Stress detection method and system based on multi-modal Gaussian kernel adaptive fusion

Abstract

The invention discloses a stress detection method and system based on multi-modal Gaussian kernel adaptive fusion. The method comprises the following steps: acquiring time sequence signals of magnetic Barkhausen noise MBN and magnetoacoustic emission signals MAE; extracting a multi-dimensional characteristic spectrum matrix, and normalizing the multi-dimensional characteristic spectrum matrix to obtain a characteristic sample; constructing a multi-modal Gaussian kernel adaptive fusion model according to the multi-modal Gaussian kernel product fusion function, and outputting a stress predicted value Z by using the multi-modal Gaussian kernel adaptive fusion model; and quantitatively evaluating the prediction effect of the multi-modal Gaussian kernel adaptive fusion model by using the evaluation function. According to the method, two sensing features of magnetic Barkhausen and magnetoacoustic emission are fused, a feature selection and fusion network architecture based on principal component analysis is constructed, an evaluation method and a model are constructed through a classification machine learning algorithm, and high-robustness, high-precision and high-repeatability detection and prediction of ferromagnetic construction stress are realized.

Description

Technical Field

[0001] The present invention relates to the field of nondestructive testing technology, and in particular to a stress detection method based on multi-modal Gaussian kernel adaptive fusion. Background Art

[0002] Ferromagnetic materials play a key role in mechanical manufacturing, infrastructure engineering, oil and gas transportation and other fields due to their good mechanical properties and mature processing technology. However, these materials are susceptible to fatigue damage, corrosion erosion and internal defects (such as cracks, pores, etc.) during service. If these hidden dangers are not identified in time, they may cause major safety accidents. Therefore, the development of advanced non-destructive testing technology is crucial to ensuring the safety performance of materials.

[0003] At present, the mainstream technology for non-destructive testing of ferromagnetic materials is mainly through ultrasonic, magnetic powder and other detection means (for macro defect detection) and traditional multi-sensor signal fusion methods. The latter is based on a linear combination of a single sensor source or a fixed mathematical model. Specifically, for magnetoacoustic emission (MAE) or Barkhausen (MBN) signals, independent prediction models are established respectively, and parameter calibration is performed through single variable regression or simple polynomial fitting; a preset structural function (such as polynomial product or linear superposition) is used for multi-sensor fusion, and its model order and parameter relationship are set by manual experience; the standard gradient descent method or Newton iteration method is used in the parameter optimization process to determine the coefficient by minimizing the training set error; the full-dimensional modeling of multi-feature signals is directly performed, but the feature dimensionality reduction preprocessing mechanism is not introduced; in the model verification stage, only the accuracy of the single fusion structure and the single sensor model is compared, and there is a lack of a systematic comparison framework for multiple types of fusion functions. In general, the above-mentioned magnetic non-destructive technology solutions have the following shortcomings:

[0004] Due to the complex surface state of ferromagnetic workpieces, the characteristic interaction form of existing technical solutions is solidified, which makes it difficult to adapt to the nonlinear coupling effects under complex working conditions, resulting in high error rate, missed detection rate and low repetition rate in stress inspection results.

[0005] Traditional polynomial and linear combination models are limited by fixed-order structures and cannot effectively capture the complex nonlinear correlation characteristics in multi-sensor data. When the data distribution is skewed or there are local extreme values, the prediction error increases significantly.

[0006] The existing exponential product model is highly sensitive to parameters and is prone to overfitting due to exponential explosion or parameter mismatch under limited training samples. It has poor generalization ability and is not robust enough to noise and outliers.

[0007] Single-dimensional data modeling that solely relies on MAE or MBN sensors lacks a collaborative compensation mechanism between multi-source features and is susceptible to environmental interference under complex working conditions, resulting in poor stability and low reliability of prediction results. Summary of the Invention

[0008] The embodiments of the present invention provide a stress detection method and system based on multimodal Gaussian kernel adaptive fusion to solve the problems of insufficient nonlinear modeling capability, poor robustness, and parameter optimization easily falling into local optimality in existing models, as well as the technical problem of difficulty in meeting the monitoring requirements for high precision and strong adaptability.

[0009] A stress detection method based on multimodal Gaussian kernel adaptive fusion, comprising:

[0010] Obtain the original response signals of the magnetic Barkhausen noise MBN and magnetic acoustic emission MAE of the ferromagnetic specimen, preprocess the original response signals to obtain the time series signals of the magnetic Barkhausen noise MBN and magnetic acoustic emission MAE;

[0011] A multidimensional characteristic spectrum matrix is extracted based on the time series signals of magnetic Barkhausen noise (MBN) and magnetic acoustic emission (MAE), and characteristic samples are obtained after normalization of the multidimensional characteristic spectrum matrix.

[0012] The feature samples are divided into a training set and a validation set. A multimodal Gaussian kernel adaptive fusion model is constructed based on the training set, the validation set, and the multimodal Gaussian kernel product fusion function. The stress prediction value Z is output using the multimodal Gaussian kernel adaptive fusion model.

[0013] The evaluation function is used to quantitatively evaluate the prediction effect of the multimodal Gaussian kernel adaptive fusion model.

[0014] A stress detection system based on multimodal Gaussian kernel adaptive fusion, comprising:

[0015] A signal acquisition module is used to obtain the original response signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE of the ferromagnetic specimen, and pre-process the original response signals to obtain the time series signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE;

[0016] The signal processing module extracts a multidimensional characteristic spectrum matrix based on the time series signals of magnetic Barkhausen noise (MBN) and magnetic acoustic emission (MAE), and obtains characteristic samples after normalizing the multidimensional characteristic spectrum matrix.

[0017] The model construction module is used to divide the feature samples into a training set and a validation set, build a multimodal Gaussian kernel adaptive fusion model based on the training set, the validation set and the multimodal Gaussian kernel product fusion function, and use the multimodal Gaussian kernel adaptive fusion model to output the stress prediction value Z;

[0018] The effect evaluation module is used to quantitatively evaluate the prediction effect of the multimodal Gaussian kernel adaptive fusion model using an evaluation function.

[0019] In the present invention, the stress detection method and system based on multimodal Gaussian kernel adaptive fusion perform best in terms of error bars, fitting effects, and accuracy in low, medium, and high stress stages. The stress prediction values are highly consistent with the actual results, and the predicted values are basically within the error limit. The system can effectively improve the stability and accuracy of the stress prediction model.

[0020] In the present invention, in response to the problem of insufficient nonlinear modeling capabilities of existing models, the present invention adaptively adjusts the mean and variance parameters to achieve global approximation of complex nonlinear relationships and improves prediction accuracy; the present invention solves the problem of poor robustness of traditional exponential / polynomial models, utilizes the smooth attenuation characteristics of Gaussian distribution to suppress noise interference, and combines the complementarity of multi-sensor data to enhance the generalization ability and stability of the model under non-ideal working conditions; the present invention optimizes the multi-objective parameter joint learning mechanism, and through the efficient collaboration of dimensionality reduction training set and L-BFGS-B algorithm, breaks through the bottleneck of traditional fusion model parameters easily falling into local optimality, achieves global rapid convergence, and ensures model reliability; the present invention constructs a multi-level comparative verification framework, solves the one-sidedness problem of single sensor information, and dynamically screens the optimal model through differential analysis of fusion prediction values and single sensor results to meet the industrial scenario's demand for high-precision and strong adaptability monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0022] Figure 1 is a flow chart of a stress detection method based on multimodal Gaussian kernel adaptive fusion in one embodiment of the present invention;

[0023] Figure 2 is a structural diagram of a stress detection system based on multi-modal Gaussian kernel adaptive fusion in one embodiment of the present invention;

[0024] Figure 3 is a structural diagram of a stress detection system based on multi-modal Gaussian kernel adaptive fusion in another embodiment of the present invention;

[0025] Figure 4 1 is a framework diagram of a stress detection system based on multi-modal Gaussian kernel adaptive fusion in one embodiment of the present invention;

[0026] Figure 5 Schematic diagram of envelopes of MAE signals and MBN signals of a stress detection method based on multi-modal Gaussian kernel adaptive fusion in one embodiment of the present invention;

[0027] Figure 6 It is a line graph of the MBN, MAE single sensor signals and the cumulative contribution rate of the principal components of the MBN and MAE splicing features in one embodiment of the present invention;

[0028] Figure 7 is a fitting diagram of stress prediction distribution of different models in one embodiment of the present invention;

[0029] Figure 8 is a comparison chart of stress prediction results of different models in one embodiment of the present invention;

[0030] Figure 9 1 is a comparison chart of model predicted values and true values in one embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0032] The stress detection method based on multimodal Gaussian kernel adaptive fusion provided by the embodiment of the present invention can be applied as follows: Figure 1 Specifically, the stress detection method based on multimodal Gaussian kernel adaptive fusion is applied in a stress detection system based on multimodal Gaussian kernel adaptive fusion, and the stress detection system based on multimodal Gaussian kernel adaptive fusion includes: Figure 2 The client and server shown communicate over a network, addressing the limitations of existing models, such as insufficient nonlinear modeling capabilities, poor robustness, and the tendency for parameter optimization to fall into local optima. This also addresses the technical challenges of meeting the requirements for high-precision and adaptable monitoring. The client, also known as the user end, is the program that corresponds to the server and provides local services to clients. The client can be installed on, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented as a standalone server or a server cluster consisting of multiple servers.

[0033] In one embodiment, if Figure 1 As shown in the figure, a stress detection method based on multi-modal Gaussian kernel adaptive fusion is provided. Figure 1 The server in the example is used as an example, and the steps are as follows:

[0034] S10, obtaining the original response signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE of the ferromagnetic specimen, and preprocessing the original response signals to obtain the time series signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE;

[0035] S20, extracting a multidimensional characteristic spectrum matrix based on the time series signals of magnetic Barkhausen noise (MBN) and magnetic acoustic emission (MAE), and obtaining characteristic samples after normalizing the multidimensional characteristic spectrum matrix;

[0036] S30, dividing the feature samples into a training set and a validation set, constructing a multimodal Gaussian kernel adaptive fusion model based on the training set, the validation set, and the multimodal Gaussian kernel product fusion function, and using the multimodal Gaussian kernel adaptive fusion model to output a stress prediction value Z;

[0037] S40. Use the evaluation function to quantitatively evaluate the prediction effect of the multimodal Gaussian kernel adaptive fusion model.

[0038] Understandably, the core principle of the stress detection method based on multimodal Gaussian kernel adaptive fusion is: starting work based on the magnetic-acoustic coupling effect, such as Figure 3 As shown, a sinusoidal excitation signal of a specific frequency is generated by an excitation source, which is amplified and input into a Helmholtz coil pair that can form a uniform alternating magnetic field.

[0039] Then, when the ferromagnetic specimen is placed in the field of the magnetic field, its internal magnetic domain structure produces a dynamic response. Driven by the external magnetic field, the magnetic domains are deflected in a directional manner, triggering magnetostriction. The microscopic deformation of the material accumulates to form a mechanical vibration wave, namely the magnetoacoustic emission effect. This signal is captured by the piezoelectric sensor; at the same time, the sudden movement of the magnetic domain wall excites discrete Barkhausen noise, which is converted into electromagnetic energy through a high-sensitivity induction coil.

[0040] Then, considering that both types of physical signals (magnetoacoustic emission effect signals and Barkhausen noise signals) are weak signals at the microvolt level, a hierarchical amplification strategy is adopted. The MAE magnetoacoustic channel is equipped with a 60dB preamplifier to improve the signal-to-noise ratio, and the Barkhausen MBN channel uses a 30dB gain dedicated conditioning circuit. The time domain waveform synchronously acquired by the acquisition card contains three key parameters: the original excitation signal, the magnetoacoustic response spectrum, and the Barkhausen pulse sequence. In addition, to eliminate electromagnetic coupling interference, a digital bandpass filter is designed specifically for the Barkhausen channel. The fundamental wave crosstalk is effectively suppressed through frequency domain feature extraction to ensure the accuracy of subsequent time-frequency analysis and characteristic parameter calculation.

[0041] In the present invention, the method innovatively integrates the dual detection mechanism of magnetoelastic effect and magnetic domain motion, and provides a multi-dimensional characterization means for material stress state evaluation through collaborative analysis of multi-physical field signals.

[0042] In one embodiment, if Figure 1and Figure 3 As shown, step S10 is: obtaining the original response signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE of the ferromagnetic specimen, preprocessing the original response signals, and obtaining the time series signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE; which includes the following sub-steps:

[0043] S101. Generate an original excitation signal based on an excitation source. After power amplification, the original excitation signal is input into a Helmholtz coil. The Helmholtz coil generates a uniform alternating magnetic field as a magnetizing excitation source. Through a magneto-mechanical-electrical coupling mechanism, it excites magnetoelastic waves and magnetic domain dynamics in the ferromagnetic specimen, thereby generating original response signals for magnetic Barkhausen noise (MBN) and magneto-acoustic emission (MAE). These original response signals provide raw data for subsequent signal processing.

[0044] S102. The original response signals of the magnetic Barkhausen noise MBN and the original response signals of the magnetoacoustic emission MAE are respectively subjected to signal amplifiers to improve the signal-to-noise ratio. It can be understood that a hierarchical amplification strategy is adopted, the MAE magnetoacoustic channel is configured with a 60dB preamplifier to improve the signal-to-noise ratio, and the Barkhausen MBN channel adopts a 30dB gain dedicated conditioning circuit.

[0045] S103, the acquisition card synchronously acquires the original excitation signal, the original response signal of the amplified magnetic Barkhausen noise MBN, and the original response signal of the amplified magnetoacoustic emission MAE; it can be understood that the time domain waveform obtained by the synchronous acquisition card includes three key parameters: the original excitation signal, the magnetoacoustic response spectrum, and the Barkhausen pulse sequence.

[0046] S104: Bandpass filters are used to filter out the noise floor of the original response signals of the amplified magnetic Barkhausen noise (MBN) and the amplified magnetic acoustic emission (MAE). The envelopes of the filtered signals are calculated, and a joint time-frequency domain analysis is performed to obtain time series signals of the magnetic Barkhausen noise (MBN) and the magnetic acoustic emission (MAE). To eliminate electromagnetic coupling interference, a digital bandpass filter is designed specifically for the Barkhausen channel. Frequency domain feature extraction effectively suppresses fundamental crosstalk, ensuring the accuracy of subsequent time-frequency analysis and characteristic parameter calculation.

[0047] In one embodiment, step S20 includes the following sub-steps:

[0048] S201. Quantitatively extract four time-domain features and four frequency-domain features from the original response signal waveforms and envelopes of magnetic Barkhausen noise (MBN) and magnetoacoustic emission (MAE).

[0049] Understandably, in order to obtain data collection of rich and diverse signal samples of ferromagnetic specimens, 70 MAE and MBN signal collections were performed, and during the collection process, a stress loading device was used to gradually apply stress from 0 MPa to 100 MPa, covering 11 collection stages, and the collected signals were properly stored. Then, MATLAB was used to perform data processing on the collected signals. Since each collected MAE / MBN signal contains nine to ten complete cycles, the complete cycle of the MAE / MBN signal can be extracted separately based on the excitation sinusoidal signal. At the same time, the envelopes are extracted from the Barkhausen signal and the magnetoacoustic emission signal, respectively, and the specific presentation form is as follows: Figure 5 shown.

[0050] S202. The four time-domain features and four frequency-domain features of magnetic Barkhausen noise (MBN) are spliced and fused into an MBN initial feature data set having eight features; the four time-domain features and four frequency-domain features of magnetic acoustic emission (MAE) are spliced and fused into an MAE initial feature data set having eight features.

[0051] It can be understood that eight eigenvalues, including peak value, root mean square value, skewness, kurtosis, center of gravity frequency, frequency band energy, frequency domain amplitude skewness, and frequency domain amplitude kurtosis, are extracted from the MAE original signal and envelope respectively. Similarly, MATLAB is used to segment the collected MBN signal into single-cycle signals, and a bandpass filter is used to filter out the background noise of the MBN signal. Then, the same eight eigenvalues are extracted from the MBN signal and envelope respectively.

[0052] S203 , performing feature concatenation and fusion on the MBN initial feature data set and the MAE initial feature data set to form a joint feature matrix with 16 features.

[0053] Understandably, the joint feature matrix can lay the foundation for subsequent feature selection and model construction. The 8 features of MAE and the 8 features of MBN extracted are spliced and fused to form a feature matrix containing 16 features (joint feature matrix), which provides a data basis for the subsequent construction of multimodal Gaussian kernel adaptive fusion model and fusion prediction model.

[0054] S204: Perform j iterations of periodic magnetization experiments on the joint feature matrix with 16 features to construct a feature dataset with j × 16 dimensions. Principal component analysis (PCA) is then used to extract the primary feature parameters, remove redundant information, and reduce the dimension of the j × 16 feature dataset to a j × 10 dimension. The j × 10-dimensional feature dataset is then normalized to obtain a 10-dimensional feature spectrum matrix, which serves as the feature sample. This step, by introducing a feature dimensionality reduction preprocessing mechanism, overcomes the bottleneck of traditional fusion model parameters that are prone to falling into local optimality, achieving rapid global convergence and ensuring model reliability.

[0055] Specifically, in another embodiment, step S204 includes the following sub-steps:

[0056] S2041, horizontally splicing the MBN initial feature data set with 8 features of the magnetic Barkhausen noise MBN and the MAE initial feature data set with 8 features of the magnetic acoustic emission MAE according to the sample order, that is, X concat = [F], forming a joint feature matrix with 16 features.

[0057] Understandably, the MBN initial feature data set with 8 features of magnetic Barkhausen noise MBN and the MAE initial feature data set with 8 features of magnetic acoustic emission MAE are used to construct two independent 8-dimensional feature vectors, and the 8-dimensional feature vectors of MAE and MBN are horizontally spliced according to the sample order to facilitate the subsequent analysis of the correlation and comprehensive characteristics between the two.

[0058] S2042, through the projection matrix W - PCA∈R^(16*10) maps the original data of the joint feature matrix to a low-dimensional space, obtaining XPCA=Xconcat.WPCA. After dimensionality reduction, the data becomes a 10-dimensional feature spectrum matrix, and the 10-dimensional feature spectrum matrix is normalized.

[0059] Understandably, since the original data has a high dimension, redundant features and high computational complexity, the unsupervised dimensionality reduction technology of principal component analysis (PCA) is used to reduce the dimensionality of the 16-dimensional data. The reduced dimensionality effectively improves the subsequent computational efficiency while retaining key information.

[0060] Understandably, if Figure 6 As shown in the figure, in the PCA dimensionality reduction process, in order to avoid the eigenvalues with larger values being overweighted and the features with smaller values being ignored, it is necessary to normalize the eigenvalue sample matrix (7722×10) after dimensionality reduction. The specific method is to normalize the data in units of each column. Figure 6 The cumulative contribution rate of each principal component of MBN, MAE single sensor signal and MBN and MAE splicing fusion features is shown in the broken line graph. Figure 6 It can be clearly seen that in the PC1-PC5 interval, the cumulative contribution rates of MAE and MBN single sensor features increase rapidly, and the cumulative contribution rates of the first five principal components of the MBN and MAE single sensor feature parameter matrices reach 97.57% and 99.61% respectively, after which the upward trend becomes very slow; while the cumulative contribution rate of MAE and MBN signal splicing fusion features increases relatively slowly, and the cumulative contribution rate reaches more than 99% at PC10.

[0061] Comprehensive analysis shows that the MAE and MBN single-sensor feature signals are primarily concentrated in the first five principal components, while the combined fusion features of the MAE and MBN signals are primarily concentrated in the first ten principal components. It is generally believed that when the cumulative contribution rate exceeds 95%, the primary information is considered retained, and the explanatory power of the original variables reaches over 95%. Using the PCA algorithm, the dimension of the combined fusion feature matrix of the MAE and MBN signals was successfully reduced from 16 to 10 dimensions. The 7722×16-dimensional MBN / MAE dataset was compressed to 7722×10 dimensions, and the new matrix contains both MAE and MBN feature information.

[0062] In another embodiment, step S30 includes the following sub-steps:

[0063] S301, apply the multimodal Gaussian kernel product fusion function to transform the feature sample’s feature Y’∈

[0064] R(7722*10) is mapped to the target parameter space, each Gaussian kernel corresponds to a weak learner, and cross-modal nonlinear coupling is achieved through kernel function multiplication. The mathematical expression of the multimodal Gaussian kernel product fusion function is:

[0065]

[0066] Among them, Y' represents the features after dimension reduction and normalization, Y′ i represents the principal component after dimensionality reduction, Z represents the stress prediction value, and λ i1 Represents the amplitude coefficient, which is used to characterize the global contribution weight of the modal feature, λ i2 represents the center offset, which is used to dynamically adjust the feature distribution to align with the target, λ i3 Represents the bandwidth coefficient, which is used to control the feature sensitivity range.

[0067] It can be understood that the above-mentioned cross-modal nonlinear coupling is achieved through kernel function multiplication, which is more suitable for eliminating the physical dimension differences of heterogeneous features (MAE / MBN) than the traditional weighted summation.

[0068] S302, pre-train the base model by the feature-kernel function matching method, and according to the mutual information maximization principle, for each principal component Y′ after dimensionality reduction i Assign the optimal Gaussian kernel and filter out the labels corresponding to the true label Z true Significantly correlated features are used to construct an initial set of base models.

[0069] Understandably, only when the principal component Y′ i and the true label Z true Mutual information MI(Y′ i ,z ture )>θ MI(θ MI =0.15), the principal component will be retained, thus constructing the initial base model set.

[0070] S303: After completing feature screening, the parameters of the base model are initialized using an equal-weight initialization method to ensure that all features have the same weight in the initial stage.

[0071] Understandably, the initialization parameters include three parts:

[0072] Initialized as the dimensionally reduced and normalized principal component Y′ i The median of This provides robust central estimates for the model.

[0073] Initialized as the principal component Y′ after dimensionality reduction and normalization i The interquartile range (IQR) of This is a robust estimation based on the interquartile range, which effectively avoids the impact of outliers on parameters.

[0074] Initialized to equal weights, that is Where N is the number of features. This equal-weight initialization ensures that all features have the same weight at the initial stage, providing a fair starting point for subsequent model optimization. Significantly correlated features are then screened to construct the initial base model set.

[0075] The above steps S302-S303 are the first stage of the staged optimization strategy for model establishment.

[0076] S304. Use the limited memory Broyden-Fletcher-Goldfarb-Shannon bounded optimization algorithm (L-BFGS-B) to collaboratively optimize the design objective function, which is used to optimize the base model parameters and suppress overfitting, and impose constraints. The mathematical expression of the objective function is:

[0077]

[0078] Among them, through the first is the mean square error (MSE) term, which is used to ensure the prediction accuracy of the model. The second term L1 is the L1 regularization term, which is used to constrain the sparsity of the amplitude coefficient. L1 regularization penalizes non-zero parameters, causing the amplitude coefficients corresponding to unimportant features in the model to approach zero, thereby achieving feature selection and model simplification, and improving the generalization ability of the model. is the reciprocal bandwidth penalty term, which is used to suppress overfitting caused by too narrow Gaussian kernel. i3Control the width of the Gaussian kernel. A smaller bandwidth will cause the kernel function to be too concentrated in a local area, increasing the complexity of the model. i3 -2 , encouraging the model to use a wider kernel function, thereby reducing the complexity of the model and preventing overfitting.

[0079] Understandably, in order to ensure the rationality of the parameters and the stability of the model, we also impose the following constraints: The first constraint: the amplitude coefficient λ i1 , is limited to the interval [0,1], ensuring that its value is always non-negative and does not exceed 1. Center offset λ i2 is limited to min(Y′ i )-δ to max(Y′ i )+δ, where Ensure that the center offset is within the reasonable range of the characteristic distribution. Bandwidth coefficient λ i3 is limited to no less than By ensuring that the kernel function width is large enough and designing the objective function and constraints, when optimizing the model parameters, not only the prediction accuracy is enhanced, but also the sparsity and generalization ability of the model are taken into account, thereby achieving more stable and reliable performance on complex data sets.

[0080] S305. Use the adaptive sample enhancement and overfitting suppression training strategy to improve the generalization ability and optimization effect of the model. Understandably, the second constraint condition is applied: the adaptive sample enhancement and overfitting suppression training strategy dynamically adjusts the sample weights through the residual sensitive sampling mechanism.

[0081] In another specific embodiment, step S305 further includes the following sub-steps:

[0082] S3051, dynamically adjust the sample weight through the residual sensitive sampling mechanism, and calculate the prediction residual ∈ j (t)=|z j (t)-z ture,j |, and according to the formula P j (t+1)∝exp(∈ j (t) / τ) to adjust the sampling weight;

[0083] S3052, adopt kernel function annealing strategy to realize gradual optimization from coarse adjustment to fine adjustment; after each round of iteration, λ i3 Bandwidth factor λ i3 According to λ i3 (t+1)=λ i3 (t)·[1-η·sigmoid(t / T max )] for dynamic contraction, setting η = 0.05 and T max=500 ensures that the bandwidth gradually decreases, enhancing the model's ability to capture detailed features;

[0084] S3053. Monitor the sliding average loss of the validation set during training When the value does not decrease for 20 consecutive rounds, the early stopping mechanism is triggered to terminate the training, and the model parameters with the lowest validation loss are saved as the final output.

[0085] It can be understood that the steps S3051-S3053 are to apply the second constraint: the adaptive sample enhancement and overfitting suppression training strategy dynamically adjusts the sample weights through the residual sensitive sampling mechanism. j (t)=|z j (t)-z ture,j |, according to the formula P j (t+1)∝exp(∈ j (t) / τ), adjust the sampling weight, and the temperature coefficient τ controls the smoothness of the weight distribution, so that the model pays more attention to samples with larger prediction errors.

[0086] Then, the kernel function annealing strategy is used to achieve gradual optimization from coarse adjustment to fine adjustment. After each round of iteration, the bandwidth parameter λ i3 (t) Press λ i3 (t+1)=λ i3 (t)·[1-η·sigmoid(t / T max )] dynamic contraction, where η = 0.05 and T max = 500 ensures that the bandwidth gradually decreases and enhances the model's ability to capture detailed features.

[0087] During training, monitor the sliding average loss of the validation set If the performance does not decrease after 20 consecutive rounds, the early stopping mechanism is triggered to terminate training, and the model parameters with the lowest validation loss are saved as the final output. This adaptive sample enhancement and overfitting suppression training strategy effectively improves the robustness and convergence efficiency of the model by dynamically adjusting sample weights, optimizing kernel function bandwidth, and strict validation monitoring.

[0088] The above steps S304-S305 are the first stage of the staged optimization strategy for model establishment.

[0089] S306. Set the parameters of the multimodal Gaussian kernel adaptive fusion model according to the experimental requirements, use the feature samples as the data set samples, set the ratio of the training set and the validation set, and use the corresponding predicted stress value Z as the output target. After training, the multimodal Gaussian kernel adaptive fusion model is obtained. It can be understood that the feature samples (a total of 7722×10 columns of data) after PCA dimensionality reduction and normalization are used as the input sample library of the multimodal Gaussian kernel adaptive fusion model, and the sample data is divided into 80% for the training set and 20% for the test set. At the same time, the corresponding stress value is used as the target. After training, the multimodal Gaussian kernel adaptive fusion model is obtained.

[0090] All the steps of step S30 construct a multimodal Gaussian kernel adaptive fusion model, which not only retains key information but also improves the generalization ability and computational efficiency of the model, and provides an effective solution for the feature processing of magnetoacoustic emission and magnetic Barkhausen signals.

[0091] In one embodiment, the step S40: using an evaluation function to quantitatively evaluate the prediction effect of the multimodal Gaussian kernel adaptive fusion model, using the following evaluation function, such as:

[0092] The first one is R-squared:

[0093] SS res =∑(y i -f i ) 2 (3)

[0094]

[0095] Among them, y i -f i Indicates 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 coefficient of determination, 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.

[0096] The second type is RMSE (Root Mean Square Error):

[0097]

[0098] Among them, f i -y i It represents the difference between the true value and the predicted value. RMSE reflects the degree of difference between the predicted value and the true value. The smaller the RMSE value, the more stable the model and the higher the reliability.

[0099] It is understandable that after the multimodal Gaussian kernel adaptive fusion model is built, 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, specifically:

[0100] In order to study the performance indicators of MAE single sensor feature fusion stress prediction model, MBN single sensor feature fusion stress prediction model and multimodal Gaussian kernel adaptive fusion model, the fitting diagram of the output results of each model test set (such as Figure 7 The following conclusions were drawn from the analysis:

[0101] Figure 7 Among them, (a) is a multimodal Gaussian kernel adaptive fusion model; (b) is a stress prediction model based on MAE single sensor feature fusion; (c) is a stress prediction model based on MBN single sensor feature fusion; (d) is a multimodal quadratic polynomial fusion model; (e) is a multimodal linear fusion model; (f) is an experiment based on a multimodal exponential fusion model.

[0102] First, we focus on the MAE and MBN single sensor feature fusion methods. Figure 7 As can be seen, the overall error bar of the MBN single sensor is large, and the linear fit is poor. Meanwhile, the MAE single sensor performs differently at different stress stages: the error bar is small but the fit is poor at low stress stages, while the error bar is large but the fit is good at high stress stages. Overall, the MAE single sensor has smaller error bars and better fit than the MBN single sensor, indicating that the MAE single sensor signal is more robust than the MBN single sensor feature fusion algorithm in stress prediction.

[0103] Secondly, for different fusion models based on multi-sensor sensing, the quadratic polynomial fusion model has smaller error bars, better fitting effect, and better overall performance, second only to the Gaussian fusion model. The linear fusion model has poor fitting effect at low stress and larger error bars at high stress, and is the worst among the four fusion models. The exponential fusion model has poor fitting effect at low stress and larger error bars. Overall, the quadratic polynomial outperforms the linear fusion and exponential fusion in terms of error and fitting, and is second only to the Gaussian fusion model. The exponential fusion has poor fitting effect at low stress and larger error bars at high stress, but its overall performance is better than the linear fusion, and the linear model has the worst performance.

[0104] Finally, we focus on the multimodal Gaussian kernel adaptive fusion model. Figure 7 As can be seen in (a), the model performs best in terms of error bars, fitting performance, and accuracy at low, medium, and high stress levels. The stress predictions are highly consistent with the actual results, and the predicted values are generally within the error limits. This demonstrates that multi-sensor feature fusion can effectively improve the stability and accuracy of stress prediction models.

[0105] In summary, different models have their own advantages and disadvantages in stress prediction, and the multimodal Gaussian kernel adaptive fusion model shows obvious advantages in overall performance.

[0106] In another specific embodiment, Figure 8 As shown in the figure, the stress prediction results of different models are compared. Figure 8 The following conclusions can be drawn from the analysis:

[0107] First, observe Figure 8 (a) This figure shows the comparison between the prediction results of the multimodal Gaussian kernel adaptive fusion model (i.e., the multi-sensor Gaussian fusion model) and MAE and MBN single-sensor. From the perspective of the root mean square error (RMSE) indicator, the multimodal Gaussian kernel adaptive fusion model has the smallest RMSE, which is only 7.4149, followed by the MAE single-sensor model, and the MBN single-sensor model has the largest RMSE. This shows that in terms of error, the multimodal Gaussian kernel adaptive fusion model performs best and has the smallest error. From the perspective of the determination coefficient (R 2 )Indicator analysis, R of multimodal Gaussian kernel adaptive fusion model 2 The largest, MAE single sensor model is followed by the MBN single sensor model R 2 This indicates that the multimodal Gaussian kernel adaptive fusion model has the highest stress prediction accuracy, followed by the MAE feature fusion model, and the MBN feature fusion model has the worst accuracy. Overall, the multimodal Gaussian kernel adaptive fusion model is more accurate than the single sensor model.

[0108] Secondly, analysis Figure 8 (b) This figure shows the comparison of prediction results of different multi-sensor models. From the perspective of RMSE index, the multi-sensor Gaussian fusion model also has the smallest RMSE value, followed by the multi-sensor quadratic polynomial fusion model and the multi-sensor exponential fusion model, and the multi-sensor linear fusion model has the largest RMSE. This shows that among the multi-sensor fusion models, the multi-sensor Gaussian fusion model has the smallest error. Looking at R 2 Indicators, R of multi-sensor Gaussian fusion model 2 The largest value is 0.942, followed by the multi-sensor quadratic polynomial fusion model and the multi-sensor exponential fusion model. The R 2 This again shows that the multi-sensor Gaussian fusion model has the highest accuracy.

[0109] In summary, whether compared with the single-sensor model or compared within the multi-sensor fusion model, the multimodal Gaussian kernel adaptive fusion model performs well in terms of error and accuracy and has obvious advantages.

[0110] In another embodiment, Figure 4As shown, the stress detection method based on multimodal Gaussian kernel adaptive fusion also includes building a fusion prediction model. The specific steps include:

[0111] The feature samples are divided into a training set and a validation set. A fusion prediction model is constructed based on the training set, validation set, and multimodal Gaussian kernel product fusion function. The L-BFGS-B optimization algorithm is used to collaboratively optimize the feature parameters, establish a nonlinear mapping relationship between multi-sensor features and stress, and output the stress prediction value Z. The mathematical expression of the Gaussian magnetic-acoustic coupling feature fusion formula is:

[0112]

[0113] Among them, Y' represents the feature after dimension reduction and normalization, Z represents the stress prediction value, and λ i1 Represents the amplitude coefficient, which is used to characterize the global contribution weight of the modal feature, λ i2 represents the center offset, λ i3 Indicates the bandwidth factor.

[0114] Understandably, the input of the fusion prediction model framework is the MAE and MBN time series signals, and the output is the stress prediction value Z. Specifically, during the fusion prediction model construction and verification phase, 80% of the feature sample data after dimensionality reduction and normalization is used as the training set. The fusion prediction model is constructed through the multimodal Gaussian kernel product fusion function, and the L-BFGS-B optimization algorithm is used to collaboratively optimize the feature parameters to establish a nonlinear mapping relationship between multi-sensor features and stress. The remaining 20% of the samples are used as the verification set. The MAE and MBN multi-feature spectrum matrices after dimensionality reduction are input into the trained fusion model, and the stress prediction value Z is output. At the same time, a comparative verification mechanism is set up:

[0115] Maintaining the same algorithmic framework, separate MAE and MBN single-sensor models were constructed. Predictions obtained using different fusion formulas were compared for dimensionality-reduced and normalized magnetic-acoustic splicing multi-feature data. By comparing the prediction results of a single-sensor Gaussian magnetic-acoustic coupling feature fusion formula model with different multi-sensor fusion formula models, the advantages of the multi-sensor Gaussian magnetic-acoustic coupling feature fusion formula model in terms of prediction accuracy, error distribution, and stability were quantitatively evaluated. This cross-validation method of multidimensional feature fusion and single-sensor models under the multimodal Gaussian kernel product fusion function model, as well as multiple comparisons between the multimodal Gaussian kernel product fusion function model and other multi-sensor fusion formula models, has formed an intelligent sensing detection system with fault-tolerant compensation capabilities.

[0116] It should be understood that the order of execution of the steps in the above embodiments does not necessarily mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0117] In one embodiment, a stress detection system based on multimodal Gaussian kernel adaptive fusion is provided. The stress detection system based on multimodal Gaussian kernel adaptive fusion corresponds one-to-one with the stress detection method based on multimodal Gaussian kernel adaptive fusion in the above embodiment. Figure 2 As shown, the stress detection system based on multimodal Gaussian kernel adaptive fusion includes:

[0118] The signal acquisition module 100 is used to obtain the original response signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE of the ferromagnetic specimen, and pre-process the original response signals to obtain the time series signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE;

[0119] The signal processing module 200 extracts a multidimensional characteristic spectrum matrix based on the time series signals of the magnetic Barkhausen noise (MBN) and the magnetic acoustic emission (MAE), and obtains characteristic samples after normalizing the multidimensional characteristic spectrum matrix.

[0120] The model construction module 300 is used to divide the feature samples into a training set and a validation set, construct a multimodal Gaussian kernel adaptive fusion model based on the training set, the validation set and the multimodal Gaussian kernel product fusion function, and use the multimodal Gaussian kernel adaptive fusion model to output the stress prediction value Z;

[0121] The effect evaluation module 400 is used to quantitatively evaluate the prediction effect of the multimodal Gaussian kernel adaptive fusion model using an evaluation function.

[0122] In the present invention, the purpose of the above-mentioned system is to overcome the problems of robustness and poor repeatability of existing stress detection technology. It 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.

[0123] To verify the accuracy of the magnetic-acoustic fusion model in stress assessment, the following process was used:

[0124] First, a set of random tensile stress experiments was designed and implemented. During the experiment, the collected MBN and MAE signals were processed through a series of processes, including signal filtering, feature extraction, and PCA dimensionality reduction. Finally, the principal component comprehensive score with a principal component contribution rate exceeding 95% was obtained. The relationship between these tensile values and the corresponding principal component comprehensive scores is as follows: Figure 9 As shown, it provides basic data for subsequent analysis.

[0125] Then, the 10 principal components of each experiment after dimensionality reduction are input into the trained magnetic-acoustic multi-sensor fusion model. Figure 9 From the comparison between the model prediction value and the true value, it can be seen that the model prediction value can more accurately reflect the trend of stress change, which preliminarily indicates that the model has a certain ability to capture the law of stress change.

[0126] To further quantitatively evaluate the model's performance, detailed calculations were performed between the actual and predicted values for 11 experimental groups. The results showed that the maximum relative error between the two was 18.25%, the minimum relative error was 0.061%, and the average relative error was 9.309%. These data demonstrate that the model has high reliability and predictive accuracy, capable of accurately predicting stress values to a certain extent.

[0127] Based on the above analysis, experimental verification and error analysis have proved that the magnetic-acoustic multi-sensor fusion model proposed in this invention (i.e., the multimodal Gaussian kernel adaptive fusion model) has better performance in stress assessment than other single-sensor stress prediction models. It not only provides a new and effective method for stress assessment, but also promotes the development of intelligent decision-making in the field of stress assessment and prediction, which has important theoretical and practical significance.

[0128] Based on the above analysis, experimental verification and error analysis, the multimodal Gaussian kernel adaptive fusion model outperforms the single sensor feature fusion model and the other three multi-sensor fusion models in performance. The reasons can be explained from the following two aspects:

[0129] First, compared with the single sensing model, this model uses the cross-modal probability density mapping characteristics of the Gaussian kernel to uniformly transform the two heterogeneous signals, MAE (acoustic emission time domain energy) and MBN (magnetic noise frequency domain envelope), into dimensionless contribution weights. i2 Dynamically align the optimal response range of each sensing feature (such as MAE main frequency offset corresponding to dislocation density threshold, MBN envelope area associated with grain size), and combine it with the bandwidth parameter λ i3 Adaptively adjusting the sensitive range effectively reduces the possibility of misjudgment by a single sensor due to local interference (such as electromagnetic noise and mechanical vibration). Experimental prediction results show that the RMSE of the multimodal Gaussian kernel adaptive fusion model improves by 12.99% and 46.73% compared to the MAE single-sensor and MBN single-sensor models, respectively.

[0130] Secondly, compared with quadratic polynomials, due to the dimensional coupling and numerical explosion problems caused by quadratic terms, the failure risk of exponential products for negative value features, and the limitation of linear products that cannot capture nonlinear correlations, the Gaussian kernel fusion model and the multimodal Gaussian kernel adaptive fusion model use probability density mapping to map MAE (mV·s) to MBN (T 2 ) are unified into dimensionless contribution weights, and the parameter λ is used i2Dynamic alignment material performance sensitive threshold (such as the critical point of magnetic domain pinning strength), λ i3 The adaptive shrinkage feature sensitive interval, combined with the nonlinear coupling mechanism in the form of product, achieved good results in the experiment, making the model more repeatable in stress measurement.

[0131] In summary, MAE and MBN only reflect the motion characteristics of a certain type of magnetic domain, and have inherent limitations in stress assessment. Furthermore, fusion methods such as quadratic terms, exponential products, and linear products each have their own shortcomings. The multimodal Gaussian kernel adaptive fusion model, on the other hand, integrates multiple signal characteristics, unifies signals into dimensionless contribution weights through probability density mapping, adaptively shrinks feature-sensitive intervals, and incorporates nonlinear coupling mechanisms, making the model more advantageous in terms of stability, accuracy, repeatability, and robustness.

[0132] For the specific definition of the stress detection system based on multimodal Gaussian kernel adaptive fusion, please refer to the definition of the stress detection method based on multimodal Gaussian kernel adaptive fusion above, which will not be repeated here. The various modules in the above-mentioned stress detection system based on multimodal Gaussian kernel adaptive fusion can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0133] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, which can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0134] Those skilled in the art will clearly understand that for the sake of convenience and brevity in description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0135] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A stress detection method based on multimodal Gaussian kernel adaptive fusion, characterized in that: include: S10, obtaining the original response signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE of the ferromagnetic specimen, and preprocessing the original response signals to obtain the time series signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE; S20, extracting a multidimensional characteristic spectrum matrix based on the time series signals of magnetic Barkhausen noise (MBN) and magnetic acoustic emission (MAE), and obtaining characteristic samples after normalizing the multidimensional characteristic spectrum matrix; S30, dividing the feature samples into a training set and a validation set, constructing a multimodal Gaussian kernel adaptive fusion model based on the training set, the validation set, and the multimodal Gaussian kernel product fusion function, and using the multimodal Gaussian kernel adaptive fusion model to output a stress prediction value Z; S40. Use the evaluation function to quantitatively evaluate the prediction effect of the multimodal Gaussian kernel adaptive fusion model.

2. The stress detection method based on multimodal Gaussian kernel adaptive fusion according to claim 1 is characterized in that: The step S10 includes: S101, generating an original excitation signal based on an excitation source, the original excitation signal is input into a Helmholtz coil after power amplification, and the Helmholtz coil generates a uniform alternating magnetic field as a magnetization excitation source, and excites magnetoelastic waves and magnetic domain dynamic behaviors of the ferromagnetic specimen through a magneto-mechanical-electric coupling mechanism, thereby generating original response signals of magnetic Barkhausen noise (MBN) and magneto-acoustic emission (MAE), respectively; S102, respectively, improving the signal-to-noise ratio of the original response signal of the magnetic Barkhausen noise MBN and the original response signal of the magnetoacoustic emission MAE through a signal amplifier; S103, the acquisition card synchronously acquires the original excitation signal, the original response signal of the amplified magnetic Barkhausen noise MBN, and the original response signal of the amplified magnetoacoustic emission MAE; S104. Use bandpass filters to filter out the background noise of the original response signal of the amplified magnetic Barkhausen noise MBN and the original response signal of the amplified magnetic acoustic emission MAE, calculate the envelope of the filtered signals, and use time-frequency domain joint analysis to obtain the time series signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE.

3. The stress detection method based on multimodal Gaussian kernel adaptive fusion according to claim 2, characterized in that: The step S20 includes: S201, quantifying and extracting four time domain features and four frequency domain features from the original response signal waveforms and envelopes of magnetic Barkhausen noise (MBN) and magnetic acoustic emission (MAE); S202, splicing and fusing the four time-domain features and the four frequency-domain features of magnetic Barkhausen noise (MBN) into an MBN initial feature data set having eight features; splicing and fusing the four time-domain features and the four frequency-domain features of magnetic acoustic emission (MAE) into an MAE initial feature data set having eight features; S203, performing feature concatenation and fusion on the MBN initial feature data set and the MAE initial feature data set to form a joint feature matrix having 16 features; S204. Perform j periodic magnetization experiment iterations on the joint feature matrix with 16 features to construct a feature data set with j×16 dimensions. Use principal component analysis (PCA) to extract main feature parameters, remove redundant information, reduce the dimension of the j×16-dimensional feature data set to j×10 dimensions, and normalize the j×10-dimensional feature data set to obtain a 10-dimensional feature spectrum matrix as a feature sample.

4. The stress detection method based on multimodal Gaussian kernel adaptive fusion according to claim 3 is characterized in that: The step S204 includes: S2041, horizontally splicing the MBN initial feature data set with 8 features of the magnetic Barkhausen noise MBN and the MAE initial feature data set with 8 features of the magnetic acoustic emission MAE according to the sample order, that is, X concat = [F], forming a joint feature matrix with 16 features; S2042. Map the original data of the joint feature matrix to a low-dimensional space through the projection matrix W_PCA∈R^(16*10), and obtain XPCA=Xconcat.WPCA. After dimensionality reduction, the data becomes a 10-dimensional feature spectrum matrix, and the 10-dimensional feature spectrum matrix is normalized.

5. The stress detection method based on multimodal Gaussian kernel adaptive fusion according to claim 3 is characterized in that: The step S30 includes: S301. Apply a multimodal Gaussian kernel product fusion function to map feature samples to the target parameter space. Each Gaussian kernel corresponds to a weak learner, and cross-modal nonlinear coupling is achieved through kernel function multiplication. The mathematical expression of the multimodal Gaussian kernel product fusion function is: Among them, Y' represents the features after dimension reduction and normalization, Y i ' represents the principal component after dimensionality reduction, Z represents the stress prediction value, λ i1 Represents the amplitude coefficient, which is used to characterize the global contribution weight of the modal feature, λ i2 represents the center offset, λ i3 represents the bandwidth factor; S302, pre-train the base model by the feature-kernel function matching method, and according to the mutual information maximization principle, for each principal component Y after dimensionality reduction i 'Assign the optimal Gaussian kernel and filter out the true label Z true Significantly correlated features are used to construct an initial set of base models; S303: After completing feature screening, the parameters of the base model are initialized using an equal-weight initialization method to ensure that all features have the same weight in the initial stage; S304. Use the limited memory Broyden-Fletcher-Goldfarb-Shannon bounded optimization algorithm (L-BFGS-B) to collaboratively optimize the design objective function, which is used to optimize the base model parameters and suppress overfitting, and impose constraints. The mathematical expression of the objective function is: Among them, through the first is the mean square error term, which is used to ensure the prediction accuracy of the model. The second term is a regularization term used to constrain the sparsity of the amplitude coefficient, and the third term It is the reciprocal bandwidth penalty term, which is used to suppress overfitting caused by too narrow Gaussian kernel; S305. Use adaptive sample enhancement and overfitting suppression training strategies to improve the generalization ability and optimization effect of the model; S306. Set the parameters of the multimodal Gaussian kernel adaptive fusion model according to experimental requirements, use the feature samples as data set samples, set the ratio of the training set and the validation set, and use the corresponding predicted stress value Z as the output target. After training, a multimodal Gaussian kernel adaptive fusion model is obtained.

6. The stress detection method based on multimodal Gaussian kernel adaptive fusion according to claim 5, characterized in that: The step S305 includes: S3051, dynamically adjust the sample weight through the residual sensitive sampling mechanism, and calculate the prediction residual ∈ j (t)=|z j (t)-z ture,j |, and according to the formula Adjust sampling weights; S3052, adopt kernel function annealing strategy to realize gradual optimization from coarse adjustment to fine adjustment; after each round of iteration, λ i3 Bandwidth factor λ i3 According to λ i3 (t+1)=λ i3 (t)·[1-η·sigmoid(t / T max )] for dynamic contraction, where η = 0.05 and T max =500 ensures that the bandwidth gradually decreases, enhancing the model's ability to capture detailed features; S3053. Monitor the sliding average loss of the validation set during training When the value does not decrease for 20 consecutive rounds, the early stopping mechanism is triggered to terminate the training, and the model parameters with the lowest validation loss are saved as the final output.

7. The stress detection method based on multimodal Gaussian kernel adaptive fusion according to claim 1, characterized in that: Also includes: The feature samples are divided into training set and validation set, and a fusion prediction model is constructed based on the training set, validation set and multimodal Gaussian kernel product fusion function. The L-BFGS-B optimization algorithm is used to collaboratively optimize the feature parameters, establish a nonlinear mapping relationship between multi-sensor features and stress, and output the stress prediction value Z.

8. A stress detection method and system based on multimodal Gaussian kernel adaptive fusion, characterized in that: include: A signal acquisition module is used to obtain the original response signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE of the ferromagnetic specimen, and pre-process the original response signals to obtain the time series signals of the magnetic Barkhausen noise MBN and the magnetic acoustic emission MAE; The signal processing module extracts a multidimensional characteristic spectrum matrix based on the time series signals of magnetic Barkhausen noise (MBN) and magnetic acoustic emission (MAE), and obtains characteristic samples after normalizing the multidimensional characteristic spectrum matrix. The model construction module is used to divide the feature samples into a training set and a validation set, build a multimodal Gaussian kernel adaptive fusion model based on the training set, the validation set and the multimodal Gaussian kernel product fusion function, and use the multimodal Gaussian kernel adaptive fusion model to output the stress prediction value Z; The effect evaluation module is used to quantitatively evaluate the prediction effect of the multimodal Gaussian kernel adaptive fusion model using an evaluation function.

Images