Multi-fidelity data fusion method based on knowledge distillation

Abstract

The invention provides a multi-fidelity data fusion method based on knowledge distillation, and the method comprises the steps: building a transformer finite element model, extracting a sample set through an optimal Latin hypercube sampling method, and enabling the data set to comprise high-fidelity magnetic field data and low-fidelity magnetic field data; and performing multi-scale feature deep fusion on the high-fidelity magnetic field data and the low-fidelity magnetic field data based on a multi-fidelity data fusion method, and performing prediction based on a constructed deep learning model to obtain transformer magnetic field distribution, the deep learning model is obtained by training the fused high-fidelity magnetic field data and low-fidelity magnetic field data. Through the multi-fidelity data fusion method, the dependence of the deep learning model on a large amount of marked high-fidelity data is reduced, and the prediction of the transformer magnetic field distribution is realized.

Description

Technical Field

[0001] The present application belongs to the technical field of transformer magnetic field distribution, and in particular, relates to a multi-fidelity data fusion method based on knowledge distillation. Background Art

[0002] In the field of modern electrical equipment, especially in the calculation of magnetic field of transformers, a series of complex challenges are faced. These challenges mainly stem from the complexity of calculation, the need for multiple time and space scales, the strong coupling nature and the nonlinear characteristics of the system. Although traditional numerical simulation techniques such as finite difference method, finite element method (FEM) and moment method have been widely used in the comprehensive performance analysis of electrical equipment since the 1960s, these methods have begun to show limitations with the increase of design variables and the complexity of model structure. For example, complex electrical equipment modeling may cause simulation time to be as long as tens or even hundreds of hours, and in severe cases may cause computer system crashes, hindering the continuation of tasks. In addition, these methods cannot establish predictive models based on previous simulation results. Whenever the geometric structure or material properties change, they need to be re-simulated and solved, which increases the computational burden and cannot meet the requirements of digital twin technology.

[0003] With the increasingly harsh application environment and continuous improvement of performance of complex and high-performance electrical equipment, the need to systematically reflect the operating status of industrial processes and equipment has become urgent. The dimensions of design parameters and performance data are getting higher and higher, which increases the complexity of calculations. In recent years, artificial intelligence technology has developed rapidly. Compared with traditional machine learning methods, deep learning has a more flexible and complex network architecture and more powerful nonlinear characterization capabilities, which can provide support for the intelligence of transformers. However, deep learning technology also faces important challenges in its application, especially for extreme working conditions and large electrical equipment. It is difficult and expensive to obtain sufficient and effective training samples. Summary of the invention

[0004] In view of this, the present application aims to propose a multi-fidelity data fusion method based on knowledge distillation, which can reduce the computing time required for data set generation and verify the effectiveness of the method through the case of magnetic field prediction of amorphous alloy transformer.

[0005] To achieve the above purpose, the technical solution of this application is implemented as follows:

[0006] This application provides a multi-fidelity data fusion method based on knowledge distillation, including:

[0007] A transformer finite element model is constructed, and a sample set is extracted by an optimal Latin hypercube sampling method, wherein the data set includes high-fidelity magnetic field data and low-fidelity magnetic field data;

[0008] The high-fidelity magnetic field data and the low-fidelity magnetic field data are deeply fused with multi-scale features based on a multi-fidelity data fusion method, and the transformer magnetic field distribution is predicted based on a constructed deep learning model, wherein the deep learning model is trained by the fused high-fidelity magnetic field data and the low-fidelity magnetic field data.

[0009] Furthermore, the transformer finite element model is constructed, and a data set containing magnetic field data of different fidelity is obtained by an optimal Latin hypercube sampling method, including:

[0010] A three-dimensional model of the transformer is established through finite element simulation software, the geometric structure parameters of the transformer are selected as variables, the sampling interval is determined according to the structural dimensions of the actual transformer, and data sampling is performed;

[0011] Based on the sampling results, by changing the geometric structure variables of the transformer, parametric scanning is performed with the help of finite element analysis method to obtain three-dimensional simulation models with different geometric structures, and the corresponding magnetic field calculation results are calculated, wherein the three-dimensional simulation model includes a high-fidelity model and a low-fidelity model;

[0012] Construct data sets based on different magnetic field calculation results.

[0013] Furthermore, the geometric structure parameters include at least window width, window height, core width and core height.

[0014] Furthermore, the deep learning model building process includes:

[0015] Data set division, randomly dividing the data set into a training set, a validation set and a test set, and performing data preprocessing, wherein the preprocessing includes matrixing, normalization and color conversion;

[0016] Establish a deep learning model, learn the complex mapping relationship between geometric structure and core magnetic field distribution through model parameter training, use training set, validation set and test set to create the model, and calculate the prediction error of the model;

[0017] The magnetic field prediction is performed through the trained deep learning model, and evaluation indicators are selected to evaluate the prediction performance of the model.

[0018] Furthermore, it also includes:

[0019] The cosine annealing learning rate based on the adaptive moment estimation optimizer is used to dynamically adjust the model learning rate. The formula is as follows:

[0020]

[0021] In the formula, Represents the range of learning rate, T cur Represents the current iteration number, T i Represents the number of iterations in a cycle.

[0022] Furthermore, the mean square error is used as a standard for evaluating the prediction accuracy of the model, and the mean square error formula includes:

[0023]

[0024] In the formula, y i and are the true value and predicted value of the i-th sample respectively, and N is the number of samples.

[0025] Furthermore, the mean absolute percentage error, accuracy and Pearson correlation coefficient are used as indicators to analyze the prediction results, including:

[0026]

[0027]

[0028] Where n True Indicates the number of pixels where the deviation between the predicted result and the true result is less than 0.1T, n all is the total number of pixels in the image, X i and Y i Represent the predicted value and true value of the i-th sample respectively, and are the average of the predicted and actual values, respectively.

[0029] Compared with the prior art, the multi-fidelity data fusion method based on knowledge distillation described in this application has the following beneficial effects:

[0030] The multi-fidelity data fusion method based on knowledge distillation described in this application aims to solve the problem of high cost of magnetic field calculation for geometrically complex transformers using finite element method. Through the multi-fidelity data fusion method, the dependence of deep learning models on a large amount of labeled high-fidelity data is reduced, and the prediction of transformer magnetic field distribution is realized. Experimental results show that the method described in this embodiment can significantly reduce the amount of high-fidelity data required to build a deep learning model and reduce the total computing time cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0032] Figure 1This is a flow chart of a multi-fidelity data fusion method based on knowledge distillation described in an embodiment of the present application;

[0033] Figure 2 This is a finite element model diagram of a single-phase C-type core transformer described in an embodiment of the present application;

[0034] Figure 3 This is a finite element model diagram of a single-phase E-type core transformer described in an embodiment of the present application;

[0035] Figure 4 A high- and low-fidelity model mesh division diagram described in the embodiments of the present application;

[0036] Figure 5 A sample graph of a training set described in an embodiment of the present application;

[0037] Figure 6 A set of sample example diagrams of the E-type core transformer data set described in the embodiment of the present application;

[0038] Figure 7 This is a diagram of a training framework based on knowledge distillation described in an embodiment of the present application;

[0039] Figure 8 This is a schematic diagram of the CI-UNet++ structure described in an embodiment of the present application. DETAILED DESCRIPTION

[0040] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0041] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be the usual meanings understood by people with ordinary skills in the field to which the present application belongs. The "first", "second" and similar words used in the embodiments of the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0042] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0043] See also Figure 1As shown, this embodiment provides a multi-fidelity data fusion method based on knowledge distillation, which specifically includes the following steps:

[0044] Step S101: construct a transformer finite element model, and extract a sample set by using an optimal Latin hypercube sampling method, wherein the data set includes high-fidelity magnetic field data and low-fidelity magnetic field data.

[0045] Specifically, in this embodiment, step S101 is described in detail as follows, including:

[0046] 1.1 Construction of finite element model

[0047] Amorphous alloy transformers have great application and promotion space due to their superior energy-saving and consumption-reducing characteristics and simple manufacturing process. Therefore, this embodiment selects a single-phase C-type core transformer as a verification case and establishes a finite element model, such as Figure 2 shown.

[0048] The structural materials and excitation conditions of the transformer are as follows: the core material is amorphous alloy, the winding material is copper, and the power supply voltage is 40 V. The parameters such as the transformer core thickness, excitation frequency, and winding turns are shown in Table 1.

[0049] Table 1 Transformer parameters

[0050] parameter Numeric Core thickness E 52mm Excitation frequency 50Hz Primary winding turns 60 turns Secondary winding turns 60 turns Window radius R1 7mm Core radius R2 18mm

[0051] The three-dimensional model of the transformer is established by finite element simulation software. The geometric structural parameters of the transformer, such as window width, window height, core width and core height, are selected as variables, and the sampling interval is determined according to the structural dimensions of the actual transformer. The sampling interval is shown in Table 2.

[0052] Table 2 Structural parameter sampling interval

[0053]

[0054]

[0055] This embodiment also establishes a single-phase E-type core transformer as an additional calculation case, and establishes a corresponding finite element model, such as Figure 3 shown.

[0056] The structural materials and excitation conditions of the transformer are as follows: the core material is soft magnet, the winding material is copper, and the power supply voltage is 25 V. The parameters of the transformer core thickness, excitation frequency, and winding turns are shown in Table 3.

[0057] Table 3 Transformer parameters

[0058] parameter Numeric Excitation frequency 50Hz Primary winding turns 300 turns Secondary winding turns 300 turns Core thickness G 20mm Winding thickness E 8mm Winding height F 46mm

[0059] 1.2 Sampling method

[0060] Optimal Latin Hypercube Sampling (OLHS) is a statistical sampling method for generating multivariate random sample sets. OLHS aims to optimize the distribution of samples so that the distance between sample points is maximized in multidimensional space, which can reduce the redundancy between samples and ensure the uniformity and representativeness of samples. Through OLHS, highly representative sample points can be selected in the parameter space, and the sample points can be organized into a sample database based on the calculation results of numerical simulation, which lays the foundation for building subsequent magnetic field prediction models and helps to improve optimization efficiency. Finally, 625 sample points were obtained according to the sampling interval in Table 2, and the specific sampling results are shown in Table 4. Based on the sampling results, COMSOL parametric scanning is used to perform three-dimensional modeling of the transformer magnetic field, and the magnetic field distribution in the XZ plane is extracted.

[0061] Table 4 Optimal Latin Hypercube Sampling Results

[0062]

[0063] Through numerical simulation with COMSOL, different mesh parameters are set to build high-fidelity model and low-fidelity model respectively. The mesh parameters of two transformer models with different fidelity are shown in Table 5. The mesh division diagram of different fidelity models is shown in Figure 4 As shown, the left side is the high-fidelity model decomposition diagram, and the right side is the low-fidelity model decomposition diagram.

[0064] Table 5 Transformer model grid parameters with different fidelity

[0065] Grid Parameters High-fidelity models Low-fidelity models Mesh Vertices 51975 684 Number of units 304170 3819 Minimum unit mass 0.08974 0.07848 Average unit quality 0.6581 0.4655

[0066] Since we are not interested in the magnetic field distribution in the air domain, we need to create a mask to mark all values ​​less than the minimum value min of the magnetic flux density on the core. Set the values ​​less than min to white, and normalize the values ​​greater than or equal to min to range between 0 and 1.

[0067]

[0068] Among them, max is set according to the maximum value of the overall magnetic flux density. Then the data matrix is ​​mapped to obtain the three color components of R, G, and B. The specific transformation rules are shown in Table 6 below.

[0069] Table 6 Color transformation rules

[0070] Data Range R G B 0<B<min 255 255 255 min≤B<max <![CDATA[255g(B) 2 ]]> <![CDATA[2g(B)-g(B) 2 ]]> <![CDATA[255(1-g(B) 2 )]]> B≥max 255 255 0

[0071] Among them, when min ≤ B < max, after obtaining the RGB three-color components through the color transformation formula, the RGB three-color components are then synthesized into an image to obtain the output part in the sample. Batch operations are performed on the matrix data to obtain the visualization result of the magnetic field distribution. Through this invertible color transformation processing method, the predicted RGB image result can be re-transformed into a specific field value distribution map. This method not only improves the interpretability after field value visualization but also can obtain the specific values of the field distribution. One of the training set samples is as follows Figure 5 as shown. The high-fidelity magnetic field calculation result is on the left, and the low-fidelity magnetic field calculation result is on the right. It can be seen that the low-fidelity calculation result shows an "irregular magnetic field distribution in the form of burrs" in the local area of the edge, which does not conform to the physical law, but the overall trend of the magnetic field distribution is roughly consistent with the high-fidelity calculation result.

[0072] Select geometric structure parameters such as the core width, core height, window width of the core, and window height of the core of the transformer as variables. As shown in Table 7, each variable is divided into 5 sub-intervals. The total number of samples is obtained by multiplying the numbers within each sub-interval of the 4 variables, 5×5×5×5 = 625.

[0073] Table 7 Transformer variable values

[0074] Structural parameters Sampling interval / mm Core width A 74-86 Core height B 64-76 Core window width C 16-24 Core window height D 46-54

[0075] The specific sampling results are shown in Table 8. Based on the sampling results, COMSOL is used to perform three-dimensional modeling of the transformer magnetic field, and the magnetic field distribution in the XZ plane is extracted.

[0076] Table 8 Optimal Latin hypercube sampling results

[0077]

[0078] Based on the sampling results, by changing the geometric structure variables of the transformer and using finite element analysis for parametric scanning, three-dimensional simulation models with different geometric structures are obtained, and the corresponding magnetic field calculation results are calculated. An example of a set of sample data is as follows Figure 6 as shown. Figure (a) is the geometric structure diagram of the transformer, and Figure (b) is the magnetic field distribution diagram after simulation calculation. Each set of data corresponds one by one.

[0079] Step S102: Perform multi-scale feature depth fusion on the high-fidelity magnetic field data and the low-fidelity magnetic field data through a multi-fidelity data fusion method, and predict the transformer magnetic field distribution based on the constructed deep learning model. Among them, the deep learning model is trained by the fused high-fidelity magnetic field data and the low-fidelity magnetic field data.

[0080] Specifically, in this embodiment, the multi-fidelity data fusion method used is completed based on the training framework of knowledge distillation. Specifically, Figure 7 As shown in the figure, in the first stage, a large amount of low-fidelity data is used to train the teacher model to obtain in-depth data understanding. In the second stage, first, the parameters of the trained teacher model are frozen to ensure that they remain unchanged during subsequent training; then, the teacher model and the student model are allowed to perform forward propagation operations on the same data to generate prediction results. In each training cycle, the student model performs backpropagation by minimizing the comprehensive loss function, and only updates the parameters of the student model, so that the student model can both imitate the behavior of the teacher model and learn and improve directly from the labeled data. After training, the performance of the student model can be evaluated on the test set to verify the effectiveness of the knowledge distillation method.

[0081] In addition, this embodiment improves the UNet++ model by introducing the Inception module and the Convolutional Block Attention Module (CBAM) to establish a CI-UNet++ model. Figure 8 The details are as follows:

[0082] Replace the standard convolutional layer in UNet++ with the Inception module to achieve deep fusion of multi-scale features in the feature extraction stage, enhance the model's ability to capture features of various scales, and ensure that the model can more comprehensively understand and process complex input data

[0083] Introducing CBAM in the decoding stage optimizes the feature recognition process by weighting the channel and spatial dimensions of the feature map. This improvement enhances the model's ability to handle areas with rich details and drastic changes in the image, thereby improving the model's recognition accuracy and prediction performance for these key areas.

[0084] The construction process of the model is described as follows:

[0085] 2.1 Model Training

[0086] The key parameters in this embodiment are configured as follows:

[0087] (1) Adopt the cosine annealing learning rate adjustment strategy based on the Adaptive Moment Estimation (Adam) optimizer. The key advantage of the cosine annealing algorithm is that it can help the model avoid falling into the local minimum and increase the possibility of finding the global minimum. The formula for its dynamically changing learning rate is:

[0088]

[0089] In the formula, Represents the range of learning rate, T cur Represents the current iteration number, T i Represents the number of iterations in a cycle.

[0090] (2) In the selection of loss function, the mean square error (MSE) is uniformly used as the standard for evaluating the prediction accuracy of the model. MSE intuitively measures the difference between the model prediction value and the true value, and is one of the most commonly used performance evaluation indicators in regression tasks. By minimizing MSE, the model can learn how to accurately predict the transformer magnetic field distribution and reduce the prediction error. The calculation formula of MSE is

[0091]

[0092] In the formula, y i and are the true value and predicted value of the i-th sample respectively, and N is the number of samples.

[0093] (3) The 625 samples obtained through OLHS were divided into training set, validation set and test set in the ratio of 8:1:1.

[0094] (4) The Optuna hyperparameter optimization framework is used to fine-tune the performance of each model to ensure the best performance on the dataset. The key hyperparameters adjusted include the initial learning rate, batch size, the optimizer's first-order momentum decay rate (Beta1), the second-order momentum decay rate (Beta2), weight decay (Weight-Decay), the dropout rate (Dropout-Rate), and the model's main path depth. The optimized hyperparameter values ​​are shown in Table 9.

[0095] Table 9 Optimal hyperparameter combination based on Optuna

[0096] Training Hyperparameters Optimized hyperparameter values Initial Learning Rate 5.1×10-3 Batch-Size 8 Beta1 0.91 Beta2 0.98 Weight-Decay 8.4×10-7 Dropout-Rate 0.15 Main path depth 10

[0097] After selecting the optimal parameters, apply this set of parameters to the model and conduct a final performance test on an independent test set. The performance on the test set can objectively reflect the model's ability to process new data, evaluate the model's possible performance in actual application scenarios, and ensure the model's practicality and reliability.

[0098] 2.2 Evaluation indicators

[0099] Mean Absolute Percentage Error (MAPE), Accuracy and Pearson Correlation Coefficient (PCC) are added as indicators for analyzing the prediction results, where ACC refers to the proportion of pixels with a deviation within 0.1T to the total number of pixels, and black pixels are inaccurate points. The value range of PCC is between -1 and 1, 1 indicates a complete positive correlation, -1 indicates a complete negative correlation, and 0 indicates no linear correlation. The calculation formulas for MAE, MAPE, ACC and PCC are:

[0100]

[0101] Where n True Indicates the number of pixels where the deviation between the predicted result and the true result is less than 0.1T, n all is the total number of pixels in the image, X i and Y i Represent the predicted value and true value of the i-th sample respectively, and are the average of the predicted and actual values, respectively.

[0102] 2.3 Analysis of prediction results

[0103] In multi-fidelity fusion scenarios, high-fidelity data is usually scarce and expensive, and the number of high-fidelity samples is generally less than that of low-fidelity samples. Therefore, the following two different sample combinations are used to simulate actual scenarios:

[0104] (1) Combination 1 uses 50 high-fidelity samples and 200 low-fidelity samples to test the performance of each method in a scenario where high-fidelity samples are relatively scarce.

[0105] (2) Combination 2 uses 250 high-fidelity samples and 250 low-fidelity samples to test the performance of each method in a scenario where high-fidelity samples are relatively sufficient.

[0106] Using the same sample combination, three multi-fidelity data fusion methods and a single-fidelity model using only high-fidelity data are used to predict the C-type core transformer dataset, and the performance of each method is evaluated by the three indicators of MSE, MAE, and MAPE. The accuracy comparison of the prediction results is shown in Table 10.

[0107] Table 10 Comparison of the accuracy of magnetic field prediction results of C-type core transformer

[0108]

[0109] From the prediction accuracy results in Table 10, it can be seen that in scenarios where high-fidelity samples are relatively scarce, the prediction accuracy of the loss function reconstruction method is relatively poor due to its simple structure and focus on learning directly from data. The transfer learning method can partially make up for the problem of insufficient data by transferring the knowledge learned from low-fidelity data, so the prediction accuracy has been improved. The knowledge distillation method, thanks to its ability to effectively utilize the teacher model's deep understanding of complex data relationships and pass it on to the simpler student model, has the highest accuracy in scenarios where data is scarce. This result shows the unique advantage of knowledge distillation in extracting and transferring implicit knowledge between models.

[0110] In the scenario where high-fidelity samples are relatively sufficient, both knowledge distillation and transfer learning methods show good prediction accuracy, reflecting that these two methods can make full use of sufficient high-fidelity data resources. In contrast, although the loss function reconstruction method can still directly benefit from data learning when data is sufficient, its prediction accuracy is relatively poor compared to the other two methods.

[0111] In summary, the prediction accuracy of knowledge distillation is relatively high, and it can perform well in scenarios with different sample sizes.

[0112] The method described in this embodiment aims to solve the problem of high cost of magnetic field calculation for transformers with complex geometric structures using the finite element method. By using a multi-fidelity data fusion method, the deep learning model is reduced in its reliance on a large amount of labeled high-fidelity data, and the transformer magnetic field distribution is predicted. Experimental results show that the method described in this embodiment can significantly reduce the amount of high-fidelity data required to build a deep learning model and reduce the total computing time cost.

[0113] It should be noted that the above describes some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0114] The embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present application.

Claims

1. A multi-fidelity data fusion method based on knowledge distillation, characterized in that: include: A transformer finite element model is constructed, and a sample set is extracted by an optimal Latin hypercube sampling method, wherein the data set includes high-fidelity magnetic field data and low-fidelity magnetic field data; The high-fidelity magnetic field data and the low-fidelity magnetic field data are deeply fused with multi-scale features based on a multi-fidelity data fusion method, and the transformer magnetic field distribution is predicted based on a constructed deep learning model, wherein the deep learning model is trained by the fused high-fidelity magnetic field data and the low-fidelity magnetic field data.

2. The method according to claim 1, characterized in that The transformer finite element model is constructed, and a data set containing magnetic field data of different fidelity is obtained by an optimal Latin hypercube sampling method, including: A three-dimensional model of the transformer is established through finite element simulation software, the geometric structure parameters of the transformer are selected as variables, the sampling interval is determined according to the structural dimensions of the actual transformer, and data sampling is performed; Based on the sampling results, by changing the geometric structure variables of the transformer, parametric scanning is performed with the help of finite element analysis method to obtain three-dimensional simulation models with different geometric structures, and the corresponding magnetic field calculation results are calculated, wherein the three-dimensional simulation model includes a high-fidelity model and a low-fidelity model; Construct data sets based on different magnetic field calculation results.

3. The method according to claim 2, characterized in that: The geometric structure parameters include at least window width, window height, core width and core height.

4. The method according to claim 1, characterized in that: The deep learning model building process includes: Data set division, randomly dividing the data set into a training set, a validation set and a test set, and performing data preprocessing, wherein the preprocessing includes matrixing, normalization and color conversion; Establish a deep learning model, learn the complex mapping relationship between geometric structure and core magnetic field distribution through model parameter training, use training set, validation set and test set to create the model, and calculate the prediction error of the model; The magnetic field prediction is performed through the trained deep learning model, and evaluation indicators are selected to evaluate the prediction performance of the model.

5. The method according to claim 4, characterized in that Also includes: The cosine annealing learning rate based on the adaptive moment estimation optimizer is used to dynamically adjust the model learning rate. The formula is as follows: In the formula, Represents the range of learning rate, T cur Represents the current iteration number, T i Represents the number of iterations in a cycle.

6. The method according to claim 4, characterized in that The mean square error is used as a standard for evaluating the prediction accuracy of the model. The mean square error formula includes: In the formula, y i and are the true value and predicted value of the i-th sample respectively, and N is the number of samples.

7. The method according to claim 4, characterized in that The mean absolute percentage error, accuracy, and Pearson correlation coefficient are used as indicators to analyze the prediction results, including: Where n True Indicates the number of pixels where the deviation between the predicted result and the true result is less than 0.1T, n all is the total number of pixels in the image, X i and Y i Represent the predicted value and true value of the i-th sample respectively, and are the average of the predicted and actual values, respectively.

Images