A parameter optimization method for electromagnetic induction detection based on multi-task learning model

Abstract

The present invention relates to an electromagnetic induction detection parameter optimization method based on a multi-task learning model, comprising the following steps: 1) constructing a sample database, including carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and thermal imaging sequence diagrams; 2) pre-processing the database, obtaining the temperature difference on the surface of each sample, establishing a data set with the surface temperature difference as key information, and dividing the samples of the data set into a training set and a test set; 3) constructing a PLE multi-task learning model, and performing training and verification to obtain the optimal surface temperature difference and the corresponding experimental parameter combination under each material parameter; 4) using a Bayesian algorithm to iteratively optimize the data set until the expected improvement value of the Bayesian optimization is lower than a set threshold, the optimization process is terminated, and an optimized neural network model is obtained. The present invention can efficiently and accurately obtain induction heating detection parameters with better detection effects.

Description

Technical Field

[0001] The present invention relates to an electromagnetic induction detection method, and in particular to an electromagnetic induction detection parameter optimization method based on a multi-task learning model, which is used for damage detection of carbon fiber composites and belongs to the technical field of carbon fiber composite damage detection. Background Art

[0002] Carbon fiber composites are widely used in aerospace, rail transit, shipbuilding and other fields. Internal defects and damages, such as delamination, cracks, and fractures, are prone to occur during their manufacturing and use. Since the damage occurs inside the composite, it is often difficult to detect. Thermal imaging detection technology is a common technology for material defect and damage detection. The use of eddy current induction heating can quickly heat up the conductive material, so that material damage can be found through thermal imaging sequence diagram analysis. Therefore, based on the conductive properties of carbon fiber, eddy current induction heating thermal imaging detection technology is also suitable for damage detection of carbon fiber composites.

[0003] However, the anisotropy and non-uniformity of carbon fiber composites make it difficult to match the detection parameters of eddy current induction heating, such as the excitation frequency, heating time, coil-sample distance, and coil input current, making it difficult to obtain the best damage detection effect. The empirical method and trial-and-error method rely on the operator's experience and repeated experiments, which are inefficient and produce unstable detection results.

[0004] Therefore, in order to solve the above problems, it is necessary to provide an innovative electromagnetic induction detection parameter optimization method based on a multi-task learning model to overcome the above defects in the prior art. Summary of the invention

[0005] The purpose of the present invention is to provide an electromagnetic induction detection parameter optimization method based on a multi-task learning model, which quickly and accurately finds the optimal electromagnetic induction heating parameters and improves the thermal imaging detection effect by constructing a multi-task learning model.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a method for optimizing electromagnetic induction detection parameters based on a multi-task learning model, which includes the following process steps:

[0007] 1) Build a sample database, including carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and thermal imaging sequence diagrams;

[0008] 2) Preprocess the database, obtain the temperature difference on the surface of each sample, establish a data set with surface temperature difference as key information, and divide the samples of the data set into training set and test set;

[0009] 3) Build a PLE multi-task learning model, and perform training and verification to obtain the optimal surface temperature difference and corresponding experimental parameter combination under each material parameter;

[0010] 4) Use the Bayesian algorithm to iteratively optimize the data set: obtain the optimal surface temperature difference predicted by the multi-task learning model, search and select new experimental parameters within the experimental parameter range, conduct experiments under the new experimental parameters, and obtain new data; the new data is added to the existing database, and the model is retrained, and the above operations are repeated until the expected improvement value of the Bayesian optimization is lower than the set threshold, the optimization process is terminated, and the optimized neural network model is obtained.

[0011] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: the step 1) is specifically:

[0012] 1-1), establish a sample database consisting of N groups of different carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and result data; wherein the carbon fiber composite material parameters include the carbon fiber ply structure sequence, number of layers, and ply angle; the electromagnetic induction heating experimental parameters include: excitation frequency, heating time, coil-sample distance, coil input current;

[0013] 1-2), by controlling the variable method, obtain N groups of sample combinations with different material parameters and experimental parameters;

[0014] 1-3), through experiments, the corresponding thermal imaging sequence diagrams of N sample combinations are obtained, so as to construct a sample database.

[0015] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: the step 2) is specifically:

[0016] 2-1), use Gaussian blur filter to reduce noise on the thermal imaging sequence of each sample;

[0017] 2-2), using Otsu's threshold segmentation method, automatically determine the optimal temperature threshold of the image and binarize the image;

[0018] 2-3) Use the Canny edge detection algorithm to identify the edges in the image, find the contours in the binary image, and determine the location and shape of the defect area;

[0019] 2-4), create a mask according to the outline of the defect area, select multiple normal areas in the image, calculate the average temperature of the defect area and the normal area respectively, and calculate their temperature difference; repeat the above operation for each frame of the image, calculate the temperature difference of each frame of the image, and select the largest temperature difference as part of the result data;

[0020] 2-5), randomly select 80% of the samples in the N groups of databases as training sets, and the remaining 20% ​​of the samples as test sets; randomly select M times to obtain M groups of training sets and validation sets.

[0021] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: the step 3) is specifically:

[0022] 3-1) Data standardization and normalization: Load the data in the database, including carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and surface temperature difference obtained by pretreatment; and standardize the material parameters and experimental parameters;

[0023] 3-2), define multiple input layers, including material parameter input layer, experimental parameter input layer and surface temperature difference input layer; input data enters the model through their respective input layers to form the initial feature vector:

[0024] ;

[0025] 3-3), build a shared layer: input feature vector, through multiple fully connected layers to form a shared feature extraction layer; these shared features are shared between different tasks and provide useful feature representation for each task:

[0026]

[0027] in, For the The output features of the layer, and Respectively The weights and biases of the layers, is the activation function; is the input vector of the input layer;

[0028] 3-4), construct task-specific layers, that is, construct two task layers, namely "surface temperature difference prediction task" and "experimental parameter optimization task"; each task has an independent task-specific layer to extract features related to the task;

[0029] 3-5) Training of multi-task learning model: The whole model is trained using M sets of training sets. The mean square error (MSE) is used as the loss function during the training process. The formula is as follows:

[0030]

[0031] in, is the actual value, is the model prediction value;

[0032] 3-6), input and output processing: input carbon fiber composite material parameters and output induction heating detection parameters;

[0033] 3-7), use the Adam optimization algorithm to optimize the model parameters and set the training stop conditions; stop training when the loss values ​​of the model on the training set and the validation set converge or meet the set conditions; calculate the mean square error and R square value of the model on the test set;

[0034] 3-8), from the M groups of training and verification results, select a set of model parameters with the smallest mean square error of the final output as the multi-task learning model.

[0035] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: in the step 3-1), the material parameters and the experimental parameters are standardized, and the standardized formula used is as follows:

[0036]

[0037] in, is the normalized input data, is the input parameter, is the mean of the input data, is the standard deviation.

[0038] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: in the steps 3-4), the output layer of the surface temperature difference prediction task is a fully connected layer, which predicts the surface temperature difference; the output layer of the experimental parameter optimization task is a fully connected layer, which outputs the experimental parameter combination; each task defines the following features:

[0039]

[0040]

[0041] in, and The surface temperature difference prediction task and the experimental parameter optimization task are respectively Output features of the layer;

[0042]

[0043]

[0044] in, is the predicted optimal surface temperature difference, is the predicted experimental parameter combination.

[0045] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: in the steps 3-7), the calculation formula of the R square value is:

[0046]

[0047] in, is the mean of the actual values.

[0048] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: the step 4) is specifically:

[0049] 4-1), using the Bayesian optimization algorithm, construct a Gaussian proxy model to approximate the objective function;

[0050] 4-2), select the expected improvement EI acquisition function and calculate the expected improvement value of each point in the entire parameter space;

[0051] 4-3), select new sampling points within the experimental parameter range according to the calculated improvement value, conduct experiments at the newly selected sampling points, obtain the carbon fiber thermal sequence diagram, and add the obtained data to the existing database as new training data, and retrain the neural network with the updated data;

[0052] 4-4), add the new data to the Gaussian proxy model, retrain the Gaussian process model, and use the updated Gaussian process model and acquisition function to select new sampling points;

[0053] 4-5), when the expected improvement value of the new sampling point Below the set threshold When , the optimization algorithm terminates.

[0054] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is further as follows: in the step 4-1), the Matérn kernel function is used in the Gaussian proxy model, and the function formula is as follows:

[0055]

[0056] in, and is the first input parameter and samples, is the length parameter scale, is the smoothness parameter, is the modified Bessel function, which has the following formula:

[0057]

[0058] in, is the order of the function, is the independent variable.

[0059] The electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention is also as follows: in the step 4-2), the formula of the EI acquisition function is as follows:

[0060]

[0061] in, is the Gaussian process model at the point The predicted mean at ; is the Gaussian process model in The standard deviation of the predictions; is the current optimal value; is the standard normal cumulative distribution function; is the standard normal probability density function.

[0062] Compared with the prior art, the present invention has the following beneficial effects: the electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention obtains the optimal surface temperature difference of each material parameter under different experimental parameters by constructing a multi-task learning model, and automatically optimizes the experimental parameters according to the optimal surface temperature difference. Compared with thermal imaging defect detection that relies on the operator's experience and repeated experiments to determine the detection parameters, the detection efficiency is greatly improved, and at the same time the accuracy of defect detection is improved, thereby achieving efficient and precise detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of the electromagnetic induction detection parameter optimization method based on the multi-task learning model of the present invention.

[0064] Figure 2 This is a flow chart of step 2) of the present invention for preprocessing the database.

[0065] Figure 3 It is an architecture diagram of the multi-task learning model in step 3) of the present invention. DETAILED DESCRIPTION

[0066] Please refer to the instruction manual Figure 1 To Attachment Figure 3 As shown, the present invention is a method for optimizing electromagnetic induction detection parameters based on a multi-task learning model, which includes the following process steps:

[0067] 1) Construct a sample database. The sample database includes carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and thermal imaging sequence diagrams. The specific method for constructing the sample database is:

[0068] 1-1), establish a sample database consisting of N groups of different carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and result data; wherein the carbon fiber composite material parameters include the carbon fiber ply structure sequence, number of layers, and ply angle; the electromagnetic induction heating experimental parameters include: excitation frequency, heating time, coil-sample distance, coil input current. In this embodiment, N=300.

[0069] 1-2), 300 groups of sample combinations with different material parameters and experimental parameters were obtained through the controlled variable method.

[0070] 1-3), through experiments, the corresponding thermal imaging sequence images of 300 sample combinations are obtained, so as to construct a sample database.

[0071] 2) Preprocess the database, obtain the temperature difference on the surface of each sample, establish a data set with surface temperature difference as the key information, and divide the samples of the data set into training set and test set.

[0072] Specifically, the preprocessing includes the following steps:

[0073] 2-1), use Gaussian blur filter to denoise the thermal imaging sequence of each sample to reduce noise.

[0074] 2-2), the optimal temperature threshold of the image is automatically determined by Otsu's threshold segmentation method, and the image is binarized.

[0075] 2-3), use the Canny edge detection algorithm to identify the edges in the image, find the contours in the binary image, and determine the location and shape of the defect area.

[0076] 2-4), create a mask according to the outline of the defect area, select multiple normal areas in the image, calculate the average temperature of the defect area and the normal area respectively, and calculate their temperature difference; repeat the above operation for each frame of the image, calculate the temperature difference of each frame of the image, and select the largest temperature difference as part of the result data.

[0077] 2-5), randomly select 80% of the samples in the 300 databases as training sets, and the remaining 20% ​​of the samples as test sets; randomly select M times to obtain M sets of training sets and validation sets. In this embodiment, M=10.

[0078] 3) Build a PLE (Progressive Layer Extraction) multi-task learning model, train and verify it, and obtain the optimal surface temperature difference and corresponding experimental parameter combination under each material parameter. The specific construction method is as follows:

[0079] 3-1), data standardization and normalization: load the data in the database, including carbon fiber composite material parameters (ply structure, number of layers, ply angle), electromagnetic induction heating experimental parameters (excitation frequency, heating time, coil and sample distance, coil input current) and surface temperature difference obtained by preprocessing; and standardize the material parameters and experimental parameters. In order to ensure the consistency of the input data, the material parameters and experimental parameters are standardized, and the standardized formula used is as follows:

[0080]

[0081] in, is the normalized input data, is the input parameter, is the mean of the input data, is the standard deviation, which indicates the degree of dispersion of the data.

[0082] 3-2), define multiple input layers, including material parameter input layer, experimental parameter input layer and surface temperature difference input layer; input data enters the model through their respective input layers to form the initial feature vector:

[0083] .

[0084] 3-3), build a shared layer: input feature vector, through multiple fully connected layers to form a shared feature extraction layer; these shared features are shared between different tasks and provide useful feature representation for each task:

[0085]

[0086] in, For the Output features of the layer; and Respectively Layer weights and biases; It is the activation function, which is responsible for mapping the linear combination in the neural network to the nonlinear space; is the input vector of the input layer.

[0087] 3-4), construct task-specific layers, that is, construct two task layers, namely "surface temperature difference prediction task" (i.e. task 1, task1) and "experimental parameter optimization task" (i.e. task 2, task2); each task has an independent task-specific layer for extracting features related to the task.

[0088] Furthermore, the output layer of the surface temperature difference prediction task is a fully connected layer, which predicts the surface temperature difference; the output layer of the experimental parameter optimization task is a fully connected layer, which outputs the experimental parameter combination; each task defines the following features:

[0089]

[0090]

[0091] in, and The surface temperature difference prediction task and the experimental parameter optimization task are respectively Output features of the layer;

[0092]

[0093]

[0094] in, is the predicted optimal surface temperature difference, is the predicted experimental parameter combination.

[0095] 3-5) Training of multi-task learning model: The whole model is trained using M sets (M=10) of training sets. The mean square error (MSE) is used as the loss function during the training process. The formula is as follows:

[0096]

[0097] in, is the actual value, is the model prediction value. In order to ensure the accuracy of the model, in this embodiment, the MSE loss function is set to converge to a threshold value less than 0.1 during the training process to ensure that the error of the model is controlled within a reasonable range.

[0098] 3-6), input and output processing: input carbon fiber composite material parameters and output induction heating detection parameters. Before optimizing the experimental parameters, the material parameters of the carbon fiber composite are first input, including the ply structure sequence, number of layers and ply angle. These parameters will be used as input data of the neural network model and combined with the eddy current induction heating experimental parameters to optimize the detection process.

[0099] After the neural network model is trained, it outputs a set of optimized induction heating detection parameters based on material parameters and surface temperature difference. These parameters include excitation frequency, heating time, distance between coil and sample, and coil input current, which will be used for further experiments and optimization.

[0100] 3-7), use the Adam optimization algorithm to optimize the model parameters and set the training stop condition; when the loss value of the model on the training set and the validation set reaches convergence or meets the set conditions, stop training; calculate the mean square error and R square value of the model on the test set; where the R square value is calculated using the following calculation formula:

[0101]

[0102] in, is the mean of the actual values.

[0103] In this embodiment, the convergence condition of the model is that the MSE of the training set and the validation set converge to below 0.1, and the R-squared value reaches above 0.9.

[0104] 3-8), from the 10 sets of training and verification results, select a set of model parameters with the smallest mean square error of the final output as the multi-task learning model.

[0105] 4) Use the Bayesian algorithm to iteratively optimize the data set: obtain the optimal surface temperature difference predicted by the multi-task learning model, search and select new experimental parameters within the experimental parameter range, conduct experiments under the new experimental parameters, and obtain new data; the new data is added to the existing database, and the model is retrained, and the above operations are repeated until the expected improvement value of the Bayesian optimization is lower than the set threshold, the optimization process is terminated, and the optimized neural network model is obtained.

[0106] The specific method of this step is:

[0107] 4-1), using the Bayesian optimization algorithm, a Gaussian proxy model is constructed to approximate the objective function. The Matérn kernel function is used in the Gaussian proxy model, and the function formula is as follows:

[0108]

[0109] in, and is the first input parameter and samples, is the length parameter scale, is the smoothness parameter. In this embodiment, the initial value of the length scale parameter is set to =1.0, set the smoothness parameter =1.5.

[0110] It is a modified Bessel function, which is used to describe the smoothness and similarity measure of the kernel function. Its formula is as follows:

[0111]

[0112] in, is the order of the function, is the independent variable.

[0113] 4-2), select the expected improvement EI acquisition function and calculate the expected improvement value of each point in the entire parameter space. The formula of the EI acquisition function is as follows:

[0114]

[0115] in, is the Gaussian process model at the point The predicted mean at ; is the Gaussian process model in The standard deviation of the predictions; is the current optimal value; is the standard normal cumulative distribution function; is the standard normal probability density function.

[0116] 4-3), select new sampling points within the experimental parameter range according to the calculated improvement value, conduct experiments on the newly selected sampling points, obtain the carbon fiber thermal sequence diagram, and add the obtained data to the existing database as new training data, and retrain the neural network with the updated data.

[0117] 4-4), add the new data to the Gaussian proxy model, retrain the Gaussian process model, and use the updated Gaussian process model and acquisition function to select new sampling points.

[0118] 4-5), when the expected improvement value of the new sampling point Below the set threshold When , the optimization algorithm terminates. In this embodiment, the threshold value is set .

[0119] The present invention uses the test set to test the trained multi-task learning model and count the offline test results. The database example is shown in Table 1 below, which includes material parameters, experimental parameters and result parameters after preprocessing.

[0120]

[0121] Table 1

[0122] The offline test results are shown in Table 2 below:

[0123]

[0124] Table 2

[0125] It can be seen that after training and testing the multi-task learning model, the model showed high accuracy on the training set and test set. In the offline test, the accuracy of the model reached 98.72%, the mean square error (MSE) was 0.04, and the R square value reached 0.92. This shows that the multi-task learning model can better predict the surface temperature difference of carbon fiber composites, and the optimized experimental parameter combination has high accuracy and stability.

[0126] In summary, by adopting the above method of the present invention, a multi-task learning model is constructed to obtain the optimal surface temperature difference and the corresponding experimental parameter combination under each material parameter. The optimal surface temperature difference is used to iteratively optimize new experimental parameters within the experimental parameter range, update the sample database and retrain the model to more accurately obtain the experimental parameter combination and improve the thermal imaging detection effect.

[0127] The above specific implementation methods are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for optimizing electromagnetic induction detection parameters based on a multi-task learning model, characterized in that: The process steps include: 1) Build a sample database, including carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and thermal imaging sequence diagrams; 2) Preprocess the database, obtain the temperature difference on the surface of each sample, establish a data set with surface temperature difference as key information, and divide the samples of the data set into training set and test set; 3) Build a PLE multi-task learning model, and perform training and verification to obtain the optimal surface temperature difference and corresponding experimental parameter combination under each material parameter; specifically: 3-1) Data standardization and normalization: Load the data in the database, including carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and surface temperature difference obtained by pretreatment; and standardize the material parameters and experimental parameters; 3-2), define multiple input layers, including material parameter input layer, experimental parameter input layer and surface temperature difference input layer; input data enters the model through their respective input layers to form the initial feature vector: ； 3-3), build a shared layer: input feature vector, through multiple fully connected layers to form a shared feature extraction layer; these shared features are shared between different tasks and provide useful feature representation for each task: in, For the The output features of the layer, and Respectively The weights and biases of the layers, is the activation function; is the input vector of the input layer; 3-4), build task-specific layers, that is, build two task layers, namely "surface temperature difference prediction task" and "experimental parameter optimization task"; each task has an independent task-specific layer to extract features related to the task; Among them, the output layer of the surface temperature difference prediction task is a fully connected layer, which predicts the surface temperature difference; the output layer of the experimental parameter optimization task is a fully connected layer, which outputs the experimental parameter combination; each task defines the following features: in, and The surface temperature difference prediction task and the experimental parameter optimization task are respectively Output features of the layer; in, is the predicted optimal surface temperature difference, is the predicted experimental parameter combination; 3-5) Training of multi-task learning model: The whole model is trained using M sets of training sets. The mean square error (MSE) is used as the loss function during the training process. The formula is as follows: in, is the actual value, is the model prediction value; 3-6), input and output processing: input carbon fiber composite material parameters and output induction heating detection parameters; 3-7), use the Adam optimization algorithm to optimize the model parameters and set the training stop conditions; stop training when the loss values ​​of the model on the training set and the validation set converge or meet the set conditions; calculate the mean square error and R square value of the model on the test set; 3-8), from the M groups of training and verification results, select a set of model parameters with the smallest mean square error of the final output as the multi-task learning model; 4) Iteratively optimize the data set using the Bayesian algorithm: obtain the optimal surface temperature difference predicted by the multi-task learning model, search and select new experimental parameters within the experimental parameter range, conduct experiments under the new experimental parameters, and obtain new data; add the new data to the existing database, and retrain the model, repeat the above operations until the expected improvement value of the Bayesian optimization is lower than the set threshold, the optimization process is terminated, and the optimized neural network model is obtained; specifically: 4-1), using the Bayesian optimization algorithm, construct a Gaussian proxy model to approximate the objective function; 4-2), select the expected improvement EI acquisition function and calculate the expected improvement value of each point in the entire parameter space; 4-3), select new sampling points within the experimental parameter range according to the calculated improvement value, conduct experiments at the newly selected sampling points, obtain the carbon fiber thermal sequence diagram, and add the obtained data to the existing database as new training data, and retrain the neural network with the updated data; 4-4), add the new data to the Gaussian proxy model, retrain the Gaussian process model, and use the updated Gaussian process model and acquisition function to select new sampling points; 4-5), when the expected improvement value of the new sampling point Below the set threshold When , the optimization algorithm terminates.

2. The electromagnetic induction detection parameter optimization method based on the multi-task learning model according to claim 1, characterized in that: The step 1) is specifically as follows: 1-1), establish a sample database consisting of N groups of different carbon fiber composite material parameters, electromagnetic induction heating experimental parameters and result data; wherein the carbon fiber composite material parameters include the carbon fiber ply structure sequence, number of layers, and ply angle; the electromagnetic induction heating experimental parameters include: excitation frequency, heating time, coil-sample distance, coil input current; 1-2), by controlling the variable method, obtain N groups of sample combinations with different material parameters and experimental parameters; 1-3), through experiments, the corresponding thermal imaging sequence diagrams of N sample combinations are obtained, so as to construct a sample database.

3. The electromagnetic induction detection parameter optimization method based on a multi-task learning model as claimed in claim 1, characterized in that: The step 2) is specifically as follows: 2-1), use Gaussian blur filter to reduce noise on the thermal imaging sequence of each sample; 2-2), using Otsu's threshold segmentation method, automatically determine the optimal temperature threshold of the image and binarize the image; 2-3) Use the Canny edge detection algorithm to identify the edges in the image, find the contours in the binary image, and determine the location and shape of the defect area; 2-4), create a mask according to the contour of the defect area, select multiple normal areas in the image, calculate the average temperature of the defect area and the normal area respectively, and calculate their temperature difference; Repeat the above operation for each frame of image, calculate the temperature difference of each frame of image, and select the largest temperature difference as part of the result data; 2-5), randomly select 80% of the samples in the N groups of databases as training sets, and the remaining 20% ​​of the samples as test sets; randomly select M times to obtain M groups of training sets and validation sets.

4. The electromagnetic induction detection parameter optimization method based on a multi-task learning model as claimed in claim 1, characterized in that: In step 3-1), the material parameters and experimental parameters are standardized, and the standardized formula used is as follows: in, is the normalized input data, is the input parameter, is the mean of the input data, is the standard deviation.

5. The electromagnetic induction detection parameter optimization method based on a multi-task learning model as claimed in claim 1, characterized in that: In the step 3-7), the calculation formula of the R square value is: in, is the mean of the actual values.

6. The electromagnetic induction detection parameter optimization method based on a multi-task learning model according to claim 1, characterized in that: In step 4-1), the Matérn kernel function is used in the Gaussian proxy model, and the function formula is as follows: in, and is the first input parameter and samples, is the length parameter scale, is the smoothness parameter, is the modified Bessel function, which has the following formula: in, is the order of the function, is the independent variable.

7. The electromagnetic induction detection parameter optimization method based on a multi-task learning model as claimed in claim 1, characterized in that: In step 4-2), the formula of the EI acquisition function is as follows: in, is the Gaussian process model at the point The predicted mean at ; is the Gaussian process model in The standard deviation of the predictions; is the current optimal value; is the standard normal cumulative distribution function; is the standard normal probability density function.

Images