A deep learning-based non-destructive testing method for silicon steel performance

Abstract

The application provides a deep learning-based nondestructive testing method for silicon steel performance, which comprises the following steps: taking the Barkhausen parameters, the tangential magnetic field strength harmonic analysis characteristic parameters, the incremental permeability characteristic parameters and the multi-frequency eddy current characteristic parameters of each measuring point of the to-be-tested silicon steel strip as the micromagnetic parameters of the corresponding measuring point; taking the tension value of the to-be-tested silicon steel strip and the linear distance between the surface of the to-be-tested silicon steel strip and the online micromagnetic detection probe as the production process parameters of each measuring point of the to-be-tested silicon steel strip; inputting the micromagnetic parameters and the production process parameters of any measuring point of the to-be-tested silicon steel strip into a quasi-Newton neural network model corresponding to the steel type of the to-be-tested silicon steel strip which has been trained; and outputting the mechanical performance parameters and the magnetic performance parameters of the measuring point by the quasi-Newton neural network model. The application can realize the online detection of the mechanical performance and the magnetic performance of the silicon steel.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of nondestructive testing of silicon steel, and particularly relates to a nondestructive testing method for silicon steel performance based on deep learning. BACKGROUND

[0002] Silicon steel sheet is one of the core materials of various electric machines, transformers and other electrical appliances, and is mainly used as the core of various electric machines, generators and transformers, and is an indispensable soft magnetic alloy in the power and electronic industries. The quality of silicon steel, the manufacturing technology of silicon steel and the product quality are one of the important indicators for measuring the production and technological development level of special steel. Effective evaluation of the product quality of silicon steel requires corresponding detection methods and standards. The finished product detection of cold-rolled silicon steel mainly includes magnetic properties, mechanical properties, magnetic aging, size, shape and surface quality and the like.

[0003] There are two kinds of magnetic property test standards. One is GB / T3655-2008, which measures the magnetic properties of electrical steel (strip) by using Epstein square. The detection process is that the 25cm Epstein square is composed of a primary coil, a secondary coil and a sample as a core, which forms a no-load transformer. The sample is assembled into a square frame by a double-tower joint, and four beams with equal length and cross-sectional area are formed. The measurement is carried out according to the alternating current characteristics. The sample width is required to be b=30mm±0.2mm, and the length is required to be l=300mm±0.5mm. The number of strips of the Epstein square experimental sample should be a multiple of 4, and 8 longitudinal and 8 transverse samples are cut from the same electrical steel for experiment. The other is GB / T13789-92, which measures the magnetic properties of single-piece electrical steel (strip). The main difference from GB / T3655-2008 is that an air gap permeameter is used instead of an Epstein square as a magnetizing mechanism. The air gap permeameter is composed of two symmetrical U-shaped yokes.

[0004] At present, silicon steel can also be detected by an on-line continuous iron loss detector, but the compensation coefficient needs to be manually changed according to different steel types during the detection process, and the measured value is not used as the delivery value of the contract, but only as a reference value for process adjustment.

[0005] The mechanical property test method of silicon steel refers to the standard GB228-2002. The tensile strength and yield strength of cold-rolled non-oriented silicon steel should meet the requirements of national standard GB2521-2008. The usual practice is to sample at the head and tail of the strip steel, then make a tensile sample according to the test method standard, and perform tensile test on the testing machine according to the requirements to obtain the corresponding detection results to represent the mechanical properties of the whole coil of strip steel.

[0006] In general, the nondestructive testing of silicon steel in the prior art must rely on the participation of manual and laboratory operation, and the measurement accuracy and efficiency need to be improved. SUMMARY

[0007] The present application aims to solve the problems in the prior art, and provides a deep learning-based multi-target optimization silicon steel performance nondestructive detection method.

[0008] The technical scheme adopted by the present application is a deep learning-based silicon steel performance nondestructive detection method, comprising the following steps:

[0009] In the production process, the tension value of the measured strip steel, the linear distance between the surface of the measured strip steel and the online micro-magnetic detection probe, and the thickness and type of the measured strip steel are obtained by measurement;

[0010] In the production process, the Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental permeability characteristic parameters and multi-frequency eddy current characteristic parameters of each measurement point of the measured strip steel are obtained by the online micro-magnetic detection probe;

[0011] The Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental permeability characteristic parameters and multi-frequency eddy current characteristic parameters of each measurement point of the measured strip steel are normalized as the micro-magnetic parameters of the corresponding measurement point;

[0012] The tension value of the measured strip steel and the linear distance between the surface of the measured strip steel and the online micro-magnetic detection probe are normalized as the production process parameters of each measurement point of the measured strip steel;

[0013] The micro-magnetic parameters and production process parameters of any measurement point of the measured strip steel are input into the trained quasi-Newton neural network model corresponding to the type and thickness of the measured strip steel; the quasi-Newton neural network model outputs the mechanical property parameters and magnetic property parameters of the measurement point;

[0014] The mechanical property parameters and magnetic property parameters of the measurement point are de-normalized to obtain the mechanical property and magnetic property data of the measurement point;

[0015] The mechanical property and magnetic property data of each measurement point on the measured strip steel form the mechanical property and magnetic property sequence of the strip steel.

[0016] In the above technical scheme, the mechanical property data includes yield strength, tensile strength, hardness and elongation; and the magnetic property data includes iron loss, magnetic induction and magnetic permeability.

[0017] In the above technical scheme, the training process of the quasi-Newton neural network model comprises:

[0018] Building a dataset: detecting the mechanical property data and magnetic property data of several strips of different steel grades and thicknesses through laboratory tests; obtaining the Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental permeability characteristic parameters and multi-frequency eddy current characteristic parameters of each strip through an online micro-magnetic detection probe; obtaining the tension value, the linear distance between the surface of each strip and the online micro-magnetic detection probe through measurement;

[0019] Delete the abnormal data in the data set, and normalize the data in the data set to obtain the micro-magnetic parameters, production process parameters and specification parameters, and mechanical property parameters and magnetic property parameters of each strip;

[0020] Based on the data set, a training sample set is constructed: different thicknesses of strips corresponding to each type of steel grade are respectively constructed into training sample sets; a single training sample information includes micro-magnetic parameters and production process parameters as model inputs, and mechanical property parameters and magnetic property parameters as training labels;

[0021] The training sample sets of strips of different thicknesses corresponding to each steel grade are respectively used to train the quasi-Newton neural network model to obtain the quasi-Newton neural network model corresponding to strips of different thicknesses under different steel grades.

[0022] In the above technical solution, the training process of the quasi-Newton neural network model further includes:

[0023] The number of hidden layer neurons of the quasi-Newton neural network model is determined;

[0024] The training parameter initialization includes setting the allowed error and the learning rate;

[0025] The training sample set is input into the quasi-Newton neural network model, the Adam algorithm is used to optimize the training process, the weights of the quasi-Newton neural network model are iteratively updated based on the training data, and the quasi-Newton neural network model is converged until the quasi-Newton neural network model is converged.

[0026] In the above technical solution, the training process of the quasi-Newton neural network model further includes:

[0027] The root mean square error (RMSE) is used to evaluate the performance of the trained quasi-Newton neural network model. The smaller the value of RMSE, the closer the predicted value of the model to the true value, the higher the accuracy of the model, and the better the performance. The parameters of the quasi-Newton neural network model are continuously adjusted until the RMSE value of the quasi-Newton neural network model meets the requirements.

[0028] The technical scheme further comprises the following steps: comparing the mechanical property and magnetic property sequence of a certain steel strip output by the quasi-Newton neural network model with the mechanical property and magnetic property data of the steel strip measured in the laboratory, and calculating the corresponding error rate; if the error rate exceeds a threshold value, the micro-magnetic parameters and production process parameters and the mechanical property parameters and magnetic property parameters of several steel strips obtained are combined to retrain and optimize the corresponding quasi-Newton neural network model.

[0029] The technical scheme further comprises the following steps: setting a time period, and after the end of each time period, the micro-magnetic parameters, production process parameters and specification parameters and the mechanical property parameters and magnetic property parameters of several steel strips obtained in the corresponding time period are combined to retrain and optimize the quasi-Newton neural network model.

[0030] In the technical scheme, the output frequency of the mechanical property and magnetic property sequence of the steel strip is the detection frequency of the online micro-magnetic detection probe.

[0031] In the technical scheme, the process of deleting abnormal data in the data set comprises the following steps: setting a threshold value of the mechanical property and magnetic property by hand, screening out abnormal values in the data set, and deleting the corresponding training sample data of the abnormal values as abnormal samples.

[0032] The application provides a computer readable storage medium, wherein a deep learning-based silicon steel performance nondestructive detection method program is stored on the computer readable storage medium, and the deep learning-based silicon steel performance nondestructive detection method program is executed by a processor to realize the steps of the deep learning-based silicon steel performance nondestructive detection method.

[0033] The application has the following beneficial effects: the quasi-Newton method is used to detect the micro-magnetic parameters of the steel strip produced online by using the online micro-magnetic detection equipment, the model calculation is performed through the classification of the steel grade and thickness, the mechanical property and magnetic property parameters of the steel strip are calculated, the detection system can realize the full-length online and meter-level force-magnetic property detection of the silicon steel, the detection efficiency is improved, and the method also provides a means for online process adjustment.

[0034] This invention classifies strip steel of different thicknesses under different steel grades and constructs corresponding quasi-Newtonian neural network models for each, improving the detection accuracy of strip steel. This invention trains the quasi-Newtonian neural network model to explore the relationship between the micromagnetic parameters and process parameters of silicon steel strip and their mechanical and magnetic properties, thereby enabling online detection of the mechanical and magnetic properties of silicon steel. During the construction of training samples, this invention obtains relevant process parameters of different strip steels from data on actual production lines, providing abundant and convenient data sources; simultaneously, outlier data in the dataset is removed to ensure the accuracy of the model. This invention determines the error rate of the model based on the real-time detection results output by the quasi-Newtonian neural network model, thereby further optimizing the model's accuracy. This invention reconstructs the training set based on the output results within a set time period to optimize and train the model, enhancing the model's detection accuracy. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the process of the present invention.

[0036] Figure 2 This is a schematic diagram comparing the measured hardness values ​​with the laboratory calculated values ​​in a specific embodiment.

[0037] Figure 3 This is a schematic diagram comparing the measured tensile strength with the laboratory calculated value in a specific embodiment.

[0038] Figure 4 This is a schematic diagram comparing the measured value and the laboratory calculated value of the yield strength in a specific embodiment.

[0039] Figure 5 This is a schematic diagram comparing the measured values ​​of iron loss with the laboratory calculated values ​​in a specific embodiment.

[0040] Figure 6 This is a schematic diagram comparing the measured value of magnetic induction with the laboratory calculated value in a specific embodiment. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but these descriptions do not constitute a limitation on the present invention.

[0042] like Figure 1 As shown, the present invention provides the following technical solution: a non-destructive testing method for silicon steel properties based on deep learning, comprising the following steps:

[0043] S1: Collect and preprocess the dataset required for predicting the mechanical and magnetic properties of silicon steel.

[0044] S2: The collected micromagnetic data and process data are normalized using standard methods;

[0045] S3: establishment of a quasi-Newton neural network model;

[0046] S4: optimization of the quasi-Newton neural network model using training sample data;

[0047] S5: prediction of the mechanical properties and magnetic properties of the measured strip steel using the trained quasi-Newton neural network model.

[0048] The present application trains the quasi-Newton neural network model corresponding to the steel grade and thickness by data acquisition, pre-processing of the data, and forming training samples, and inputs the micro-magnetic parameters of the strip steel detected online and the production process parameters of the strip steel into the optimized quasi-Newton neural network model corresponding to the type and thickness, to obtain the mechanical properties and magnetic properties of the strip steel being produced.

[0049] In the step S1, the data acquisition includes the micro-magnetic parameters obtained by measurement of the online micro-magnetic probe. The online micro-magnetic probe is a 3MA probe produced by Fraunhofer Company in Germany. The process parameters of the strip steel production and the specification parameters of the strip steel are obtained by a secondary system on the production line. The detection results of the laboratory mechanical properties and magnetic properties of the strip steel are obtained by laboratory tests.

[0050] Specific micro-magnetic parameters include Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental magnetic permeability characteristic parameters, and multi-frequency eddy current characteristic parameters obtained by detection of the strip steel by the probe; process parameters of the strip steel production include the tension value of the strip steel at the detection position of the probe and the height of the probe from the strip steel; specification parameters of the strip steel include the thickness of the strip steel and the classification of the strip steel; and detection results of the laboratory mechanical properties of the strip steel include yield strength, tensile strength, hardness, and elongation, and detection results of the magnetic properties include iron loss, magnetic induction, and magnetic permeability.

[0051] In the step S1, the pre-processing of the data refers to obtaining training samples after eliminating abnormal data and standardizing the collected data. The elimination of abnormal data refers to deleting useless or missing data, screening out abnormal values in the original data by manually setting thresholds of the mechanical properties and magnetic properties, and deleting all corresponding sample data as abnormal samples, i.e., corresponding micro-magnetic parameters, process parameters of the strip steel production, and specification parameters of the strip steel.

[0052] In the step S2, the standardization and normalization processing refers to taking 80% of the original sample data as a training set and the remaining 20% as a test set, and normalizing all data. Specifically, for different variable data, the following formula is used:

[0053]

[0054] The data is limited in the interval [0, 1], wherein x is the original data, x' is the standardized data, x min , x max are the maximum and minimum of the original data.

[0055] In the step S3, the establishment of the quasi-Newton neural network model and the setting of the initial value of the parameter include the following specific steps:

[0056] S31, the data samples are classified according to different thicknesses corresponding to different steel grades; different thicknesses correspond to different steel grade types, and the strip steel data belonging to a certain steel grade type and in a certain thickness interval form a sample set. Different thickness interval division methods are adopted for different types of steel, which is specifically based on the properties of different types of steel.

[0057] S32, the Barkhausen parameters, the tangential magnetic field strength harmonic analysis characteristic parameters, the incremental magnetic permeability characteristic parameters and the multi-frequency eddy current characteristic parameters detected by the online micro-magnetic detection probe, and the tension value and the height of the probe from the strip steel are used as input layer neurons.

[0058] S33, the laboratory inspection data yield strength, tensile strength, hardness and elongation, iron loss, magnetic induction and magnetic permeability are used as output layer neurons.

[0059] S34, the neuron excitation function f(x) is set: the implicit layer excitation function is tanh(x), and the excitation function of the output layer is purelin(x);

[0060] S35, the Nguyen-Widrow algorithm is used to initialize the weight and threshold value of the neural network prediction model, and the maximum network iteration number is 1000.

[0061] In the step S4, the training and optimization of the quasi-Newton neural network model include the optimization of the implicit layer neuron value, the quasi-Newton algorithm neural network and the prediction model. The sample sets formed in the step S31 are respectively divided into corresponding training sets and test sets, and the quasi-Newton neural network model is trained and tested by using the training sets and test sets of different types respectively, so as to obtain the optimized quasi-Newton neural network model corresponding to different strip steel thicknesses under different steel types.

[0062] The specific process of the training and optimization of the quasi-Newton neural network model in the step S4 is as follows:

[0063] S41, the number of implicit layer neurons is determined by the empirical formula , wherein, N i is the number of input layer neurons, N0 is the number of input layer neurons, and N sis the sample number of the training set, and is an arbitrary variable, usually in the range of 2-10.

[0064] S42, training parameter initialization, including setting the allowable error and learning rate.

[0065] S43, input the training sample into the quasi-Newton neural network model, optimize the training process using the Adam algorithm, and update the weight of the quasi-Newton network based on the training data until the model converges.

[0066] S44, the calculated mechanical properties and magnetic properties of silicon steel are subjected to inverse normalization processing, specifically: y'=x min +y(x max -x min ), wherein y' is the actual value after inverse normalization, y is the output value of the model, and x min , x max are the maximum and minimum values of the model input data.

[0067] The network training uses the root mean square error (RMSE) index to evaluate the performance of the trained model, which is the square root of the expected value of the square of the difference between the predicted value and the true value on the test set, that is:

[0068]

[0069] Where n is the number of data, xi is the true value, and pi is the predicted value. The smaller the value of RMSE, the closer the predicted value of the model to the true value, the higher the accuracy of the model, and the better the performance.

[0070] In step S5, the prediction of the mechanical properties and magnetic properties of silicon steel is performed by selecting a quasi-Newton neural network model for prediction based on the steel type and thickness of the measured strip steel fed back by the production line secondary system. The Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental permeability characteristic parameters and multi-frequency eddy current characteristic parameters detected by the online micro-magnetic detection probe, as well as the tension value and height data of the probe from the strip steel fed back by the production line secondary system are subjected to normalization processing, and then input into the selected quasi-Newton neural network model. The output value of the quasi-Newton neural network model is subjected to inverse normalization processing to obtain the mechanical properties and magnetic properties sequence of the length direction of the silicon steel being produced, and the output frequency corresponds to the detection frequency of the micro-magnetic detection probe.

[0071] During the operation of the present application, the performance data measured in the laboratory and the online prediction data are compared, and the corresponding error rate is calculated. If the error rate exceeds the threshold value, the model is retrained and optimized based on the latest historical data to ensure the accuracy of the prediction.

[0072] Preferably, the model can also be optimized in combination with recent historical data at a fixed period to improve the prediction accuracy of the model.

[0073] The application is further illustrated below by experimental data of specific embodiments.

[0074] The specific embodiment provides a deep learning-based multi-objective optimization silicon steel performance nondestructive detection method, including the following steps:

[0075] Step 1: Collecting online micromagnetic signal parameters of different steel types of 0.35mm-thick strip steel produced continuously in a non-oriented silicon steel production line, tension parameters of the strip steel, and simultaneously collecting laboratory mechanical and magnetic performance offline detection data of the strip steel.

[0076] Step 2: The micromagnetic detection probe detects 41 groups of micromagnetic parameters at each detection point, including 7 Barkhausen parameters (maximum amplitude, average amplitude in one excitation cycle, remanence point amplitude, magnetic field strength of the coercive field at the maximum amplitude, magnetic field strength of the coercive field at 25% of the maximum amplitude, magnetic field strength of the coercive field at 50% of the maximum amplitude, magnetic field strength of the coercive field at 75% of the maximum amplitude), 11 tangential magnetic field strength harmonic analysis characteristic parameters (amplitude and phase of 3, 5 and 7 harmonics, 3, 5, 7 and 9 harmonic amplitude sum, deformation coefficient, coercive force, harmonic amplitude at the zero point of the hysteresis loop and the steady-state voltage of the electromagnetic coil), 7 incremental permeability characteristic parameters (maximum amplitude, average amplitude in one excitation cycle, remanence point amplitude, width of the permeability curve at the maximum amplitude, 25% of the amplitude, 50% of the amplitude, 75% of the amplitude), and 16 multi-frequency eddy current characteristic parameters (real part, imaginary part, amplitude and phase of the coil impedance signal under 4 different eddy current frequencies). The sampling strip steel is provided with only 2 detection points online, which are the collection points at the beginning and end of the strip steel.

[0077] Each detection point has laboratory mechanical and magnetic performance data. The 41 groups of micromagnetic parameters, strip steel tension, thickness and mechanical and magnetic performance detection parameters are integrated into a table by using the related functions of EXCEL, and the samples with missing data and repeated samples are removed, the average value of multiple measurements is used to replace the abnormal data, and the sample data is formed.

[0078] Step 3: All sample data is divided into several categories according to the steel type and thickness. Each category of sample data is used to train the quasi-Newton neural network model. 80% of the sample data in each category of sample data is randomly selected as the training sample set, and the remaining 20% of the sample data is used as the test sample set.

[0079] Step 4: All sample data is normalized according to the formula .

[0080] Step 5: Using empirical formulas Determine the number of neurons in the hidden layer, where N i N0 is the number of neurons in the output layer, with a value of 1, meaning each mechanical and magnetic property is trained separately. N0 is the number of neurons in the input layer, with a value of 43. s This represents the number of samples in the training set, and α is set to 2. In this specific embodiment, the number of hidden layers is 2, and the number of neurons is set to 30.

[0081] Step 6: Set the neuron activation function f(x): the hidden layer activation function is tanh(x), the output layer activation function is purelin(x), and the weights and thresholds of the neural network prediction model are initialized using the Nguyen-Widrow algorithm. The maximum number of network iterations is 1000, the allowable error is set to 0.001, and the learning rate is 0.01.

[0082] Step 7: Input the normalized training sample sets of each type into the corresponding quasi-Newton neural network model, use the Adam algorithm to optimize the training process, and iteratively update the weights of the quasi-Newton network based on the training data until the model converges.

[0083] The root mean square error (RMSE) metric is used to evaluate the performance of a trained model. RMSE is the square root of the expected value of the squared difference between the predicted and actual values ​​on the test set.

[0084] Where n refers to the number of data points, x i p is the true value i The RMSE value is the predicted value. The smaller the RMSE value, the closer the model's predicted value is to the true value, the higher the model's accuracy, and the better its performance. Continuously adjust the model parameters until the model's RMSE value reaches the required level.

[0085] Step 8: Apply the trained quasi-Newtonian neural network models to the actual production line. The micromagnetic detection probe is installed perpendicular to the strip surface at a distance of 1mm from the strip surface to detect micromagnetic signals on the running strip. The 41 micromagnetic parameters collected in each group and the online process parameters (strip tension and probe distance from the strip) are input into the quasi-Newtonian neural network model corresponding to the steel type and thickness of the running strip. After inverse normalization, the output of the quasi-Newtonian neural network model is used to predict the mechanical and magnetic properties of a certain point of the strip.

[0086] The model output is inversely normalized, specifically: y' = x min +y(x max -x min ), where y' is the actual value after denormalization, y is the output value of the quasi-Newton neural network model, and x min xmax The maximum and minimum values ​​of the input data for the quasi-Newtonian neural network model.

[0087] Combination Figures 2-6 As shown in Table 1, the online prediction data of this specific embodiment is similar to the corresponding performance data measured in the laboratory, which indicates that the measurement accuracy of the present invention is high.

[0088] Table 1. Comparison of Measurement Accuracy in Specific Embodiments

[0089]

[0090]

[0091] Step 9: Continuously compare the performance data measured in the laboratory with the online prediction data, calculate the corresponding error rate. If the error rate exceeds the threshold, retrain and optimize the model using the most recent historical data to ensure prediction accuracy. If the accuracy can be guaranteed within a month, optimize the model monthly using recent historical data to improve the model's prediction precision.

[0092] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A non-destructive testing method for silicon steel properties based on deep learning, characterized in that: Includes the following steps: During the production process, the tension value of the strip to be tested, the straight-line distance between the surface of the strip to be tested and the online micro-magnetic detection probe, as well as the thickness and steel type of the strip to be tested are obtained; During the production process, Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental permeability characteristic parameters, and multi-frequency eddy current characteristic parameters of each measurement point of the strip under test are obtained through an online micro-magnetic detection probe. The Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental magnetic permeability characteristic parameters and multi-frequency eddy current characteristic parameters of each measurement point of the strip under test are normalized and used as the micro-magnetic parameters of the corresponding measurement points. The tension value of the strip to be tested and the straight-line distance between the surface of the strip to be tested and the online micro-magnetic detection probe are normalized and used as the production process parameters for each measurement point of the strip to be tested. Input the micro-magnetic parameters and production process parameters of any measurement point of the strip to be tested into the pre-trained quasi-Newton neural network model corresponding to the steel type and thickness of the strip to be tested; the quasi-Newton neural network model outputs the mechanical property parameters and magnetic property parameters of the measurement point; The mechanical and magnetic properties of the measurement point are inversely normalized to obtain the mechanical and magnetic property data of the measurement point. The mechanical and magnetic properties data of each measurement point on the strip to be tested form a sequence of the mechanical and magnetic properties of the strip. The training process of the quasi-Newtonian neural network model includes: Dataset construction: Mechanical and magnetic properties of several strips of different steel grades and thicknesses were tested in the laboratory; Barkhausen parameters, tangential magnetic field strength harmonic analysis characteristic parameters, incremental permeability characteristic parameters, and multi-frequency eddy current characteristic parameters of each strip were obtained through an online micro-magnetic detection probe; the tension value of each strip at the online micro-magnetic detection probe and the straight-line distance between its surface and the online micro-magnetic detection probe were measured. Remove outlier data from the dataset and normalize the data to obtain the micro-magnetic parameters, production process parameters, specification parameters, mechanical property parameters, and magnetic property parameters for each strip. Training sample sets are constructed based on the dataset: training sample sets are constructed for strip steel of different thicknesses corresponding to each type of steel; the information of a single training sample includes micro-magnetic parameters and production process parameters as input to the model, and mechanical performance parameters and magnetic property parameters as training labels; The quasi-Newton neural network models were trained using training sample sets of strip steel of different thicknesses corresponding to different steel grades, thus obtaining the quasi-Newton neural network models corresponding to strip steel of different thicknesses under different steel grades.

2. The method according to claim 1, characterized in that: The mechanical property data includes yield strength, tensile strength, hardness, and elongation; the magnetic property data includes iron loss, magnetic induction, and permeability.

3. The method according to claim 1, characterized in that: The training process of a quasi-Newtonian neural network model also includes: Determine the number of hidden layer neurons in the quasi-Newtonian neural network model; Initialize training parameters, including setting the allowable error and learning rate; The training sample set is input into the quasi-Newton neural network model, and the Adam algorithm is used to optimize the training process. The weights of the quasi-Newton neural network model are iteratively updated based on the training data until the quasi-Newton neural network model converges.

4. The method according to claim 3, characterized in that: The training process of a quasi-Newtonian neural network model also includes: The root mean square error (RMSE) metric is used to evaluate the performance of the trained quasi-Newton neural network model. The smaller the RMSE value, the closer the model's prediction is to the true value, the higher the model's accuracy, and the better its performance. The parameters of the quasi-Newton neural network model are continuously adjusted until the RMSE value of the quasi-Newton neural network model meets the requirements.

5. The method according to claim 1, characterized in that: It also includes the following steps: The mechanical and magnetic property sequences of a strip steel output by the quasi-Newton neural network model are compared with the mechanical and magnetic property data of the same strip steel obtained by laboratory measurement, and the corresponding error rate is calculated. If the error rate exceeds the threshold, the corresponding quasi-Newton neural network model is retrained and optimized by combining several micro-magnetic parameters, production process parameters, mechanical property parameters and magnetic property parameters of the strip steel that have been obtained.

6. The method according to claim 1, characterized in that: The process also includes the following steps: setting a time period, and after each time period, combining the micro-magnetic parameters, production process parameters, mechanical property parameters and magnetic property parameters of the strip steel obtained within the corresponding time period, retraining and optimizing the corresponding quasi-Newton neural network model.

7. The method according to claim 1, characterized in that: The output frequency of the mechanical and magnetic properties sequence of the strip is the detection frequency of the online micro-magnetic detection probe.

8. The method according to claim 3, characterized in that: The process of deleting outlier data in a dataset includes: manually setting thresholds for mechanical and magnetic properties to filter out outliers in the dataset, and deleting their corresponding training sample data as outlier samples.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a deep learning-based non-destructive testing method program for silicon steel properties. When the deep learning-based non-destructive testing method program for silicon steel properties is executed by a processor, it implements the steps of the deep learning-based non-destructive testing method for silicon steel properties as described in any one of claims 1-8.

Images