscale symbiosis matrix; and a homogeneity of the texture eigenvalues of the grayscale symbiosis matrix. The method for training a model for predicting workpiece characteristics as described in claim 1 further includes a data preprocessing step before the model building step, in which the extreme values of the product history of the workpiece are removed through the machine learning module. In the training method for a model for predicting workpiece characteristics as described in claim 5, in the data preprocessing step, the machine learning module uses a mutual information to perform feature selection on the product history of the workpiece to obtain the optimal feature subset, wherein the equation of the mutual information is: in and is a random variable, and Respectively indicate and The marginal probability mass function of express and The joint probability mass function of . In the method for training a model for predicting workpiece characteristics as described in claim 4, in the data preprocessing step, a hyperparameter range of the machine learning module is set. As described in claim 1, the training method of a model for predicting workpiece characteristics, in the image capture step, the workpiece of the layer-by-layer images includes a plurality of tensile test bars and a plurality of annular test pieces, the tensile test bars are used to detect a tensile strength, and the annular test pieces are used to detect a magnetic permeability and an iron loss. In the method for training a model for predicting workpiece characteristics as described in claim 1, in the model building step, the machine learning module uses a gradient boosting decision tree for calculation. A training system for a model for predicting workpiece characteristics, the training system comprising: a photosensitive coupling element, configured to obtain multiple layer-by-layer images of a workpiece during a printing process of forming a workpiece through a machine; a conversion module, configured to calculate through a gray-level co-occurrence matrix to obtain multiple eigenvalues of the layer-by-layer images, the gray-level co-occurrence matrix having the equation: in is the gray-level symbiosis matrix, is the grayscale value of the reference pixel, is the grayscale value of the adjacent pixels, each layer of image The size is , is a column of the gray-level symbiosis matrix. is the column shift, is the row of the gray-level symbiosis matrix, for row displacement; an integration module configured to integrate the feature values and multiple process parameters of the workpiece in the printing process to form a product history of the workpiece; a machine learning module configured to process the product history of the workpiece to generate a training model; and an evaluation module configured to evaluate whether the training model is an optimal model through a model evaluation indicator and report a result.