Model training method, workpiece temperature prediction method, system, device and medium

Abstract

The application discloses a model training method, a workpiece temperature prediction method, a system, equipment and a medium. The training method comprises the following steps: obtaining a historical training set of a flattening process, wherein the historical training set comprises induction heating parameters, physical parameters, front surface temperatures and non-front surface temperatures of standard workpieces at the same time; inputting the induction heating parameters, the physical parameters and the front surface temperatures into a workpiece temperature prediction model to obtain predicted non-front surface temperatures of the standard workpieces; calculating a loss value according to the predicted non-front surface temperatures and the obtained non-front surface temperatures, and adjusting parameters of the workpiece temperature prediction model according to the loss value until a convergence condition is met. The workpiece temperature prediction model can predict internal temperatures and back surface temperatures of the workpiece in real time according to the front surface temperatures, the induction heating parameters and the physical parameters of the workpiece in the flattening process, greatly improves the controllability of flattening, and realizes precise flattening.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-frequency flattening, in particular to a model training method, a workpiece temperature prediction method, a system, a device and a medium. BACKGROUND

[0002] High-frequency flattening is a process method for controlling structural deformation, which can improve flatness and reduce fire work load. In recent years, it has been applied to ship section construction, total section and loading stage. Its working principle is: the induced magnetic field generated by the power coil acts on the workpiece to be straightened, then a strong induced current, i.e. eddy current, is generated in the workpiece to be straightened, the energy generated by the eddy current is converted into heat energy, so that the workpiece to be straightened is heated to eliminate the internal stress of the heated area of the workpiece to be straightened, and at the same time, when the heated area is cooled, the surrounding uniformly shrinks, so that the heated area becomes flat.

[0003] The high-frequency flattening process has a high demand for workpiece temperature. The side of the workpiece to be straightened heated directly by the induction coil is regarded as the front side of the workpiece to be straightened, and the front side temperature, the back side temperature and the internal temperature of the workpiece to be straightened need to be monitored in real time according to the actual process standard to control the heating depth, so as to accurately perform the high-frequency flattening process. However, since electromagnetic induction heating is a complex nonlinear process with large delay, the prior art can only use some temperature measuring devices to measure the workpiece to be straightened, such as installing an infrared temperature sensor on the induction coil. However, this method can only measure the front side temperature of the workpiece to be straightened, and the internal temperature and the back side temperature of the workpiece to be straightened cannot be monitored during the heating process. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the defect that the workpiece back side temperature and internal temperature are difficult to monitor during the flattening process in the prior art, and to provide a model training method, a workpiece temperature prediction method, a system, a device and a medium.

[0005] The present application solves the above technical problems by the following technical solutions:

[0006] The first aspect of the present application provides a model training method, which comprises:

[0007] obtaining a historical training set of the flattening process, wherein the historical training set comprises the induction heating parameters, the physical parameters, the front side temperature and the non-front side temperature of the standard workpiece at the same time;

[0008] inputting the induction heating parameters, the physical parameters and the front side temperature into a workpiece temperature prediction model to obtain the predicted non-front side temperature of the standard workpiece;

[0009] The loss value is calculated according to the predicted non-front temperature and the obtained non-front temperature, and the parameters of the workpiece temperature prediction model are adjusted according to the loss value until a convergence condition is met.

[0010] Preferably, the induction heating parameters include heating power, induction magnetic field frequency and / or heating time.

[0011] Preferably, the physical property parameters include workpiece material properties, workpiece specifications and / or workpiece structures.

[0012] Preferably, the non-front temperature includes back temperature and / or internal temperature.

[0013] Preferably, the model training method further comprises:

[0014] The trained workpiece temperature prediction model is subjected to lightweight processing.

[0015] The second aspect of the present application provides a workpiece temperature prediction method, and the steps of the workpiece temperature prediction method comprise:

[0016] Obtaining induction heating parameters, physical property parameters and front temperature of a workpiece to be flattened in a flattening process;

[0017] Inputting the induction heating parameters, physical property parameters and front temperature of the workpiece to be flattened into a workpiece temperature prediction model to obtain predicted non-front temperature of the workpiece to be flattened;

[0018] The workpiece temperature prediction model is obtained by the model training method of the present application.

[0019] The third aspect of the present application provides a model training system, and the model training system comprises a first obtaining module, a first prediction module and a training module:

[0020] The first obtaining module is used to obtain a historical training set of a flattening process, wherein the historical training set comprises induction heating parameters, physical property parameters, front temperature and non-front temperature of a standard workpiece at the same time;

[0021] The first prediction module is used to input the induction heating parameters, the physical property parameters and the front temperature into a workpiece temperature prediction model to obtain predicted non-front temperature of the standard workpiece;

[0022] The training module is used to calculate a loss value according to the predicted non-front temperature and the obtained non-front temperature, and adjust parameters of the workpiece temperature prediction model according to the loss value until a convergence condition is met.

[0023] The fourth aspect of the present application provides a workpiece temperature prediction system, and the workpiece temperature prediction system comprises a second obtaining module and a second prediction module:

[0024] The second acquisition module is configured to acquire the induction heating parameters, physical parameters, and front surface temperature of the workpiece to be leveled in a leveling process.

[0025] The second prediction module is configured to input the induction heating parameters, physical parameters, and front surface temperature of the workpiece to be leveled into a workpiece temperature prediction model to obtain a predicted non-front surface temperature of the workpiece to be leveled.

[0026] The workpiece temperature prediction model is obtained by the model training system.

[0027] A fifth aspect of the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the model training method or the workpiece temperature prediction method.

[0028] A sixth aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the model training method or the workpiece temperature prediction method.

[0029] On the basis of common sense in the art, the above-mentioned preferred conditions can be combined arbitrarily to obtain preferred examples of the present application.

[0030] The positive progress effect of the present application is that:

[0031] The workpiece temperature prediction model is trained by selecting a suitable training set, specifically, the induction heating parameters, physical parameters, and front surface temperature of the standard workpiece at each time point in the leveling process are input into the workpiece temperature prediction model to obtain a predicted non-front surface temperature of the standard workpiece at the corresponding time point; and the model parameters of the workpiece temperature prediction model are trained according to the predicted non-front surface temperature and the acquired non-front surface temperature, and the prediction result of the model is more accurate. The workpiece temperature prediction model of the present application can predict the internal temperature and back surface temperature of the workpiece in real time according to the front surface temperature, induction heating parameters, and physical parameters of the workpiece in the leveling process, greatly improving the controllability of leveling and realizing precise leveling. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The figure is a flowchart of the model training method of embodiment 1 of the present application.

[0033] Figure 2 The figure is a flowchart of the workpiece temperature prediction method of embodiment 2 of the present application.

[0034] Figure 3 The figure is a structural diagram of the model training system of embodiment 3 of the present application.

[0035] Figure 4A structural schematic diagram of a workpiece temperature prediction system according to Embodiment 4 of the present application.

[0036] Figure 5 A structural schematic diagram of an electronic device according to Embodiment 5 of the present application. DETAILED DESCRIPTION

[0037] The present application will be further described by way of examples, but the present application is not limited to the examples.

[0038] Embodiment 1

[0039] The present embodiment provides a model training method, as shown in the figure, the model training method comprises: Figure 1

[0040] S1, obtaining a historical training set of a flattening process, wherein the historical training set comprises induction heating parameters, physical property parameters, front surface temperature and non-front surface temperature of a standard workpiece at the same time.

[0041] In the present embodiment, the induction heating parameters can include heating power, induction magnetic field frequency and / or heating time. As an optional implementation, the heating power, induction magnetic field frequency and heating time can be read by the induction heating device in the flattening process, and the induction heating parameters are usually sequence data preset by the induction heating device.

[0042] In the present embodiment, the model training method can be applied in a high-frequency flattening device, which includes an induction heating device and a flattening device. The flattening process refers to heating the workpiece by using the electromagnetic heating device to make the workpiece generate an induced current flow to heat, and then using the flattening device to perform flattening work according to the workpiece temperature. Flattening includes various modes, such as deep penetration heating mode and surface heating mode. Among them, the deep penetration heating mode refers to heating the front surface of the workpiece and stabilizing the front surface temperature at the Curie temperature, and maintaining for a period of time, and the heat is transferred to the bottom surface of the workpiece through heat conduction to realize temperature penetration and achieve flattening effect. The surface heating mode refers to heating the front surface of the workpiece, controlling the heating depth, and ensuring that the front surface temperature of the workpiece is greater than the back surface temperature of the workpiece to achieve the flattening effect. For different modes and different workpieces, the electromagnetic heating device will also adjust the corresponding parameters over time during the flattening.

[0043] ​In the embodiment, the physical property parameters can include workpiece material properties, workpiece specifications and / or workpiece structures. As an optional implementation, the workpiece material properties, workpiece specifications and workpiece structures can be read from design drawings. Since the physical property parameters of the workpiece are fixed parameters that do not change over time, they need to be processed as time-varying sequence data. The workpiece material properties include carbon steel, alloy steel, cast iron and various alloys, etc. The workpiece structure is usually plate-shaped, i.e., the workpiece specifications are the length, width and thickness of the workpiece. When the workpiece has interfaces on both sides, the number of interfaces is further obtained. As an optional implementation, when the workpiece structure has interfaces, the material properties and specifications of the interfaces also need to be further obtained.

[0044] In the embodiment, the heating position of the electromagnetic heating device is usually at the center of the front surface of the workpiece. During the flattening process, the temperature gradually spreads from the center of the front surface to the periphery over time, and different positions on the front surface have different temperatures. As an optional implementation, the temperatures of different points in a certain area on the front surface are obtained at each time, and the positions and temperatures of the corresponding points are recorded, such as taking the center of the front surface as the origin, and obtaining the temperature distribution of the point positions in a certain range, such as 1 cm, 2 cm, 3 cm, etc., away from the center of the front surface. It should be noted that the distances between two points can be the same or different, such as obtaining the temperature distribution of the point positions in a certain range, such as 1 cm, 3 cm, 7 cm, etc., away from the center of the front surface.

[0045] As an optional implementation, the temperature data of the workpiece is collected by some temperature collection devices. For example, an infrared temperature sensor is installed at the heating position of the electromagnetic induction heating device, and the radiation heat effect during the heating of the workpiece is used to collect the temperature of the front surface of the standard workpiece, so that the detected temperature data is particularly accurate. In addition, when the electromagnetic induction heating device is working, the front surface temperature of the workpiece can be automatically obtained by the infrared temperature sensor installed on the electromagnetic induction heating device. Of course, the embodiment is not limited to using an infrared temperature sensor to collect the temperature.

[0046] In the embodiment, the non-front surface temperature can include the back surface temperature and / or the internal temperature. The back surface temperature refers to the temperature sequence data of the corresponding position of the heating position of the electromagnetic induction heating device on the back surface of the workpiece. The internal temperature refers to the temperature sequence data of the corresponding position of the heating position of the electromagnetic induction heating device at a certain depth from the front surface of the workpiece. The depth can be one-third to two-thirds of the distance from the front surface.

[0047] As an optional implementation, the back surface temperature of the workpiece can be collected by an infrared temperature sensor or a handheld temperature gun arranged on the back surface of the standard workpiece. The internal temperature of the workpiece can be collected by embedding a thermocouple temperature sensor in the standard workpiece.

[0048] As an optional implementation, the pandas library (a data processing library) in python (a computer programming language) is used to clean the obtained raw data. Specifically, the isnull function (a computer function representing judging whether the data contains null values) is used to find the missing values, and then the ffill function (a computer function representing backward filling) is used to fill the missing values longitudinally with the previous value, or the bfill function (a computer function representing forward filling) is used to fill the missing values longitudinally with the next value; the Z-Score method is used to judge the outliers, and a threshold M is set, when the Z-Score of the data is greater than the set threshold M, it is judged as an outlier, and then the outliers are deleted or replaced; the duplicated function (a computer function representing finding and displaying repeated values in the data) is used to judge the repeated values, and then the repeated values are deleted; finally, the rolling function (a time window function) is used to split the data according to a fixed time step to obtain a plurality of training sample data.

[0049] As an optional implementation, the principal component analysis (PCA) is used to extract the data features of the training sample data, so as to filter out the main data. First, the training sample data is linearly transformed by linear function normalization, so that the result is mapped to the range of [0, 1], realizing the equal proportion scaling of the training sample data; second, the covariance matrix is calculated to identify the correlation; then the eigenvectors and eigenvalues of the covariance matrix are calculated to identify the principal components of the data; then the eigenvectors are calculated and sorted in descending order of eigenvalues, so as to find the principal components in order of importance; finally, the transpose of the training sample data is multiplied by the transpose of the eigenvectors to reorient the data from the original axis to the axis represented by the principal components.

[0050] As an optional implementation, after extracting the data features, the variance filtering method (VarianceThreshold) is used for feature selection. First, the eigenvalue variance of each data feature is calculated; then, a variance filtering threshold N is preset, and the features with a variance lower than the variance filtering threshold N are filtered out by the Variance Threshold in sklearn (a machine learning library); finally, the remaining data features are selected, and the training set and the test set are divided according to 7:3 to prepare for subsequent machine learning, wherein the training set is the historical training set of the embodiment.

[0051] S2, input the induction heating parameters, physical property parameters and front surface temperature into the workpiece temperature prediction model to obtain the predicted non-front surface temperature of the standard workpiece.

[0052] As an optional implementation, the inductive heating parameters, the physical property parameters and the front surface temperature in the flattening process are input into the workpiece temperature prediction model in chronological order to obtain the predicted non-front surface temperature of the standard workpiece corresponding to each time point.

[0053] In the embodiment, the workpiece temperature prediction model can be an RNN (Recurrent Neural Network) network model, an LSTM (Long Short Term Memory) network model or a GRU (Gated Recurrent Unit) network model, etc.

[0054] S3, calculate the loss value according to the predicted non-front surface temperature and the obtained non-front surface temperature, and adjust the parameters of the workpiece temperature prediction model according to the loss value until the convergence condition is met.

[0055] In the embodiment, the loss value is the cumulative loss of the flattening process, that is, the loss values of the non-front surface temperatures and the obtained non-front surface temperatures at each time point are added up with different weight values, and the sum of the weight values is 1. In the embodiment, the non-front surface temperature can be at least one of the back surface temperature and the internal temperature, and when the non-front surface temperature includes the back surface temperature and the internal temperature, the loss value includes the loss between the predicted back surface temperature and the obtained back surface temperature and the loss between the predicted internal temperature and the obtained internal temperature.

[0056] As an optional implementation, the convergence condition can be to set a very small loss threshold, repeatedly train and continuously adjust the hyperparameters and learning parameters of the workpiece temperature prediction model until the loss value is lower than the set loss threshold, and obtain the trained workpiece temperature prediction model.

[0057] As an optional implementation, in the training process of the workpiece temperature prediction model, a lightweight processing is further included to facilitate subsequent deployment. For example, TensorFlow Lite (a deep learning framework) is used to quantize, prune, etc. the trained workpiece temperature prediction model, so as to make the workpiece temperature prediction model lightweight and prepare for subsequent embedding. Specifically, first, the workpiece temperature prediction model is fully quantized, that is, the weights, activation values and input values are all quantized to int8 (8 binary bits make up a byte), and all model operation operations are executed under int8 to achieve the best quantization effect; then the workpiece temperature prediction model is pruned, and the weights close to 0 in the workpiece temperature prediction model are set to 0 during training. After repeated training until the model parameters reach the target sparsity, the redundant parameters are deleted, and the workpiece temperature prediction model will be easier to compress.

[0058] As an optional implementation, the lightweight workpiece temperature prediction model is deployed. First, appropriate hardware and embeddable software systems are selected; second, appropriate deployment methods are selected according to the hardware and software; then, the lightweight workpiece temperature prediction model is exported to a deployable format and deployed on the selected hardware and software; finally, the performance and accuracy of the deployed workpiece temperature prediction model are tested by a test set, and optimization and adjustment are performed.

[0059] In this embodiment, a suitable training set is selected to train the workpiece temperature prediction model. Specifically, the inductive heating parameters, physical parameters and front surface temperature of the standard workpiece at each time during the flattening process are input into the workpiece temperature prediction model to obtain the predicted non-front surface temperature of the standard workpiece at the corresponding time; the model parameters of the workpiece temperature prediction model are trained according to the predicted non-front surface temperature and the obtained non-front surface temperature, and the model prediction result is more accurate.

[0060] Embodiment 2

[0061] This embodiment provides a workpiece temperature prediction method, as shown in Figure 2 The workpiece temperature prediction method comprises:

[0062] S21, obtaining the inductive heating parameters, physical parameters and front surface temperature of the workpiece to be flattened in the flattening process.

[0063] S22, inputting the inductive heating parameters, physical parameters and front surface temperature of the workpiece to be flattened into the workpiece temperature prediction model to obtain the predicted non-front surface temperature of the workpiece to be flattened.

[0064] The workpiece temperature prediction model is obtained by the model training method of Embodiment 1. In this embodiment, the non-front surface temperature can include at least one of the back surface temperature and the internal temperature, and the workpiece temperature prediction method can be applied in a high-frequency flattening device.

[0065] As an optional implementation, after the workpiece temperature prediction model is put into use, the error value of the actual required temperature and the predicted temperature is determined according to the flattening effect, and the related parameters of the workpiece temperature prediction model are continuously optimized according to the error value.

[0066] The workpiece temperature prediction model of this embodiment can predict the internal temperature and the back surface temperature of the workpiece in real time according to the front surface temperature, the induction heating parameters and the physical property parameters of the workpiece during the flattening process, greatly improving the controllability of flattening and realizing precise flattening.

[0067] Embodiment 3

[0068] This embodiment provides a model training system, as shown in Figure 3 The model training system includes a first acquisition module 11, a first prediction module 12 and a training module 13.

[0069] The first acquisition module 11 is configured to acquire a historical training set of the flattening process, wherein the historical training set includes the induction heating parameters, the physical property parameters, the front surface temperature and the non-front surface temperature of the standard workpiece at the same time.

[0070] In this embodiment, the induction heating parameters can include the heating power, the induction magnetic field frequency and / or the heating time. As an optional implementation, the heating power, the induction magnetic field frequency and the heating time can be read by the induction heating device during the flattening process, and the induction heating parameters are usually the sequence data preset by the induction heating device.

[0071] In this embodiment, the model training system can be applied in a high-frequency flattening device, which includes an induction heating device and a flattening device. The flattening process refers to heating the workpiece by using the electromagnetic heating device to make the workpiece generate an induction current flow to heat, and then using the flattening device to perform flattening operation according to the temperature of the workpiece. Flattening includes various modes, such as deep penetration heating mode and surface heating mode. The deep penetration heating mode refers to heating the front surface of the workpiece and stabilizing the front surface temperature at the Curie temperature, and maintaining for a period of time, and the heat is transferred to the bottom surface of the workpiece through heat conduction to realize temperature penetration and achieve the flattening effect. The surface heating mode refers to heating the front surface of the workpiece, controlling the heating depth and ensuring that the front surface temperature of the workpiece is greater than the back surface temperature of the workpiece to achieve the flattening effect. For different modes and different workpieces, the electromagnetic heating device will also adjust the corresponding parameters over time during the flattening.

[0072] In the embodiment, the physical property parameters can include workpiece material properties, workpiece specifications and / or workpiece structures. As an optional implementation, the workpiece material properties, workpiece specifications and workpiece structures can be read from design drawings. Since the physical property parameters of the workpiece are fixed parameters that do not change over time, the first acquisition module 11 needs to process them into time-varying sequence data. The workpiece material properties include carbon steel, alloy steel, cast iron and various alloys, etc. The workpiece structure is usually plate-shaped, i.e., the workpiece specifications are the length, width and thickness of the workpiece. When the workpiece has interfaces on both sides, the number of interfaces is further acquired. As an optional implementation, when the workpiece structure has interfaces, the material properties and specifications of the interfaces also need to be further acquired.

[0073] In the embodiment, the heating position of the electromagnetic heating device is usually at the center of the front surface of the workpiece. During the flattening process, the temperature gradually spreads from the center of the front surface to the periphery over time, and different positions on the front surface have different temperatures. As an optional implementation, the first acquisition module 11 acquires the temperatures of different points in a certain area on the front surface at each time and records the positions and temperatures of the corresponding points. For example, the temperature distribution of the point positions within a certain range, such as 1 cm, 2 cm, 3 cm, etc., from the center of the front surface is acquired. It should be noted that the distances between two points can be the same or different, such as the temperature distribution of the point positions within a certain range, such as 1 cm, 3 cm, 7 cm, etc., from the center of the front surface.

[0074] As an optional implementation, the first acquisition module 11 acquires the temperature data of the workpiece through some temperature acquisition devices. For example, the first acquisition module 11 installs an infrared temperature sensor at the heating position of the electromagnetic induction heating device and acquires the temperature of the front surface of the standard workpiece by using the radiation heat effect during workpiece heating, so that the detected temperature data is particularly accurate. In addition, when the electromagnetic induction heating device is working, the first acquisition module 11 can automatically acquire the temperature of the front surface of the workpiece through the infrared temperature sensor installed on the electromagnetic induction heating device. Of course, the embodiment is not limited to using an infrared temperature sensor to acquire the temperature.

[0075] In the embodiment, the non-front surface temperature can include the back surface temperature and / or the internal temperature. The back surface temperature refers to the temperature sequence data of the corresponding position of the heating position of the electromagnetic induction heating device on the back surface of the workpiece. The internal temperature refers to the temperature sequence data of the corresponding position of the heating position of the electromagnetic induction heating device within a certain depth from the front surface of the workpiece. The depth can be one-third to two-thirds from the front surface.

[0076] As an optional implementation, the back surface temperature of the workpiece can be collected by an infrared temperature sensor or a handheld temperature gun arranged on the back surface of the standard workpiece, and the internal temperature of the workpiece can be collected by embedding a thermocouple temperature sensor in the standard workpiece.

[0077] As an optional implementation, the first acquisition module 11 uses the pandas library (a data processing library) in python (a computer programming language) to perform data cleaning on the acquired raw data. Specifically, the first acquisition module 11 uses the isnull (a computer function indicating judging whether the data contains null values) function to find missing values, and then the first acquisition module 11 uses the ffill (a computer function indicating backward filling) function to fill the missing values longitudinally with the previous value or uses the bfill (a computer function indicating forward filling) function to fill the missing values longitudinally with the next value; the first acquisition module 11 uses the Z-Score (standard score) method to judge the outliers, sets a threshold M, and when the Z-Score of the data is greater than the set threshold M, it is judged as an outlier, and then the outliers are deleted or replaced; the first acquisition module 11 uses the duplicated (a computer function indicating finding and displaying repeated values in the data) function to judge the repeated values, and then deletes the repeated values; finally, the first acquisition module 11 uses the rolling (a time window function) function to split the data according to a fixed time step to obtain a plurality of training sample data.

[0078] As an optional implementation, the first acquisition module 11 uses the principal component analysis (PCA) method to extract the data features of the training sample data, thereby screening out the main data. First, the first acquisition module 11 performs linear transformation on the training sample data through linear function normalization, so that the result is mapped to the range of [0, 1], realizing equal proportion scaling of the training sample data; second, the first acquisition module 11 calculates the covariance matrix to identify the correlation; then the characteristic vectors and eigenvalues of the covariance matrix are calculated to identify the principal components of the data; thereafter, the first acquisition module 11 calculates the characteristic vectors and sorts them in descending order of eigenvalues, so as to find the principal components in order of importance; finally, the first acquisition module 11 multiplies the transpose of the training sample data by the transpose of the characteristic vector to reorient the data from the original axis to the axis represented by the principal components.

[0079] As an optional implementation, after extracting the data features, the first acquisition module 11 performs feature selection by using a variance filtering method (Variance Threshold). First, the first acquisition module 11 calculates the variance of the feature value of each data feature. Then, the first acquisition module 11 presets a variance filtering threshold N, and filters out the features with a variance lower than the variance filtering threshold N by using the VarianceThreshold in sklearn (a machine learning library). Finally, the first acquisition module 11 selects the remaining data features, and divides them into a training set and a test set according to a ratio of 7:3 to prepare for subsequent machine learning, wherein the training set is the historical training set of the present embodiment.

[0080] The first prediction module 12 is configured to input the induction heating parameters, the physical parameters, and the front surface temperature into the workpiece temperature prediction model to obtain the predicted non-front surface temperature of the standard workpiece. As an optional implementation, the first prediction module 12 inputs the induction heating parameters, the physical parameters, and the front surface temperature in the flattening process into the workpiece temperature prediction model in chronological order to obtain the predicted non-front surface temperature of the standard workpiece corresponding to each time point.

[0081] In the present embodiment, the workpiece temperature prediction model can be a model framework such as an RNN (Recurrent Neural Network) network model, an LSTM (Long Short-Term Memory) network model, or a GRU (Gated Recurrent Unit) network model. The predicted non-front surface temperature at each time point is fused with the induction heating parameters, the physical parameters, and the front surface temperature input at the next time point, and then is sent as new input data into the workpiece temperature prediction model for neural network calculation, which is repeatedly looped until the input data at the last time point is calculated.

[0082] The training module 13 is configured to calculate a loss value according to the predicted non-front surface temperature and the obtained non-front surface temperature, and adjust the parameters of the workpiece temperature prediction model according to the loss value until a convergence condition is met.

[0083] In the present embodiment, the loss value is the cumulative loss of one flattening process, that is, the training module 13 assigns different weight values to the loss values of the non-front surface temperatures and the obtained non-front surface temperatures at each time point for cumulative addition, wherein the sum of the weight values is 1. In the present embodiment, the non-front surface temperature can be at least one of the back surface temperature and the internal temperature. When the non-front surface temperature includes the back surface temperature and the internal temperature, the loss value includes the loss between the predicted back surface temperature and the obtained back surface temperature and the loss between the predicted internal temperature and the obtained internal temperature. Therefore, the training module 13 needs to continuously iterate and optimize the model parameters of the workpiece temperature prediction model so that both of the two losses are minimized.

[0084] As an optional implementation, the convergence condition can be to set a small loss threshold, and the training module 13 repeatedly trains and continuously adjusts the hyperparameters and learning parameters of the workpiece temperature prediction model until the loss value is lower than the set loss threshold, and the trained workpiece temperature prediction model is obtained.

[0085] As an optional implementation, during the training process of the workpiece temperature prediction model, the training module 13 is also used for lightweight processing of the workpiece temperature prediction model to facilitate subsequent deployment. For example, the training module 13 uses TensorFlow Lite (a deep learning framework) to quantize, prune, etc. the trained workpiece temperature prediction model to make the workpiece temperature prediction model lightweight and prepare for subsequent embedding. Specifically, first, the training module 13 performs full integer quantization on the workpiece temperature prediction model, that is, the weights, activation values and input values are all quantized to int8 (8 binary bits make up a byte), and all model operation operations are executed under int8 to achieve the best quantization effect; then the training module 13 prunes the workpiece temperature prediction model, and during the training process, the weights close to 0 in the workpiece temperature prediction model are set to 0, and after repeated training until the model parameters reach the target sparsity, the training module 13 deletes the redundant parameters, and the workpiece temperature prediction model will be easier to compress.

[0086] As an optional implementation, the training module 13 is also used for deploying the lightweight workpiece temperature prediction model. First, the training module 13 selects appropriate hardware and embeddable software systems; second, the training module 13 selects appropriate deployment methods according to the hardware and software; then, the training module 13 exports the lightweight workpiece temperature prediction model to a deployable format and deploys it to the selected hardware and software; finally, the training module 13 tests the performance and accuracy of the deployed workpiece temperature prediction model through the test set, and optimizes and adjusts it.

[0087] Embodiment 4

[0088] The embodiment provides a workpiece temperature prediction system, as shown in Figure 4 The workpiece temperature prediction system comprises a second acquisition module 21 and a second prediction module 22.

[0089] The second acquisition module 21 is used for acquiring the induction heating parameters, physical parameters and front surface temperature of the workpiece to be flattened in the flattening process.

[0090] The second prediction module 22 is used for inputting the induction heating parameters, physical parameters and front surface temperature of the workpiece to be flattened into the workpiece temperature prediction model to obtain the predicted non-front surface temperature of the workpiece to be flattened.

[0091] The workpiece temperature prediction model is obtained through the model training system of Example 3. In this example, the non-front temperature may include at least one of the back temperature and the internal temperature, and the workpiece temperature prediction system can be applied in a high-frequency leveling device.

[0092] As an optional implementation, after the workpiece temperature prediction model is put into use, the second prediction module 22 determines the error value between the actual required temperature and the predicted temperature based on the leveling effect, and continuously optimizes the relevant parameters of the workpiece temperature prediction model based on the error value.

[0093] Example 5

[0094] This embodiment provides an electronic device, such as... Figure 5 As shown, the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the model training method of Embodiment 1 or the workpiece temperature prediction method of Embodiment 2. As an optional implementation, the model training method and the workpiece temperature prediction method can be implemented on the same electronic device or on two different electronic devices. In this embodiment, the electronic device can be a high-frequency leveling device. Figure 5 The electronic device 30 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0095] like Figure 5 As shown, the electronic device 30 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including memory 32 and processor 31).

[0096] Bus 33 includes a data bus, an address bus, and a control bus.

[0097] The memory 32 may include volatile memory, such as random access memory (RAM) 321 and / or cache memory 322, and may further include read-only memory (ROM) 323.

[0098] The memory 32 may also include a program / utility 325 having a set (at least one) of program modules 324, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0099] The processor 31 performs various function applications and data processing, such as the model training method of embodiment 1 or the workpiece temperature prediction method of embodiment 2, by running the computer programs stored in the memory 32.

[0100] The electronic device 30 can also communicate with one or more external devices 34 such as a keyboard, a pointing device, etc. through an input / output (I / O) interface 35. Also, the model generation device 30 can communicate with one or more networks such as a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet, through a network adapter 36. As Figure 5 illustrated, the network adapter 36 communicates with the other modules of the model generation device 30 through the bus 33. It should be appreciated that although not shown, other hardware and / or software modules could be used in conjunction with the model generation device 30 such as, but not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0101] It should be noted that although several means / modules or sub-means / modules are mentioned in the foregoing detailed description, such division is merely exemplary and not mandatory. Indeed, according to an implementation of the present application, the features and functionalities of two or more means / modules described above can be embodied in one means / module. Conversely, the features and functionalities of one means / module described above can be further divided into several means / modules.

[0102] Embodiment 6

[0103] The present embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the model training method of embodiment 1 or the workpiece temperature prediction method of embodiment 2.

[0104] More specifically, the readable storage medium can include, but is not limited to, a portable disc, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0105] In a possible implementation, the present application can also be implemented in the form of a program product, which includes program codes for causing a terminal device to perform the model training method of embodiment 1 or the workpiece temperature prediction method of embodiment 2 when the program product is run on the terminal device.

[0106] programmable logic arrays, field programmable gate arrays, programmable logic devices, microcode, etc. In this manner, the program code can implement a virtual machine that operates in response to execution of the program code by the operation of a virtual machine instruction counter.

[0107] Although the present application has been described in connection with the embodiments thereof, it will be understood that the application is capable of further modifications. This application is intended to cover any variations, uses or adaptations of the application other than those shown in the specification.

Claims

1. A model training method, characterized in that, The model training method comprises: obtaining a historical training set of the flattening process, wherein the historical training set comprises induction heating parameters, physical property parameters, front surface temperature and non-front surface temperature of a standard workpiece at the same time; the front surface temperature is obtained by collecting the radiation heat effect of the standard workpiece during heating by a temperature collection device; the non-front surface temperature comprises back surface temperature and internal temperature, the back surface temperature refers to temperature sequence data of a position corresponding to a heating position of an electromagnetic induction heating device on a back surface of the standard workpiece, and the internal temperature refers to temperature sequence data of a position corresponding to the heating position of the electromagnetic induction heating device at a certain depth from the front surface inside the standard workpiece; the back surface temperature is obtained by collecting the temperature by a temperature sensor arranged on the back surface of the standard workpiece, and the internal temperature is obtained by collecting the temperature by a temperature sensor embedded in the standard workpiece; the heating position of the electromagnetic induction heating device belongs to the center of the front surface of the standard workpiece; the induction heating parameters comprise heating power, induction magnetic field frequency and / or heating time; the front surface temperature of the standard workpiece is obtained by collecting the temperature by an infrared temperature sensor arranged at the heating position of the electromagnetic induction heating device; inputting the induction heating parameters, the physical property parameters and the front surface temperature into a workpiece temperature prediction model to obtain predicted non-front surface temperature of the standard workpiece; calculating a loss value according to the predicted non-front surface temperature and the obtained non-front surface temperature, and adjusting parameters of the workpiece temperature prediction model according to the loss value until a convergence condition is met.

2. The model training method of claim 1, wherein, The physical property parameters comprise workpiece material properties, workpiece specifications and / or workpiece structures.

3. The model training method of claim 1, wherein, The model training method further comprises: performing lightweight processing on the trained workpiece temperature prediction model.

4. A workpiece temperature prediction method characterized by, The workpiece temperature prediction method comprises the following steps: obtaining induction heating parameters, physical property parameters and front surface temperature of a workpiece to be flattened during the flattening process; inputting the induction heating parameters, the physical property parameters and the front surface temperature of the workpiece to be flattened into a workpiece temperature prediction model to obtain predicted non-front surface temperature of the workpiece to be flattened; The workpiece temperature prediction model is obtained by the model training method according to any one of claims 1-3.

5. A model training system, comprising: The model training system comprises a first obtaining module, a first prediction module and a training module: The first acquisition module is configured to acquire a historical training set of a flattening process, wherein the historical training set comprises induction heating parameters, physical property parameters, a front surface temperature and a non-front surface temperature of a standard workpiece at the same time; the front surface temperature is acquired by a temperature acquisition device using a radiation heat effect when the standard workpiece is heated; the non-front surface temperature comprises a back surface temperature and an internal temperature, the back surface temperature refers to temperature sequence data of a position corresponding to a heating position of an electromagnetic induction heating device on a back surface of the standard workpiece, and the internal temperature refers to temperature sequence data of a position corresponding to the heating position of the electromagnetic induction heating device at a certain depth from a front surface of the standard workpiece; the back surface temperature is acquired by a temperature sensor arranged on the back surface of the standard workpiece, and the internal temperature is acquired by a temperature sensor embedded in the standard workpiece; the heating position of the electromagnetic induction heating device is the center of the front surface of the standard workpiece; the induction heating parameters comprise a heating power, an induction magnetic field frequency and / or a heating time; the front surface temperature of the standard workpiece is acquired by an infrared temperature sensor arranged at the heating position of the electromagnetic induction heating device; and the first prediction module is configured to input the induction heating parameters, the physical property parameters and the front surface temperature into a workpiece temperature prediction model to obtain a predicted non-front surface temperature of the standard workpiece. The training module is configured to calculate a loss value according to the predicted non-front surface temperature and the acquired non-front surface temperature, and adjust parameters of the workpiece temperature prediction model according to the loss value until a convergence condition is met.

6. A workpiece temperature prediction system characterized by, The workpiece temperature prediction system comprises a second acquisition module and a second prediction module. The second acquisition module is configured to acquire induction heating parameters, physical property parameters and a front surface temperature of a workpiece to be flattened in a flattening process. The second prediction module is configured to input the induction heating parameters, the physical property parameters and the front surface temperature of the workpiece to be flattened into a workpiece temperature prediction model to obtain a predicted non-front surface temperature of the workpiece to be flattened. The workpiece temperature prediction model is obtained by the model training system according to claim 5.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the model training method according to any one of claims 1-3 or the workpiece temperature prediction method according to claim 4.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the model training method according to any one of claims 1-3 or the workpiece temperature prediction method according to claim 4.

Images