Magnetocal multi-sensor feature fusion stress detection system and detection method

Abstract

The present invention provides a magnetic-acoustic multi-sensing feature fusion stress detection system and detection method, wherein the Helmholtz coil is connected to a magnetic-acoustic sensor and a magnetic Barkhausen sensor; the Helmholtz coil is connected to a data acquisition card through a signal power amplifier; the magnetic-acoustic sensor is connected to the data acquisition card through a magnetic-acoustic signal amplifier; the magnetic Barkhausen sensor is connected to the data acquisition card through a Barkhausen signal amplifier; the Barkhausen noise signal output of the data acquisition card is also connected to a bandpass filter; the signal power amplifier generates a sinusoidal power signal and passes it into the Helmholtz coil. The present invention overcomes the problems of robustness and poor repeatability of existing stress detection technologies, proposes a fusion of two sensing features based on magnetic Barkhausen and magnetic-acoustic emission, constructs a feature selection and fusion network architecture based on principal component analysis, and constructs an evaluation method and model through a classification machine learning algorithm, thereby realizing high robustness, high precision, and high repeatability detection and prediction of ferromagnetic stress.

Description

Technical Field

[0001] The invention provides a magnetic-acoustic multi-sensing feature fusion stress detection system and a detection method, belonging to the technical field of ferromagnetism detection. Background Art

[0002] Ferromagnetic materials are widely used in important fields such as pressure vessels, aerospace and amusement facilities in my country. China is a major steel producer. According to statistics, my country's steel production in 2022 will reach more than 1 billion tons, and the export volume will reach more than 50 million tons. The steel production and export volume are both the first in the world. It is worth noting that steel structural parts will be subjected to repeated stress loads during service. Over time, there will be stress concentration in the components, which will eventually lead to accidents such as fracture. In fact, stress and stress concentration are one of the main causes of damage and failure of service components and engineering accidents. Evaluating the internal stress state of equipment and structures and their evolution laws, especially measuring and analyzing the damage and destruction of equipment and structures caused by stress, is an important basis for evaluating the structural life prediction of equipment parts and components and conducting strength and safety assessments on them, and has considerable social and economic benefits.

[0003] Traditional magnetic memory, magnetic Barkhausen, and hysteresis loop detection technologies have low sensitivity, poor feature redundancy, robustness, stability, and repeatability, and cannot accurately evaluate the stress state of microscopic regions of ferromagnetic materials.

[0004] With the development of multi-sensor fusion technology and feature fusion technology, multi-sensor feature fusion technology evaluation can achieve high robustness, high precision and rapid detection of internal stress of ferromagnetic metal components. However, at present, multi-sensor fusion technology is still limited to the initial stage, and the research on multi-sensor fusion features and their stress evaluation is blank at home and abroad.

[0005] According to the damage of the tested component, stress detection technology can be divided into two categories: non-destructive and destructive. Destructive stress detection is a mechanical method, including blind hole method, slotting method, etc. Destructive methods will inevitably cause damage to the product, which is inconsistent with the quality acceptance requirements of most projects. Therefore, under the premise of not affecting the performance and chemical properties of the components, non-destructive testing to complete qualitative and internal defect inspection of the components under inspection has attracted more and more attention. Non-destructive stress testing refers to the ability to evaluate the stress condition of a component by establishing a model of the change of applied stress and internal physical quantities of the material. Non-destructive testing is mainly based on physical methods, including conventional X-ray diffraction, ultrasonic method, Barkhausen noise (MBN) detection method, magneto-acoustic emission (MAE) detection method, etc.

[0006] Domestic and foreign scholars have made significant contributions to the characterization of ferromagnetic component materials through Barkhausen and magnetoacoustic emission signals. Although there are still few application cases of stress detection using Barkhausen and magnetoacoustic emission techniques, these magnetic non-destructive technologies have great potential value. At present, the surface state of ferromagnetic workpieces is usually complex, which will also lead to high error rates, missed detection rates and low repetition rates in stress inspection results. The shortcomings of existing technologies are as follows:

[0007] 1) Existing literature has made many contributions to the nondestructive testing of material stress, but there is little research on the stress correlation and prediction of Barkhausen testing technology and magnetic acoustic emission testing technology, and there is even less research on the combined analysis of the two. In addition, the stress prediction results of most nondestructive testing technologies are relatively poor. Therefore, combining magnetic nondestructive testing technology for research has become a key issue in improving the reliability of stress testing.

[0008] 2) Magnetoacoustic emission signals belong to acoustic detection, with good repeatability and robustness, but relatively low sensitivity. Barkhausen stress detection has high sensitivity, but low skin depth, poor repeatability, and low robustness. How to effectively combine the advantages of magnetoacoustic emission and magnetic Barkhausen detection technologies and overcome the weaknesses of each detection technology is an important path to improve stress prediction.

[0009] 3) Traditional single-sensor detection has the characteristics of poor stability, high randomness and low robustness. The multi-sensor information fusion framework has strong adaptability, taking into account the advantages of each sensor, strong stability and good robustness. Summary of the invention

[0010] The present invention provides a magnetic-acoustic multi-sensing feature fusion stress detection system and detection method to solve the following technical problems:

[0011] 1) Through the magnetic Barkhausen-magnetic acoustic emission multi-sensor detection technology, the shortcomings of poor stability and high randomness of single sensor are solved, and the robustness of stress detection is improved.

[0012] 2) Through the principal component analysis method, multiple characteristic parameters are correlated and analyzed, and the principal components are extracted for stress-magnetic-acoustic multi-feature modeling, which reduces the dimension of the feature vector and the redundancy of feature information, and reduces the complexity of the stress assessment and prediction model.

[0013] 3) By extracting 5 main features from Barkhausen and magnetoacoustic emission respectively and performing multi-feature fusion, the fused features are input into the extreme gradient boosting algorithm in ensemble learning for stress-multi-feature modeling training, thereby establishing a magnetoacoustic multi-feature stress assessment model, which solves the problem of low stability and robustness of a single feature and achieves stress prediction with high sensitivity, high precision and high robustness.

[0014] The specific technical solution is:

[0015] A magnetic-acoustic multi-sensing feature fusion stress detection system, including: a Helmholtz coil, a signal power amplifier, a magnetic-acoustic sensor, a magnetic Barkhausen sensor, a Barkhausen signal amplifier, a magnetic-acoustic signal amplifier, a data acquisition card, and a bandpass filter;

[0016] The signal power amplifier is connected to the Helmholtz coil;

[0017] The magnetic field magnetizes the workpiece being measured;

[0018] The magnetoacoustic sensor is connected to the data acquisition card via a magnetoacoustic signal amplifier;

[0019] The signal power amplifier is connected to the data acquisition card;

[0020] The magnetic Barkhausen sensor is connected to the data acquisition card through the Barkhausen signal amplifier;

[0021] The Barkhausen noise signal output collected by the data acquisition card is connected to a bandpass filter;

[0022] The Helmholtz coil generates an AC magnetic field, which magnetizes the workpiece being measured. Since the ferromagnetic material contains a large number of magnetic domains, the magnetic domains rotate under the action of the external magnetic field.

[0023] During the rotation of magnetic domains, the length of the material will change, resulting in a magnetostrictive effect. During the magnetostrictive process, the material releases energy and generates stress waves, namely magnetoacoustic emission. The magnetoacoustic sensor picks up the stress wave and converts it into an electrical signal. The magnetoacoustic signal is detected by the magnetoacoustic emission sensor and sent to the signal amplifier. After being amplified by 60dB, it is connected to the data acquisition card.

[0024] At the same time, the magnetic domain wall jumps under the external magnetic field and generates a pulse signal, which is obtained by a magnetic Barkhausen sensor. The magnetic Barkhausen signal obtained by the magnetic Barkhausen is sent to a 30dB signal amplifier and then connected to a data acquisition card.

[0025] The data acquisition card synchronously obtains the excitation signal, magnetoacoustic signal and magnetic Barkhausen signal, and performs subsequent analysis and processing;

[0026] In order to eliminate the direct coupling signal of the excitation signal in the Barkhausen detection coil, a bandpass filter is used to filter the collected Barkhausen signal.

[0027] The magnetic-acoustic multi-sensor feature fusion stress detection method comprises the following steps:

[0028] (1) The Barkhausen noise signal and magnetoacoustic signal of the workpiece under the action of an external magnetic field are obtained through a magnetoacoustic multi-sensor feature fusion stress detection system, that is, the original MBN and MAE signals;

[0029] (2) Calculate and obtain the envelope signal from the MBN and MAE original signals, and extract 10 characteristic parameters from the original signals and the envelope; after repeated magnetization for j cycles, a j×10 characteristic matrix can be obtained;

[0030] (3) In order to reduce redundant information and the dimension of the feature matrix, the principal component analysis (PCA) method is used to extract the main feature parameters, remove redundant information, and compress the j×10 feature spectrum matrix to j×5 dimensions;

[0031] (4) After feature extraction and feature matrix dimensionality reduction, the j×5 MAE feature spectrum matrix and the j×5 MBN feature spectrum matrix are connected to form a new j×10 feature spectrum matrix; the new feature spectrum matrix contains both MAE features and MBN features;

[0032] (5) 70% of the samples are used as the extreme gradient boosting algorithm in ensemble learning to train stress-magnetoacoustic multi-feature modeling; the remaining 30% of the samples are used to verify the accuracy of the model; at the same time, the dimension-reduced MAE or MBN single sensor feature spectrum matrix is ​​also used in the extreme gradient boosting algorithm in ensemble learning to train stress-single sensor multi-feature modeling; the stress prediction values ​​obtained by multi-sensor and single-sensor are compared to compare the advantages and disadvantages of the stress prediction performance of the three models.

[0033] The present invention overcomes the problems of robustness and poor repeatability of existing stress detection technologies, proposes a fusion of two sensing features based on magnetic Barkhausen and magnetoacoustic emission, constructs a feature selection and fusion network architecture based on principal component analysis, and constructs an evaluation method and model through a classification machine learning algorithm, thereby achieving high robustness, high precision, and high repeatability detection and prediction of ferromagnetic stress. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a magnetic-acoustic multi-sensor feature fusion stress detection system;

[0035] Figure 2 A stress assessment framework for magneto-acoustic multi-sensing signature fusion;

[0036] Figure 3a are the Barkhausen signal, envelope and excitation signal;

[0037] Figure 3b It is the magnetoacoustic emission signal, envelope and excitation signal;

[0038] Figure 4a Paul lines of Barkhausen signals under different stresses;

[0039] Figure 4b Paul lines of magnetoacoustic emission signals under different stresses;

[0040] Figure 5is the cumulative contribution rate of the principal components of MBN and MAE signal characteristics;

[0041] Figure 6 This is the XGBoost algorithm flow chart;

[0042] Figure 7 Modeling process for xgboost;

[0043] Figure 8 This is a schematic diagram of sampling with replacement;

[0044] Figure 9a It is a stress prediction model based on MBN multi-feature fusion;

[0045] Figure 9b It is a stress prediction model based on MAE multi-feature fusion;

[0046] Fig.9c It is a stress prediction model based on the fusion of magnetic and acoustic multi-sensor features;

[0047] Fig.10a It is the stress prediction diagram based on the MBN multi-feature parameter fusion model;

[0048] Fig.10b It is a stress prediction diagram based on MAE multi-feature parameter fusion;

[0049] Fig.10c It is the stress prediction diagram of the magneto-acoustic multi-sensor characteristic parameter fusion model;

[0050] Fig.11 Comparison of prediction results between single signal feature model and magnetic-acoustic fusion model;

[0051] Fig.12 This is a comparison chart of the model prediction value and the true value. DETAILED DESCRIPTION

[0052] The specific technical solutions of the present invention are described in conjunction with the accompanying drawings and embodiments.

[0053] The system of the present invention is as follows Figure 1 The magnetic-acoustic multi-sensing feature fusion stress detection system includes: a Helmholtz coil, a signal power amplifier, a magnetic-acoustic sensor, a magnetic Barkhausen sensor, a Barkhausen signal amplifier, a magnetic-acoustic signal amplifier, a data acquisition card, and a band-pass filter.

[0054] The signal power amplifier is connected to the Helmholtz coil;

[0055] The magnetic field magnetizes the workpiece being measured;

[0056] The magnetoacoustic sensor is connected to the data acquisition card via a magnetoacoustic signal amplifier;

[0057] The signal power amplifier is connected to the data acquisition card;

[0058] The magnetic Barkhausen sensor is connected to the data acquisition card through the Barkhausen signal amplifier;

[0059] The Barkhausen noise signal output collected by the data acquisition card is connected to a bandpass filter;

[0060] The working process of the system is as follows: First, the signal power amplifier generates a sinusoidal power signal and passes it into the Helmholtz coil. The coil generates an AC magnetic field, and the AC magnetic field magnetizes the workpiece to be measured. Since the ferromagnetic material contains a large number of magnetic domains, the magnetic domains rotate under the action of an external magnetic field. During the rotation of the magnetic domains, the length of the material will change, resulting in a magnetostrictive effect. During the magnetostrictive process, the material needs to release energy to generate a stress wave, namely magnetoacoustic emission. The magnetoacoustic sensor can pick up the stress wave and convert it into an electrical signal. At the same time, the magnetic domain wall jumps under the external magnetic field and generates a pulse signal, which can be obtained using a magnetic Barkhausen sensor. Since the magnetoacoustic emission signal and the magnetic Barkhausen signal generated by the workpiece to be measured are very weak, they cannot be directly recognized by the data acquisition card. After the magnetoacoustic signal is detected by the magnetoacoustic emission sensor, it is sent to the signal amplifier, amplified by 60dB, and connected to the data acquisition card. Similarly, the magnetic Barkhausen signal obtained by the magnetic Barkhausen is sent to the 30dB signal amplifier and then connected to the data acquisition card. The data acquisition card synchronously obtains the excitation signal, magnetic acoustic signal and magnetic Barkhausen signal, and performs subsequent analysis and processing. In order to eliminate the direct coupling signal of the excitation signal in the Barkhausen detection coil, this system uses a bandpass filter to filter the collected Barkhausen signal.

[0061] The magnetic-acoustic multi-sensor feature fusion stress detection method adopts the multi-sensor signal feature fusion method, such as Figure 2 As shown in Figure 2, the input of the proposed fusion framework is the MAE and MBN time series signals, and the output is the stress prediction value. The framework contains the following steps:

[0062] (1) The Barkhausen noise signal and magnetoacoustic signal of the workpiece under the action of an external magnetic field are obtained through a magnetoacoustic multi-sensor feature fusion stress detection system, that is, the original MBN and MAE signals;

[0063] (2) Calculate and obtain the envelope signal from the MBN and MAE original signals, and extract 10 characteristic parameters from the original signals and the envelope; after repeated magnetization for j cycles, a j×10 characteristic matrix can be obtained;

[0064] (3) In order to reduce redundant information and the dimension of the feature matrix, the principal component analysis (PCA) method is used to extract the main feature parameters, remove redundant information, and compress the j×10 feature spectrum matrix to j×5 dimensions;

[0065] (4) After feature extraction and feature matrix dimensionality reduction, the j×5 MAE feature spectrum matrix and the j×5 MBN feature spectrum matrix are connected to form a new j×10 feature spectrum matrix; the new feature spectrum matrix contains both MAE features and MBN features;

[0066] (5) 70% of the samples are used as the extreme gradient boosting algorithm in ensemble learning to train stress-magnetoacoustic multi-feature modeling; the remaining 30% of the samples are used to verify the accuracy of the model; at the same time, the dimension-reduced MAE or MBN single sensor feature spectrum matrix is ​​also used in the extreme gradient boosting algorithm in ensemble learning to train stress-single sensor multi-feature modeling; the stress prediction values ​​obtained by multi-sensor and single-sensor are compared to compare the advantages and disadvantages of the stress prediction performance of the three models.

[0067] Before establishing the extreme gradient boosting model in ensemble learning, data samples need to be prepared. The MAE and MBN signals of the pure iron sheet are collected using the MAE / MBN detection part of the magnetic-acoustic detection system. In order to obtain enough sample data, 60 MAE and MBN signals are collected and stored, from 0MPa to 100MPa, with a total of 21 acquisition stages. After storage, MATLAB is used to process the signal. Each time the MAE / MBN signal is collected, there are three to four complete cycles. The complete cycle of the MAE / MBN signal can be extracted separately based on the excitation sinusoidal signal, and the envelopes can be extracted from the Barkhausen signal and the magnetoacoustic emission signal, respectively, as shown in Figure 2. Figure 3a and Figure 3b shown.

[0068] Then, 10 eigenvalues ​​including peak value, peak time, rise time, fall time, ring count, energy, root mean square, skewness, kurtosis, and multi-scale entropy were extracted from the original MAE signal and envelope respectively, and 3929 signal data samples were obtained by removing blank or abnormal signals, and the correlation coefficient of the eigenvalues ​​of the signal data samples was reduced. Similarly, MATLAB was used to segment the collected MBN signal into a single-cycle signal, and the MBN signal background noise was filtered out by a bandpass filter. Then, 10 eigenvalues ​​including peak value, ring count, MBN energy, root mean square, skewness, kurtosis, peak-to-peak value, peak factor, pulse factor, and waveform factor were extracted from the MBN signal and envelope respectively, and 3929 samples were obtained by removing abnormal values.

[0069] In order to subsequently establish a relationship model between magnetoacoustic signal characteristics and stress, whenever the pure iron sheet increases the tensile stress by 5 MPa, multiple MAE and MBN signals are obtained according to the 0 MPa stress method, and the MAE and MBN envelopes under different stresses are obtained, such as Figure 4a As shown in Figure 4. According to a similar method, the MAE and MBN features of different stresses are extracted respectively, and the MAE and MBN characteristic value spectra under different stresses can be obtained.

[0070] If all 10 characteristic parameters extracted from the MAE and MBN signals in the previous part are used as training data sets, the speed and complexity of model training will be greatly increased. Due to the high correlation between different characteristic parameters, it is necessary to select appropriate characteristic parameters and sub-data sets for training the stress prediction model based on machine learning. Tables 1 and 2 are the Pearson correlation coefficients between the characteristic parameters of MAE and MBN, respectively. The correlation coefficients between the two characteristic parameters are greater than 0.8, which are underlined and bolded.

[0071] Table 1 Pearson correlation coefficients between Barkhausen characteristic parameters

[0072]

[0073] As can be seen from Table 1, there is a high correlation between the peak value and the peak-to-valley value and the pulse factor. This means that these three features reflect some similar characteristic information of the dynamic behavior of the magnetic domain. The correlation coefficient between the peak-to-valley value and the pulse factor is 0.91, which means that the two characteristic parameters change with stress in a similar manner. There is also redundant information between some other characteristic parameters. As can be seen from Table 1, there is also a high correlation between some characteristic parameters of magnetoacoustic emission. In MBN and MAE, redundant features contribute little to magnetic signal-stress modeling, and also increase the computational complexity and complexity of magnetoacoustic multi-sensor feature fusion modeling. Therefore, it is necessary to reduce the redundant features of magnetic signals and reduce the dimension of the characteristic spectrum.

[0074] Table 2 Pearson correlation coefficients between characteristic parameters of magnetoacoustic emission

[0075]

[0076] The principal component analysis method is an unsupervised dimensionality reduction technique that reduces the dimension of a data set by converting a large data set into a smaller dimensional data set. The main information of the original data set is still contained in the dimensionality reduction process. Therefore, the PCA algorithm not only eliminates redundant information, but also improves the efficiency of model training and simplifies the characteristic spectrum matrix. Due to the difference in the dimensions between eigenvalues, in order to avoid the eigenvalues ​​with larger values ​​being overweighted and the features with smaller values ​​being ignored during the PCA dimensionality reduction process, the sample matrix of the eigenvalues ​​of the sample MAE / MBN signal (3929×10) is firstly standardized in each column, and then the PCA program is used to reduce the dimension to obtain the comprehensive score of each principal component. Figure 5 It is a line chart of the cumulative contribution rate of each principal component of MBN and MAE characteristic parameters. Figure 5 It can be seen that in the PC1-PC5 interval, the cumulative contribution rate rises rapidly, and then rises very slowly. The cumulative contribution rates of the first five principal components of the MBN and MAE feature parameter matrices reach 96.33% and 97.54% respectively.

[0077] From the above analysis, we can see that the MAE and MBN characteristic signals are mainly concentrated in the first five principal components. It is generally believed that when the cumulative contribution rate exceeds 95%, the main information can be considered to be retained, and the explanatory power of the original variable reaches more than 95%. After being processed by the PCA algorithm, the dimension of the magnetic signal characteristic matrix is ​​reduced from 10 dimensions to 5 dimensions. The 3929×10-dimensional MBN / MAE data set is compressed to 3929×5 dimensions.

[0078] In the fusion process, the MBN and MAE matrices of 3929×5 obtained after PCA processing are merged to obtain a new matrix of 3929×10, which contains both MAE feature information and MBN feature information.

[0079] The feature samples after PCA dimension reduction have a total of 3929×5 columns of data as the input sample library of XGBoost. After the sample data is standardized, 70% is taken as the training set and 30% is taken as the test set. The present invention applies the XGBoost algorithm. The ensemble learning algorithm is essentially to build multiple individual learners, and through a certain strategy, multiple individual learners are organized to complete the task, and the prediction effect is better than that of a single learning learner. In ensemble learning, the individual algorithms that make up the entire ensemble learning can be called "weak learners". In essence, the weak learner is the base model constructed by a single algorithm, and the integration is to organize the base models to jointly determine the results. Figure 6 This is the flow chart of the XGBoost algorithm. First, the data sample library x is used as the training target and a weak learner (base model) f is generated after a round of training. 1 The purpose is to minimize the objective function and then calculate the residual L between the predicted value and the true value 1 As the next round of training target. After N rounds of training, N weak learners (base models) are generated, and the combination is the final strong learner f.

[0080] In terms of XGBoost algorithm application, the author Chen Tianqi established a special call library called xgboost, and its modeling flow chart is shown in Figure 7:

[0081] XGBoost algorithm is essentially an ensemble algorithm. Ensemble algorithm refers to building multiple weak learners on the data for aggregation. For all boost ensemble algorithms, after each weak learner is added, its prediction effect will be better than the previous model. However, if the number of iterations of weak learners is increased to gradually improve the prediction effect, the processing of input data is inevitable. In the boost ensemble algorithm, in order to prevent overfitting, it is necessary to have bootstrap sampling (bootstrap) on the input data, which ensures that each learner (each tree) is a model of a different series. If there is no sampling, the weak learners in the ensemble are all from the same series, which loses the meaning of the ensemble algorithm. At the same time, the feedback mechanism marks the samples with poor prediction effect last time, so that the probability of extracting wrong samples next time increases, which makes the next weak learner become an expert in conquering the previous wrong samples. As the learner becomes stronger, these samples will gradually predict correctly, so that after adding a weak learner iteration, its prediction effect will be improved. Figure 8 Schematic diagram of sampling with replacement for the XGBoost algorithm.

[0082] After setting the XGBoost parameters according to the experimental requirements, the MBN and MAE signal related data are used as data set samples, and the corresponding stress values ​​are used as targets. The proportion of sample training set and test set is set, and the XGBoost model is obtained after training.

[0083] After the magnetic-acoustic multi-sensor fusion modeling is completed, the data of the test group is used to verify the performance indicators of the model and the effect of the technology of the present invention.

[0084] (1) Comparison of model results

[0085] In order to study the performance indicators of the MAE single sensor feature fusion stress prediction model, the MBN single sensor feature fusion stress prediction model and the magnetic-acoustic multi-sensor feature fusion prediction model, the scatter distribution of the test set output results is shown in Figure 9. The test results are linearly fitted, and the 25% error upper and lower limits are defined as the maximum error allowed for the prediction results, which is used to measure the accuracy and dispersion of the test results. It can be seen from the figure:

[0086] 1) In the low stress stage, the error of the MAE and MBN single sensor feature fusion methods is relatively large, and a large number of scatter plots are distributed outside the 25% error limit. In addition, the dispersion of the MBN single sensor prediction results is greater than that of the MAE method, indicating that MAE has better robustness in stress prediction than the MBN feature fusion algorithm;

[0087] 2) When the stress at 9a in the figure exceeds 85MPa, the MBN prediction results are consistent with the actual results. Figure 9bWhen the medium stress exceeds 60MPa, the MAE stress prediction results are consistent with the actual results. The above shows that when the stress is large, the stress prediction results using MAE and MBN single-sensor stress are relatively ideal;

[0088] 3) Fig.9c It can be seen that the stress prediction values ​​and actual results are well matched in the low, medium and high stress stages by using the magnetic-acoustic multi-sensor stress fusion feature, and the predicted values ​​are basically within the error limit. This shows that the fusion of multi-sensor features improves the stability and accuracy of the stress prediction model.

[0089] also, Fig.10a , Fig.10b and Fig.10c It is a box plot drawn using the prediction results of three models. The box covers 50% of the stress prediction values ​​at each stage. The upper and lower bottoms of the box are the third quartile (the 75% value of all data in ascending order) and the first quartile (the 25% value of all data in ascending order). The distance between the upper and lower bottoms is also called the interquartile range (IQR). The edge of the box plot is set to 1.5 times the IQR. The height of the box can reflect the discrete distribution of the stress prediction values ​​at each stage to a certain extent. The flatter the box, the smaller the data volatility and the more stable the data.

[0090] As can be seen from the figure, Fig.10a The middle box is the longest. Fig.10b The box is second, Fig.10c The box in is the shortest, which proves that the prediction value volatility of the magnetic-acoustic multi-sensor feature fusion model is the smallest, followed by the MAE model and the MBN model. At the same time, the MBN feature fusion model obtains the most abnormal points in the stress prediction value, followed by the MAE feature fusion model, and the magnetic-acoustic multi-sensor feature fusion model predicts the least abnormal values. Similarly, Fig.10a The 1.5 IQR range is the largest, followed by Fig.10b , the smallest range is Fig.10c From the above analysis, it can be seen that for stress prediction, the model that combines all eigenvalues ​​of magnetic-acoustic has the highest prediction accuracy and the best stability, which proves that the fusion of MAE signal and MBN signal eigenvalues ​​has a better effect on stress assessment of ferromagnetic materials.

[0091] (2) Model evaluation

[0092] After using machine learning to build a model, evaluation indicators are needed to evaluate model performance. These are values ​​used to quantify the prediction effect of the established model. The commonly used evaluation functions for regression prediction are as follows:

[0093] 1) R squared / r 2 (R-square):

[0094] SSres =∑(y i -f i ) 2 (1)

[0095]

[0096]

[0097] where y i -f i Represents the difference between the true value and the predicted value. Indicates the difference between the true value and its mean. R squared / r 2 It is the determination coefficient, which refers to the percentage of the relationship between the predicted value and the true value. The larger the R square value, the better the model fit and the higher the accuracy.

[0098] 2) MSE (Mean Square Error):

[0099]

[0100] where f i -y i It also represents the difference between the true value and the predicted value. MSE reflects the degree of difference between the predicted value and the true value. The smaller the MSE value, the more stable the model is and the higher the reliability is.

[0101] Comparison of prediction results between the magnetic-acoustic fusion model and the single signal feature model Fig.11 As shown in the figure. As can be seen from the figure, the magnetic-acoustic multi-sensor feature fusion model has the largest R-squared, followed by the MBN feature fusion model, and the smallest is the MAE feature fusion model. This shows that the magnetic-acoustic multi-sensor feature fusion model has the highest stress prediction accuracy, followed by the MBN feature fusion model, and the MAE feature fusion model has the worst accuracy. The magnetic-acoustic multi-sensor feature fusion model has the smallest MSE, followed by the MAE feature fusion model, and the largest is the MBN feature fusion model. This shows that the test set results of all features of magnetic-acoustic fusion have the highest accuracy, with the R-squared value reaching the highest 0.9778, and the model has the highest stability. The test set MSE value is 16.185, which is much smaller than the MSE value of the single feature model.

[0102] (3) Model verification

[0103] In order to further verify the accuracy of stress assessment of the magnetic-acoustic fusion model, 15 groups of random tensile stress experiments were designed. The collected MBN and MAE signals were also subjected to signal filtering, feature extraction, and PCA dimensionality reduction to obtain the principal component comprehensive score with a principal component contribution rate of more than 95%. The specific tensile force value and the obtained principal component comprehensive score are shown in Figure 2. Fig.12 shown.

[0104] The 10 principal component comprehensive scores after dimension reduction in each group of experiments are input into the trained magnetic-acoustic fusion xgboost model. Fig.12 A comparison between the model prediction value and the true value is given. The model prediction value reflects the trend of stress change more accurately. By calculating 15 groups of experimental true values ​​and predicted values, the maximum relative error between the two is 12%, the minimum relative error is 3.9%, and the average relative error is 5.7%, which proves that the model has high reliability and prediction accuracy.

[0105] From the above comparison, it can be seen that the magnetic-acoustic multi-sensor fusion model proposed in the present invention has better performance than other single-sensor stress prediction models, and the proposed method promotes the development of intelligent decision-making in stress assessment and prediction.

[0106] 1) Model explanation

[0107] The main reasons why the magnetic-acoustic multi-sensor feature fusion model proposed in the present invention has better performance than the single-sensor feature fusion model are as follows:

[0108] 1) MBN noise mainly comes from the discontinuous movement of the 180° magnetic domain wall. The interaction between stress and the magnetic domain wall originates from the mutual balance between magnetoelastic energy and demagnetization energy. On the one hand, magnetoelastic energy promotes the jumping of MBN. On the other hand, the demagnetization effect caused by stress reduces the movement speed of the 180° magnetic domain and the MBN signal. Therefore, the Barkhausen signal is closely related to stress, which is the main reason why MBN is used for stress assessment and prediction, and it is also the reason why MBN technology has high sensitivity and accuracy in stress prediction.

[0109] 2) Stress-strain plays a role of pinning and anti-pinning, and hinders the movement of magnetic domains. This pinning and anti-pinning is random, resulting in high randomness in the movement of magnetic domain walls and MBN jumps. In addition, due to the skin effect, MBN mainly characterizes the movement information of magnetic domains with a depth of tens to hundreds of microns on the surface of the material. Due to the rough surface of industrial materials, the flatness of different parts of the sample is different, which makes the movement characteristics of the magnetic domain walls and MBN in different regions have certain differences, which further leads to the deterioration of the repeatability of the MBN signal. The MAE signal mainly comes from the stress wave released by the magnetostrictive effect generated by the 90° magnetic domain reversal. MAE can not only measure the motion state of the surface magnetic domain wall, but also reflect the internal or deep motion characteristics of the material. The MAE signal is the result of the average effect of the movement of magnetic domains at different depths of the material. The surface roughness has little effect on the magnetoacoustic signal. Therefore, magnetoacoustic emission has good repeatability in stress measurement.

[0110] In summary, MAE and MBN only reflect the motion characteristics of a certain type of magnetic domains, and have natural limitations in stress assessment. The magnetoacoustic multi-sensor multi-feature fusion model proposed in the present invention not only includes the surface, sub-surface and internal 90° and 180° magnetic domain dynamic characteristics. Therefore, the magnetoacoustic multi-sensor signal obtains more comprehensive characteristics that reflect the motion characteristics of the internal magnetic domains of the material. The magnetoacoustic multi-sensor feature fusion technology proposed in the present invention not only retains the high sensitivity and accuracy of the MBN technology in stress assessment, but also retains the advantages of high repeatability and robustness of the MAE method in stress prediction.

Claims

1. Magnetic-acoustic multi-sensing feature fusion stress detection system, It is characterized in that include: Helmholtz coil, signal power amplifier, magnetoacoustic sensor, magnetic Barkhausen sensor, Barkhausen signal amplifier, magnetoacoustic signal amplifier, data acquisition card, bandpass filter; The signal power amplifier is connected to the Helmholtz coil; The magnetic field magnetizes the workpiece being measured; The magnetoacoustic sensor is connected to the data acquisition card via a magnetoacoustic signal amplifier; The signal power amplifier is connected to the data acquisition card; The magnetic Barkhausen sensor is connected to the data acquisition card through the Barkhausen signal amplifier; The Barkhausen noise signal output collected by the data acquisition card is connected to a bandpass filter; The Helmholtz coil generates an AC magnetic field, which magnetizes the workpiece being measured. Since the ferromagnetic material contains a large number of magnetic domains, the magnetic domains rotate under the action of the external magnetic field. During the rotation of magnetic domains, the length of the material will change, resulting in a magnetostrictive effect. During the magnetostrictive process, the material releases energy and generates stress waves, namely magnetoacoustic emission. The magnetoacoustic sensor picks up the stress wave and converts it into an electrical signal. The magnetoacoustic signal is detected by the magnetoacoustic emission sensor and sent to the signal amplifier. After being amplified by 60dB, it is connected to the data acquisition card. At the same time, the magnetic domain wall jumps under the external magnetic field and generates a pulse signal, which is obtained by using a magnetic Barkhausen sensor; the magnetic Barkhausen signal obtained by the magnetic Barkhausen is sent to a 30dB signal amplifier and then connected to a data acquisition card; The data acquisition card synchronously obtains the excitation signal, magnetoacoustic signal and magnetic Barkhausen signal, and performs subsequent analysis and processing; In order to eliminate the direct coupling signal of the excitation signal in the Barkhausen detection coil, a bandpass filter is used to filter the collected Barkhausen signal.

2. Magnetocal multi-sensor feature fusion stress detection method, It is characterized in that The magnetic-acoustic multi-sensor feature fusion stress detection system according to claim 1 comprises the following steps: ( 1) Obtain the Barkhausen noise signal and magnetoacoustic emission signal of the workpiece under the action of an external magnetic field through the magnetoacoustic multi-sensor feature fusion stress detection system, that is, the original MBN and MAE signals; (2) Calculate and obtain the envelope signal from the MBN and MAE original signals, and extract 10 characteristic parameters from the original signals and the envelope; after repeated magnetization for j cycles, a j×10 characteristic matrix can be obtained; (3) In order to reduce redundant information and the dimension of the feature matrix, the principal component analysis (PCA) method is used to extract the main feature parameters, remove redundant information, and compress the j×10 feature spectrum matrix to j×5 dimensions; (4) After feature extraction and feature matrix dimensionality reduction, the j×5 MAE feature spectrum matrix and the j×5 MBN feature spectrum matrix are connected to form a new j×10 feature spectrum matrix; the new feature spectrum matrix contains both MAE features and MBN features; (5) 70% of the samples are used as the extreme gradient boosting algorithm in ensemble learning to train stress-magnetoacoustic multi-feature modeling; the remaining 30% of the samples are used to verify the accuracy of the model; at the same time, the dimension-reduced MAE or MBN single sensor feature spectrum matrix is ​​also used in the extreme gradient boosting algorithm in ensemble learning to train stress-single sensor multi-feature modeling; the stress prediction values ​​obtained by multi-sensor and single-sensor are compared to compare the advantages and disadvantages of the stress prediction performance of the three models.

Images