Surface crack detection method and system based on array surface wave

Abstract

The present invention provides a surface crack detection method and system based on array surface waves, which relate to the field of crack detection technology. The method comprises: performing noise reduction and enhancement processing on the original signal collected from the workpiece and establishing an environmental compensation mechanism; extracting key sound path features and crack features from the processed signal; inputting the features into a neural network to predict the crack size; and combining the microscopic dynamics model for theoretical verification. The final detection result is obtained by fusing the predicted result with the theoretical result.

Description

Technical Field

[0001] The present invention relates to the technical field of crack detection, and in particular to a surface crack detection method and system based on array surface waves. Background Art

[0002] As a non-destructive testing method, surface wave testing technology is widely used in surface and near-surface defect detection of metal, composite materials and other structures. The principle is to use the characteristics of ultrasonic wave propagation on the material surface to identify and evaluate defects by analyzing the reflection, scattering and other signals generated by the interaction between surface waves and defects. Compared with body wave testing, surface wave testing is more sensitive to surface defects and has a wider detection range.

[0003] Traditional surface wave detection methods mainly rely on manual experience for signal analysis and defect identification, and have problems such as poor noise resistance, difficulty in defect quantification, and insufficient expansion and prediction capabilities.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a surface crack detection method and system based on array surface waves, which can at least solve some of the problems existing in the prior art.

[0006] According to a first aspect of an embodiment of the present invention, a surface crack detection method based on array surface waves is provided, comprising:

[0007] The reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected are collected, and the reflection signal matrix is ​​probabilistically reconstructed by a variational autoencoder to obtain a latent variable distribution. A Gaussian mixture model is constructed based on the latent variable distribution to obtain a preliminary denoised signal. The preliminary denoised signal is input into a multi-scale densely connected residual network, and an enhanced signal matrix is ​​obtained through a parallel spatial pyramid pooling module and a channel attention module. The environmental compensation feature tensor is generated through a spatiotemporal graph attention neural network.

[0008] Performing a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extracting a valid signal interval in combination with a region growing algorithm, constructing a multi-scale morphological filter bank in the valid signal interval to identify key feature points, extracting a first sound path and a second sound path from the key feature points, training a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and performing variational modal decomposition on the enhanced signal matrix to obtain a multidimensional crack feature vector;

[0009] A hybrid neural network is constructed based on the causal attention mechanism. The compensated sound path value and the multidimensional crack feature vector are input into the hybrid neural network to calculate the initial crack size prediction value and the crack extension trend prediction tensor. The initial crack size prediction value is input into the microdynamic model to obtain the theoretical extension prediction value. The crack extension trend prediction tensor and the theoretical extension prediction value are fused through the unscented Kalman filter algorithm to obtain an accurate detection result.

[0010] In an optional embodiment,

[0011] The reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected are collected, and the reflection signal matrix is ​​probabilistically reconstructed by a variational autoencoder to obtain a latent variable distribution. A Gaussian mixture model is constructed based on the latent variable distribution to obtain a preliminary denoised signal. The preliminary denoised signal is input into a multi-scale densely connected residual network, and an enhanced signal matrix is ​​obtained through a parallel spatial pyramid pooling module and a channel attention module. The environmental compensation feature tensor is generated through a spatiotemporal graph attention neural network.

[0012] Collect ultrasonic phased array scanning data from the surface of the workpiece to be inspected to obtain a reflection signal matrix, and collect surface temperature field, stress field and acoustic emission signals to obtain an environmental physical feature matrix group;

[0013] The reflected signal matrix is ​​input into a variational autoencoder. The encoder consists of a three-layer convolutional neural network. Each convolution layer is connected to a batch normalization layer and a ReLU activation function. The last layer outputs a mean vector and a logarithmic variance vector. Random noise is introduced in the sampling stage and a reparameterization method is used to convert it into a randomly generated network to obtain a latent variable distribution. The decoder uses a symmetric three-layer deconvolution network to reconstruct the signal. The loss function is constructed based on the reconstruction error and the KL divergence for training.

[0014] A Gaussian mixture model is constructed based on the latent variable distribution. The expectation maximization algorithm is used to optimize the parameters. The posterior probability of the sample is calculated and the component parameters are updated. The abnormal signal is identified and eliminated through conditional probability to obtain a preliminary noise reduction signal.

[0015] The preliminary denoised signal is input into a multi-scale densely connected residual network, passed through dense block full connection, a feature compression layer is set, and multi-scale features are extracted through spatial pyramid pooling;

[0016] Global pooling is performed on multi-scale features to obtain channel statistical features, and feature weight coefficients are obtained through multi-layer perceptron and sigmoid function, which are multiplied with the original feature map to obtain the enhanced signal matrix;

[0017] A spatiotemporal graph structure is constructed based on the environmental physical feature matrix group. The temporal and spatial attention scores are calculated through the spatiotemporal graph attention neural network. Information fusion is performed to obtain the spatiotemporal correlation coefficient, and the node features are weighted and updated to generate the environmental compensation feature tensor.

[0018] In an optional embodiment,

[0019] A Gaussian mixture model is constructed based on the latent variable distribution. The expectation maximization algorithm is used to optimize the parameters. The posterior probability of the sample is calculated and the component parameters are updated. The initial noise reduction signal is obtained by identifying and eliminating abnormal signals through conditional probability.

[0020] A Gaussian mixture model is constructed based on the latent variable distribution. Multiple K-means clustering analyses are performed. The silhouette coefficient is calculated through cross-validation. The number of clusters corresponding to the maximum silhouette coefficient is selected as the number of components. The cluster center is set as the initial mean vector. The initial covariance matrix is ​​generated based on the degree of sample dispersion within the cluster. The initial mixing coefficient is set based on the proportion of cluster samples.

[0021] The expectation maximization algorithm is used to optimize the parameters. In the expectation step, the likelihood probability of the sample point under different Gaussian components is calculated, multiplied by the mixing coefficient to obtain the numerator term, and normalized by the sum of the numerator terms of all components to obtain the posterior probability of the sample belonging to different components;

[0022] In the maximization step, the mean vector is updated with the posterior probability as the weight, the outer product matrix of the deviation between the sample point and the mean is calculated and the covariance matrix is ​​updated with the weighted average, and the mixing coefficient is updated based on the average value of the posterior probability;

[0023] Record the parameter changes and stop the iteration when the convergence condition is met or the maximum number of iterations is reached. If the convergence condition is not met and the maximum number of iterations is not exceeded, return to the desired step and continue the iteration.

[0024] The conditional probability of each component of the new input detection signal is calculated, and the signals with probabilities less than the abnormality judgment threshold are marked and eliminated to obtain the preliminary noise reduction signal.

[0025] In an optional embodiment,

[0026] Performing a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extracting an effective signal interval in combination with a region growing algorithm, constructing a multi-scale morphological filter bank in the effective signal interval to identify key feature points, extracting a first sound path and a second sound path from the key feature points, training a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and performing variational mode decomposition on the enhanced signal matrix to obtain a multi-dimensional crack feature vector, including:

[0027] A multi-level filtering structure is constructed to perform wavelet scattering transform. Directional filters are used to collect features of different scales. The filtered signal is convolved with the mother wavelet function to obtain wavelet coefficients. Modulus and average operations are performed on the coefficients to obtain local feature coefficients. The feature coefficients of each level are superimposed according to energy size to construct a multi-scale invariant feature map.

[0028] A seed point is selected in the high-energy area of ​​the feature map, and the adjacent area is searched outward from the seed point. The grayscale co-occurrence frequency of pixels in the area is calculated to obtain the grayscale co-occurrence matrix. The contrast, correlation and entropy values ​​are calculated based on the matrix to construct a texture similarity function. The cost function is constructed by combining the local energy density and iteratively optimized until convergence. The optimal segmentation threshold is obtained to extract the effective signal interval.

[0029] Morphological filter banks are constructed using structural elements of different scales. The dilation-erosion operation is performed to obtain the opening result, and then the erosion-dilation operation is performed to obtain the closing result. The results are cascaded by scale and a dynamic window is set to compare point by point to extract local extreme points.

[0030] Calculating the forward and backward differences of the signal to obtain first- and second-order difference sequences and curvature change rates, analyzing topological features through a persistent homology graph, identifying key feature points through weighted combination with extreme points, extracting the first sound path corresponding to the two largest peaks, performing normalization preprocessing on the environmental compensation feature tensor, inputting it into a Gaussian process regression model, optimizing kernel function parameters using the maximum likelihood method to construct a compensation function, and calculating the compensated sound path based on the first sound path;

[0031] The signal is decomposed by the alternating direction multiplier method, the Hilbert transform is calculated to construct the analytical signal, the spectrum envelope and center frequency are updated, and the modal function group is obtained by iterating until convergence under bandwidth constraints;

[0032] Construct analytical signals to extract instantaneous features and construct Hilbert spectrum, calculate the second-order Rayleigh entropy of signal probability distribution, reconstruct the arrangement pattern to calculate the entropy value, and combine to construct crack feature vector.

[0033] In an optional embodiment,

[0034] The signal is decomposed by the alternating direction multiplier method, the Hilbert transform is calculated to construct the analytical signal, the spectrum envelope and center frequency are updated, and the modal function group is obtained by iterating until convergence under the bandwidth constraint.

[0035] Determining the number of sub-bands to be decomposed based on the spectral characteristics of the enhanced signal matrix, dividing the entire spectral range into a plurality of sub-band intervals according to logarithmic intervals, selecting a point with a local spectral energy maximum within the sub-band interval as an initial center frequency of the current sub-band, substituting the initial center frequency into a Gaussian window function and constructing a frequency domain response characteristic of a bandpass filter, setting an initial bandwidth parameter proportional to the initial center frequency and constructing a bandpass filter bank;

[0036] Filtering the enhanced signal matrix using the bandpass filter bank to obtain an initial mode function corresponding to each sub-band, performing a Hilbert transform on the initial mode function to obtain an analytical signal, and extracting an instantaneous amplitude and an instantaneous phase from the analytical signal, wherein the instantaneous amplitude represents the energy distribution at different moments, and the instantaneous phase represents the frequency modulation characteristics;

[0037] performing bicubic spline interpolation on the instantaneous amplitude to obtain a smoothed spectrum envelope, calculating a spectrum cross-correlation function between adjacent modal components based on the smoothed spectrum envelope to obtain a spectrum overlap, and when the spectrum overlap exceeds a preset overlap threshold, adjusting the position of the initial center frequency and reducing the degree of spectrum aliasing;

[0038] Setting upper and lower bandwidth constraints to control the spectral bandwidth of the modal component based on the initial center frequency; when the spectral bandwidth of the modal component exceeds the upper and lower bandwidth constraints, performing bandwidth correction by scaling the smoothed spectrum envelope; and calculating the spectral energy contribution of the modal component to determine a bandwidth constraint weight;

[0039] The relative changes in the center frequency and bandwidth of each modal component are calculated. When the relative changes are lower than the convergence threshold, the iteration is stopped. The modal components are kept in complete oscillation form and band-limited characteristics in the time domain and frequency domain respectively to obtain an adaptive modal function group.

[0040] In an optional embodiment,

[0041] A hybrid neural network is constructed based on a causal attention mechanism. The compensated sound path value and the multi-dimensional crack feature vector are input into the hybrid neural network to calculate an initial crack size prediction value and a crack growth trend prediction tensor. The initial crack size prediction value is input into a microscopic dynamics model to obtain a theoretical growth prediction value. The crack growth trend prediction tensor and the theoretical growth prediction value are fused through an unscented Kalman filter algorithm to obtain an accurate detection result, including:

[0042] The encoder of the hybrid neural network is constructed based on the convolutional layers, batch normalization layers, and rectified linear units with leakage terms in the multi-layer convolution module. Different filter kernels are used to compensate the sound path value and the multi-dimensional crack feature vector to extract features. The batch normalization layer standardizes the feature distribution. The decoder of the hybrid neural network is constructed through a bidirectional long short-term memory network of the forward hidden layer state and the reverse hidden layer state.

[0043] Construct a query matrix, a key matrix, and a value matrix for the input feature sequence. Take the dot product of the query matrix and the key matrix to get the attention score. After softmax normalization, take the weighted sum with the value matrix to get the context vector. Mask the attention scores at future moments.

[0044] Multi-scale feature maps are extracted through convolution kernels of different sizes. The global average pooling response and maximum pooling response of the channel dimension are calculated and input into a shared multi-layer perceptron to obtain the channel attention weight. The spatial attention weight is obtained by pooling and convolution in the spatial dimension and multiplied by the multi-scale feature map to obtain the multi-scale fusion weight.

[0045] An external memory matrix is ​​established to store historical feature information. The cosine similarity between the input feature vector and each row of the external memory matrix is ​​calculated. The memory row with the highest similarity is selected and weighted averaged to update the memory matrix. The initial crack size prediction value and the crack growth trend prediction tensor are output.

[0046] The hole nucleation is controlled by the stress state, and a nonlinear damage evolution equation related to the number of cycles, stress amplitude and average stress is established. The initial predicted value of the crack size is substituted into the integral to obtain the theoretical expansion prediction value.

[0047] A symmetric sampling point set prediction is generated to obtain the predicted mean and covariance. The crack growth trend prediction tensor and the theoretical growth prediction value are fused through the observation equation and the Kalman gain matrix to obtain the crack size probability density function. The high-order statistical moments and information entropy of the crack size probability density function are calculated to obtain the reliability assessment result. The prediction interval is corrected to obtain the accurate detection result.

[0048] In an optional embodiment,

[0049] The stress state is used to control the nucleation of the cavity. A nonlinear damage evolution equation related to the number of cycles, stress amplitude, and average stress is established. The initial predicted value of the crack size is substituted and integrated to obtain the theoretical expansion prediction value, including:

[0050] Obtain the principal stress components, equivalent stress, and hydrostatic pressure at each location of the material to construct stress state parameters. Characteristic decomposition is used to obtain the principal stresses and principal stress directions. These stresses are then compared with the critical stress for void nucleation, and a time accumulation effect is introduced to control the nucleation of new voids and mark potential nucleation points.

[0051] Extract local stress field information of potential nucleation points, calculate stress gradient to characterize heterogeneity, establish a mapping relationship between stress state and pore deformation rate, introduce shape factor to construct anisotropic growth criterion, and calculate local strain increment based on constitutive relationship to determine pore volume change;

[0052] The interaction between adjacent holes was analyzed based on the hole spacing threshold criterion. The stress field interference was calculated using the stress superposition principle. A stress distribution model in the ligament area was constructed to analyze the stress concentration effect. The stress triaxiality parameter was introduced to establish the hole aggregation criterion, and adjacent holes that met the criterion were merged.

[0053] A nonlinear damage evolution equation related to the number of cycles, stress amplitude, and average stress is established. A damage acceleration factor and damage threshold criterion are introduced. The initial predicted value of the crack size is substituted and high-order numerical integration is used within discrete subintervals to calculate the stress distribution, determine the location of the hole nucleation, update the hole size, determine the aggregation behavior, accumulate fatigue damage, and calculate the state quantity at the next moment.

[0054] The microscopic hole evolution and fatigue damage accumulation are converted into macroscopic crack extension to establish a multi-scale quantitative relationship, and the prediction results are analyzed to obtain a theoretical extension prediction value. The second aspect of the embodiment of the present invention provides a surface crack detection system based on array surface waves, comprising:

[0055] The first unit is used to collect the reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected, probabilistically reconstruct the reflection signal matrix through a variational autoencoder to obtain a latent variable distribution, construct a Gaussian mixture model based on the latent variable distribution to obtain a preliminary denoised signal, input the preliminary denoised signal into a multi-scale densely connected residual network, obtain an enhanced signal matrix through a parallel spatial pyramid pooling module and a channel attention module, and generate an environmental compensation feature tensor through a spatiotemporal graph attention neural network;

[0056] The second unit is configured to perform a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extract a valid signal interval in combination with a region growing algorithm, construct a multi-scale morphological filter bank in the valid signal interval to identify key feature points, extract a first sound path and a second sound path from the key feature points, train a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and perform variational mode decomposition on the enhanced signal matrix to obtain a multidimensional crack feature vector;

[0057] The third unit is used to construct a hybrid neural network based on the causal attention mechanism, input the compensated sound range value and the multi-dimensional crack feature vector into the hybrid neural network to calculate the initial crack size prediction value and the crack extension trend prediction tensor, input the initial crack size prediction value into the microdynamic model to obtain the theoretical extension prediction value, and fuse the crack extension trend prediction tensor and the theoretical extension prediction value through the unscented Kalman filter algorithm to obtain an accurate detection result.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] According to a fourth aspect of the embodiments of the present invention,

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0065] The present invention effectively extracts and enhances crack reflection signals through various signal processing methods, including a multi-scale densely connected residual network, a channel attention module, a wavelet scattering transform, and a multi-scale morphological filter bank. This reduces noise interference, thereby improving the accuracy and reliability of crack detection. A multidimensional crack feature vector is constructed using methods such as variational mode decomposition, Hilbert spectrum, second-order Rayleigh entropy, and permutation entropy, which can more comprehensively describe crack characteristics and further improve detection accuracy. An environmental compensation feature tensor is constructed by combining environmental physical characteristics (temperature field, stress field, and acoustic emission signal). A Gaussian process regression model is used to perform nonlinear compensation for the crack acoustic path, effectively eliminating the influence of environmental factors on crack detection. A hybrid neural network based on a causal attention mechanism, a cross-scale feature fusion module, and a dynamic memory enhancement module can effectively extract long-term dependent features of crack propagation. Combined with a microscopic dynamics model and an unscented Kalman filter algorithm, this method achieves accurate prediction of crack propagation trends. By calculating high-order statistical moments and information entropy of the crack size probability density function, a reliability assessment of crack detection results is provided, providing a more reliable basis for safety assessment of engineering structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Schematic diagram of the process of a surface crack detection method based on array surface waves according to an embodiment of the present invention;

[0067] Figure 2 This is a comparative analysis diagram of the time-frequency characteristics of a surface crack detection method based on array surface waves according to an embodiment of the present invention;

[0068] Figure 3 This is a comparison diagram of the model prediction of the hole evolution of the surface crack detection method based on array surface waves according to an embodiment of the present invention;

[0069] Figure 4 Schematic diagram of the structure of a surface crack detection system based on array surface waves according to an embodiment of the present invention. DETAILED DESCRIPTION

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0071] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0072] Figure 1 FIG. 1 is a flow chart of a surface crack detection method based on array surface waves according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0073] S1. Collect the reflection signal matrix and the environmental physical feature matrix corresponding to the workpiece to be detected, probabilistically reconstruct the reflection signal matrix using a variational autoencoder to obtain a latent variable distribution, construct a Gaussian mixture model based on the latent variable distribution to obtain a preliminary denoised signal, input the preliminary denoised signal into a multi-scale densely connected residual network, obtain an enhanced signal matrix using a parallel spatial pyramid pooling module and a channel attention module, and generate an environmental compensation feature tensor using a spatiotemporal graph attention neural network;

[0074] In an optional embodiment,

[0075] The reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected are collected, and the reflection signal matrix is ​​probabilistically reconstructed by a variational autoencoder to obtain a latent variable distribution. A Gaussian mixture model is constructed based on the latent variable distribution to obtain a preliminary denoised signal. The preliminary denoised signal is input into a multi-scale densely connected residual network, and an enhanced signal matrix is ​​obtained through a parallel spatial pyramid pooling module and a channel attention module. The environmental compensation feature tensor is generated through a spatiotemporal graph attention neural network.

[0076] Collect ultrasonic phased array scanning data from the surface of the workpiece to be inspected to obtain a reflection signal matrix, and collect surface temperature field, stress field and acoustic emission signals to obtain an environmental physical feature matrix group;

[0077] The reflected signal matrix is ​​input into a variational autoencoder. The encoder consists of a three-layer convolutional neural network. Each convolution layer is connected to a batch normalization layer and a ReLU activation function. The last layer outputs a mean vector and a logarithmic variance vector. Random noise is introduced in the sampling stage and a reparameterization method is used to convert it into a randomly generated network to obtain a latent variable distribution. The decoder uses a symmetric three-layer deconvolution network to reconstruct the signal. The loss function is constructed based on the reconstruction error and the KL divergence for training.

[0078] A Gaussian mixture model is constructed based on the latent variable distribution. The expectation maximization algorithm is used to optimize the parameters. The posterior probability of the sample is calculated and the component parameters are updated. The abnormal signal is identified and eliminated through conditional probability to obtain a preliminary noise reduction signal.

[0079] The preliminary denoised signal is input into a multi-scale densely connected residual network, passed through dense block full connection, a feature compression layer is set, and multi-scale features are extracted through spatial pyramid pooling;

[0080] Global pooling is performed on multi-scale features to obtain channel statistical features, and feature weight coefficients are obtained through multi-layer perceptron and sigmoid function, which are multiplied with the original feature map to obtain the enhanced signal matrix;

[0081] A spatiotemporal graph structure is constructed based on the environmental physical feature matrix group. The temporal and spatial attention scores are calculated through the spatiotemporal graph attention neural network. Information fusion is performed to obtain the spatiotemporal correlation coefficient, and the node features are weighted and updated to generate the environmental compensation feature tensor.

[0082] The dense block is part of a densely connected network in deep learning, which improves gradient transfer efficiency and feature reuse capabilities through dense connections.

[0083] Data acquisition was performed on the surface of the workpiece to be inspected. An ultrasonic phased array probe was used to scan the workpiece surface in a grid-scanning pattern, with a scanning interval of 1 mm and a sampling frequency of 100 MHz. The acquired surface wave reflection signal matrix had a dimension of 512 × 512, with each sampling point containing 1024 time series sampling points. Simultaneously, an infrared thermal imager was used to collect surface temperature data at a sampling frequency of 10 Hz; a stress and strain tester was used to collect stress data at a sampling frequency of 1 kHz; and an acoustic emission sensor array was used to collect acoustic emission signals at a sampling frequency of 5 MHz.

[0084] The collected reflection signal matrix is ​​processed using a variational autoencoder. The variational autoencoder uses a three-layer convolutional architecture. The first layer has a 3×3 kernel size, a stride of 2, and outputs 64 feature maps. The second layer has a 3×3 kernel size, a stride of 2, and outputs 128 feature maps. The third layer has a 3×3 kernel size, a stride of 2, and outputs 256 feature maps. Each convolution layer is followed by a batch normalization layer with a momentum of 0.9 and an epsilon of 0.001, followed by a Reluctant Unit (ReLU) activation function. The final encoder layer outputs a 128-dimensional mean vector and log-variance vector, respectively.

[0085] During the sampling phase, random noise from a standard normal distribution is introduced, transforming the deterministic network into a randomly generated network through reparameterization techniques. The decoder uses a three-layer deconvolutional network with a symmetrical structure to the encoder, restoring feature dimensions layer by layer. Training uses a batch size of 64, a learning rate of 0.001, and 100 epochs.

[0086] A Gaussian mixture model was constructed based on the obtained latent variable distribution, with the number of Gaussian components set to 5. Parameter optimization was performed using the expectation-maximization algorithm, with 100 iterations and a convergence threshold of 0.001. For newly input detection signals, the conditional probability of each Gaussian component was calculated. Samples with probabilities below 0.01 were identified as anomalous signals and removed.

[0087] The denoised signal is fed into a multi-scale densely connected residual network. The network consists of four dense blocks, each containing six convolutional layers with a growth rate of 32. A feature compression layer between dense blocks reduces the number of channels to half. The parallel spatial pyramid pooling module is configured at four scales, with pooling kernel sizes of 1×1, 2×2, 4×4, and 8×8, respectively.

[0088] In the channel attention module, a shared multilayer perceptron (MLP) employs a two-layer structure. The number of neurons in the first layer is 1 / 16 of the number of channels, and the second layer restores the original number of channels. After global average pooling and max pooling are performed on the feature map, it is processed by the shared MLP. The outputs of the two branches are added and then passed through the sigmoid function to obtain the weight coefficient.

[0089] A spatiotemporal graph structure is constructed to process environmental physical features. The temporal attention layer uses an 8-head attention mechanism with a sliding window size of 10. The spatial attention layer calculates similarity based on the Euclidean distance between nodes, and the attention scores are normalized using a softmax function. Node features are weighted and updated based on the fused spatiotemporal correlation coefficients to generate an environmental compensation feature tensor.

[0090] In this embodiment, through the combined application of variational autoencoders and Gaussian mixture models, adaptive noise reduction of detection signals is achieved, the signal-to-noise ratio of the signal is effectively improved, and the impact of environmental noise on the detection results is reduced. A multi-scale densely connected residual network combined with a channel attention mechanism is adopted to achieve adaptive extraction and enhancement of multi-scale features, improve feature expression capabilities, and enhance the detection system's ability to recognize defects of different scales. The introduction of a spatiotemporal graph attention neural network to process environmental physical characteristics realizes dynamic compensation of detection results by environmental factors such as temperature fields and stress fields, thereby improving the adaptability and reliability of the detection system in complex environments.

[0091] In an optional embodiment,

[0092] A Gaussian mixture model is constructed based on the latent variable distribution. The expectation maximization algorithm is used to optimize the parameters. The posterior probability of the sample is calculated and the component parameters are updated. The initial noise reduction signal is obtained by identifying and eliminating abnormal signals through conditional probability.

[0093] A Gaussian mixture model is constructed based on the latent variable distribution. Multiple K-means clustering analyses are performed. The silhouette coefficient is calculated through cross-validation. The number of clusters corresponding to the maximum silhouette coefficient is selected as the number of components. The cluster center is set as the initial mean vector. The initial covariance matrix is ​​generated based on the degree of sample dispersion within the cluster. The initial mixing coefficient is set based on the proportion of cluster samples.

[0094] The expectation maximization algorithm is used to optimize the parameters. In the expectation step, the likelihood probability of the sample point under different Gaussian components is calculated, multiplied by the mixing coefficient to obtain the numerator term, and normalized by the sum of the numerator terms of all components to obtain the posterior probability of the sample belonging to different components;

[0095] In the maximization step, the mean vector is updated using the posterior probability as the weight, the outer product matrix of the deviation between the sample point and the mean is calculated and the covariance matrix is ​​updated based on the weighted average, and the mixing coefficient is updated based on the average value of the posterior probability; the parameter changes are recorded, and the iteration is stopped when the convergence condition is met or the maximum number of iterations is reached. If the convergence condition is not met and the maximum number of iterations is not exceeded, the iteration is returned to the expectation step to continue;

[0096] The conditional probability of each component of the new input detection signal is calculated, and the signals with probabilities less than the abnormality judgment threshold are marked and eliminated to obtain the preliminary noise reduction signal.

[0097] The silhouette coefficient is an indicator for evaluating clustering effects, used to measure the separation and closeness of data points between categories, and typically ranges from -1 to 1. The expectation-maximization algorithm is an iterative optimization algorithm used to estimate parameters in models containing latent variables, and is widely used in clustering and probability models. The transposed matrix is ​​a new matrix obtained by exchanging the rows and columns of the original matrix, and has important applications in mathematical computing, signal processing, and other fields.

[0098] To determine the optimal number of Gaussian components in the mixture model, multiple K-means cluster analyses were performed on the underlying distribution of the signal. K-means cluster analysis is a commonly used data clustering method. Its goal is to partition data points into clusters such that data points within a cluster are as similar as possible and data points between clusters are as dissimilar as possible. Each cluster analysis uses a different number of clusters, and the corresponding silhouette coefficient is calculated using cross-validation. The silhouette coefficient is used to assess the quality of the clustering results; a higher value indicates better clustering. The number of clusters with the highest silhouette coefficient is selected as the number of components in the Gaussian mixture model. For example, if a K-means cluster analysis is performed on a set of signal data with the number of clusters set to 2, 3, and 4, and the corresponding silhouette coefficients are 0.6, 0.8, and 0.7, respectively, then the number of clusters is selected as 3. After determining the number of components, the cluster centers are set as the initial mean vectors of each Gaussian component. An initial covariance matrix is ​​generated based on the degree of dispersion of the sample points within each cluster. A greater degree of dispersion results in a larger covariance matrix value. The initial mixing coefficient is set based on the proportion of samples in each cluster. For example, if a cluster contains 50% of the sample points, its corresponding initial mixing coefficient is 0.5.

[0099] The parameters of the Gaussian mixture model are iteratively optimized using the expectation-maximization algorithm. In the expectation step, the posterior probability of each sample point belonging to a different Gaussian component is calculated. For each sample point, the likelihood probability of the sample point under different Gaussian components is calculated based on the mean vector and covariance matrix of the current Gaussian component. The numerator of the likelihood probability is multiplied by the current mixing coefficient of each component, and the denominator is normalized using the sum of the numerator terms of all Gaussian components as the denominator to obtain the posterior probability of each sample belonging to a different Gaussian component.

[0100] In the maximization step, the mean vector, covariance matrix, and mixing coefficients of each component are updated based on the posterior probabilities. The posterior probabilities of the sample points are used as weighting coefficients to calculate the sample means weighted by the posterior probabilities, thereby updating the mean vectors of each component. The deviation vector between each sample point and the updated mean vector is calculated and multiplied by the transposed matrix corresponding to the current sample point to obtain the outer product matrix. The posterior probability weighted average of the outer product matrices of all sample points is performed to obtain the updated covariance matrix. The mixing coefficients are updated based on the average of the posterior probabilities of all sample points for each component.

[0101] After each iteration, the change in model parameters is recorded. Parameter optimization is complete when the parameter change over multiple consecutive iterations is less than the preset convergence threshold. If convergence conditions are not met and the maximum number of iterations has not been exceeded, the algorithm returns to the desired step and continues iterating. For example, if the convergence threshold is set to 0.001 and the maximum number of iterations is set to 100, the algorithm stops if the parameter change over five consecutive iterations is less than 0.001, or if the number of iterations reaches 100.

[0102] For a newly input detection signal, the conditional probability of the signal under different components is calculated based on the optimized Gaussian component parameters. The conditional probability is compared with the preset anomaly determination threshold. If the conditional probability under all Gaussian components is less than the anomaly determination threshold, the detection signal is marked as an anomaly signal and removed from the original signal sequence to obtain a preliminary noise reduction signal. For example, the anomaly determination threshold is set to 0.01. If the conditional probability of a new input signal under all Gaussian components is less than 0.01, it is marked as an anomaly signal and removed.

[0103] In this embodiment, the Gaussian mixture model is used to accurately simulate the potential distribution of the signal, so that abnormal signals can be identified more accurately, thereby improving the noise reduction accuracy. The optimal Gaussian mixture model parameters are automatically learned according to different signal data without human intervention. It has strong adaptability, can effectively identify abnormal signals, and has good robustness.

[0104] S2. performing a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extracting a valid signal interval in combination with a region growing algorithm. Constructing a multi-scale morphological filter bank in the valid signal interval to identify key feature points, extracting a first sound path and a second sound path from the key feature points, training a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and performing variational modal decomposition on the enhanced signal matrix to obtain a multi-dimensional crack feature vector.

[0105] In an optional embodiment,

[0106] Performing a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extracting an effective signal interval in combination with a region growing algorithm, constructing a multi-scale morphological filter bank in the effective signal interval to identify key feature points, extracting a first sound path and a second sound path from the key feature points, training a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and performing variational mode decomposition on the enhanced signal matrix to obtain a multi-dimensional crack feature vector, including:

[0107] A multi-level filtering structure is constructed to perform wavelet scattering transform. Directional filters are used to collect features of different scales. The filtered signal is convolved with the mother wavelet function to obtain wavelet coefficients. Modulus and average operations are performed on the coefficients to obtain local feature coefficients. The feature coefficients of each level are superimposed according to energy size to construct a multi-scale invariant feature map.

[0108] A seed point is selected in the high-energy area of ​​the feature map, and the adjacent area is searched outward from the seed point. The grayscale co-occurrence frequency of pixels in the area is calculated to obtain the grayscale co-occurrence matrix. The contrast, correlation and entropy values ​​are calculated based on the matrix to construct a texture similarity function. The cost function is constructed by combining the local energy density and iteratively optimized until convergence. The optimal segmentation threshold is obtained to extract the effective signal interval.

[0109] Morphological filter banks are constructed using structural elements of different scales. The dilation-erosion operation is performed to obtain the opening result, and then the erosion-dilation operation is performed to obtain the closing result. The results are cascaded by scale and a dynamic window is set to compare point by point to extract local extreme points.

[0110] Calculating the forward and backward differences of the signal to obtain first- and second-order difference sequences and curvature change rates, analyzing topological features through a persistent homology graph, identifying key feature points through weighted combination with extreme points, extracting the first sound path corresponding to the two largest peaks, performing normalization preprocessing on the environmental compensation feature tensor, inputting it into a Gaussian process regression model, optimizing kernel function parameters using the maximum likelihood method to construct a compensation function, and calculating the compensated sound path based on the first sound path;

[0111] The signal is decomposed by the alternating direction multiplier method, the Hilbert transform is calculated to construct the analytical signal, the spectrum envelope and center frequency are updated, and the modal function group is obtained by iterating until convergence under bandwidth constraints;

[0112] Construct analytical signals to extract instantaneous features and construct Hilbert spectrum, calculate the second-order Rayleigh entropy of signal probability distribution, reconstruct the arrangement pattern to calculate the entropy value, and combine to construct crack feature vector.

[0113] The mother wavelet function is the core function of the wavelet transform, which defines the basis function shape and frequency resolution of the analysis signal. The seed point is the reference point initialized in the region growing algorithm, which is used to guide the region expansion process. The pixel gray value co-occurrence frequency is a statistical feature that describes the image texture characteristics and reflects the spatial distribution law of the grayscale relationship between pixels. The dynamic search window refers to a window that dynamically adjusts the search range according to the algorithm requirements, which is used to improve computational efficiency and accuracy. The first-order difference sequence is a sequence composed of the differences between adjacent points in the time series, which is used to reflect the changing trend of the signal. The second-order difference sequence is a sequence composed of the differences between adjacent first-order differences in the time series, which is used to analyze the acceleration or rate of change characteristics of the signal. The persistent homology graph is a topological structure graph used to describe the changes of the signal over time or space, reflecting the stability and evolution of the homology characteristics. The alternating direction multiplier method is an optimization algorithm that iteratively solves complex optimization problems by decomposing them into easy-to-solve sub-problems.

[0114] The collected raw signal is preprocessed, such as removing noise and baseline drift, to obtain an enhanced signal matrix. Taking a signal containing 1000 sampling points as an example, assume that the dimension of the enhanced signal matrix after preprocessing is 1000x1.

[0115] A multi-level filtering structure is constructed to perform a wavelet scattering transform on the enhanced signal matrix. For example, a three-level filtering structure is selected, corresponding to low-frequency, medium-frequency, and high-frequency features, respectively. At each level, directional filters are used to extract features of different scales. Assuming the number of directional filters is 8, 8 filtered signals can be obtained at each level. These filtered signals are convolved with the selected mother wavelet function (such as the Morlet wavelet) to obtain wavelet coefficients. Modulo and average operations are performed on the wavelet coefficients to obtain local feature coefficients. Finally, the local feature coefficients of each level are superimposed in sequence according to the energy size to construct a multi-scale invariant feature map. Assuming that the dimensions of the local feature coefficients obtained at the three levels are 1000x8, 500x8, and 250x8, respectively, the final multi-scale invariant feature map has a dimension of 1000x24.

[0116] Add the multi-scale invariant feature map to the region growing algorithm. Select seed points in the high-energy area of ​​the feature map. Assume that 5 seed points are selected. With the seed point as the center, expand outward layer by layer to search for adjacent areas. Count the co-occurrence frequency of pixel grayscale values ​​in adjacent areas and construct a grayscale co-occurrence matrix. Calculate the contrast, correlation, and entropy of the grayscale co-occurrence matrix, construct a texture similarity function, and calculate the local energy density function. Construct a cost function based on the texture similarity function and the local energy density function, and iteratively update the parameters until convergence to obtain the optimal segmentation threshold. Assume that the optimal segmentation threshold is 0.8. Segment and extract the effective signal interval in the enhanced signal matrix based on this threshold. Assume that the length of the extracted effective signal interval is 500.

[0117] Construct a multiscale morphological filter bank within the valid signal range. For example, a multiscale morphological filter bank can be constructed using 3x3, 5x5, and 7x7 structuring elements. Dilation and erosion operations are performed on the signal at each scale to obtain an opening result. Erosion and dilation operations are then performed on the opening result to obtain a closing result. The opening and closing results are concatenated according to scale, and a dynamic search window is set. Assume that the dynamic search window size is 10. Within the dynamic search window, the signal amplitude is compared point by point, and the local maximum and local minimum points are extracted as local extreme points.

[0118] Calculate the forward and backward differences of the signal sequence to obtain a first-order difference sequence. Calculate the forward and backward differences of the first-order difference sequence to obtain a second-order difference sequence. Calculate the curvature change rate based on the first-order difference sequence and the second-order difference sequence. Construct a persistent homology graph to analyze the signal topology and obtain topological features. Assign weight coefficients to the local extreme points, curvature change rate, and topological features, respectively, and perform weighted superposition to identify key feature points. Assume that 10 key feature points are identified. Extract the first and second sound paths corresponding to the two peaks with the largest amplitudes from the sequence of key feature points. Assume that the first sound path is 100 and the second sound path is 200.

[0119] The environmental compensation feature tensor is input into the Gaussian process regression model. Assume that the environmental compensation feature tensor contains information such as temperature, humidity, and pressure, and has a dimension of 1000x3. Zero-mean and normalization preprocessing is performed on the environmental compensation feature tensor. The scale and noise parameters of the Gaussian kernel function are iteratively optimized using maximum likelihood estimation to construct a nonlinear compensation function. The first and second sound paths are substituted into the nonlinear compensation function to calculate the compensated sound path value.

[0120] Perform variational modal decomposition on the enhanced signal matrix using the alternating direction multiplication method. Initialize the center frequency of each modal component. Bandpass filter the signal in the frequency domain to obtain the initial modal functions. Calculate the Hilbert transform of the initial modal functions and construct the analytical signal. Update the spectral envelope and center frequency of each modal component. Iterate and optimize until convergence, setting frequency bandwidth constraints, to obtain the adaptive modal function set. Assume that three modal functions are obtained.

[0121] An analytical signal is constructed for the adaptive modal function group, and the instantaneous phase and amplitude are extracted. The instantaneous frequency is calculated using the Hilbert transform and combined with the instantaneous amplitude to construct the Hilbert spectrum. The probability distribution characteristics of the signal sequence are statistically analyzed to construct a probability density function, and the second-order Rayleigh entropy is calculated. The signal sequence is reconstructed to generate permutation patterns, and the occurrence probabilities of different patterns are calculated to obtain the permutation entropy. The Hilbert spectrum, second-order Rayleigh entropy, and permutation entropy are combined to construct a multidimensional crack feature vector.

[0122] In this embodiment, through wavelet scattering transform and multi-scale morphological filtering, more comprehensive and detailed crack features can be extracted, and the interference of noise and environmental factors is reduced. The nonlinear compensation function is constructed using the Gaussian process regression model, which can effectively compensate for the influence of environmental factors on the crack sound path and improve the accuracy of the sound path measurement. By combining multiple features such as the Hilbert spectrum, second-order Rayleigh entropy and permutation entropy, a multidimensional crack feature vector is constructed, which provides richer feature information for crack identification and classification.

[0123] In an optional embodiment,

[0124] The signal is decomposed by the alternating direction multiplier method, the Hilbert transform is calculated to construct the analytical signal, the spectrum envelope and center frequency are updated, and the modal function group is obtained by iterating until convergence under the bandwidth constraint.

[0125] Determining the number of sub-bands to be decomposed based on the spectral characteristics of the enhanced signal matrix, dividing the entire spectral range into a plurality of sub-band intervals according to logarithmic intervals, selecting a point with a local spectral energy maximum within the sub-band interval as an initial center frequency of the current sub-band, substituting the initial center frequency into a Gaussian window function and constructing a frequency domain response characteristic of a bandpass filter, setting an initial bandwidth parameter proportional to the initial center frequency and constructing a bandpass filter bank;

[0126] Filtering the enhanced signal matrix using the bandpass filter bank to obtain an initial mode function corresponding to each sub-band, performing a Hilbert transform on the initial mode function to obtain an analytical signal, and extracting an instantaneous amplitude and an instantaneous phase from the analytical signal, wherein the instantaneous amplitude represents the energy distribution at different moments, and the instantaneous phase represents the frequency modulation characteristics;

[0127] performing bicubic spline interpolation on the instantaneous amplitude to obtain a smoothed spectrum envelope, calculating a spectrum cross-correlation function between adjacent modal components based on the smoothed spectrum envelope to obtain a spectrum overlap, and when the spectrum overlap exceeds a preset overlap threshold, adjusting the position of the initial center frequency and reducing the degree of spectrum aliasing;

[0128] Setting upper and lower bandwidth constraints to control the spectral bandwidth of the modal component based on the initial center frequency; when the spectral bandwidth of the modal component exceeds the upper and lower bandwidth constraints, performing bandwidth correction by scaling the smoothed spectrum envelope; and calculating the spectral energy contribution of the modal component to determine a bandwidth constraint weight;

[0129] The relative changes in the center frequency and bandwidth of each modal component are calculated. When the relative changes are lower than the convergence threshold, the iteration is stopped. The modal components are kept in complete oscillation form and band-limited characteristics in the time domain and frequency domain respectively to obtain an adaptive modal function group.

[0130] The Gaussian window function is a weighted function used for localized analysis of signals and has good time-frequency resolution. The bicubic spline interpolation is an interpolation method implemented by fitting bicubic polynomials and is used to smooth data and fill missing values. The smoothed spectrum envelope is a spectrum shape obtained by filtering and smoothing techniques and is used to reflect the overall frequency distribution characteristics of the signal.

[0131] Analyze the spectral characteristics of the enhanced signal matrix to determine the number of sub-bands to be decomposed. Choose an appropriate number of sub-bands based on the frequency range and complexity of the signal. For example, divide the entire spectrum into eight sub-bands at logarithmic intervals.

[0132] Within each sub-band, find the point with the maximum local spectral energy and use it as the initial center frequency of the sub-band. For example, in the first sub-band [10Hz, 20Hz], find the point with the maximum spectral energy at 15Hz and use it as the initial center frequency of the sub-band.

[0133] A bandpass filter bank is constructed using a Gaussian window function combined with the initial center frequency and bandwidth parameters. The initial bandwidth parameter can be set proportional to the initial center frequency, for example, 0.2 times the center frequency. For example, for the first subband, with a center frequency of 15 Hz and an initial bandwidth of 3 Hz, a Gaussian window function with a center frequency of 15 Hz and a bandwidth of 3 Hz is constructed as a bandpass filter.

[0134] The enhanced signal matrix is ​​filtered using the constructed bandpass filter bank to obtain the initial mode function corresponding to each subband. For example, a bandpass filter with a center frequency of 15 Hz is used to filter the signal to obtain the initial mode function corresponding to the first subband. A Hilbert transform is then performed on each initial mode function to obtain the analytical signal. The instantaneous amplitude and phase are extracted from the analytical signal. The instantaneous amplitude reflects the energy distribution at different times, while the instantaneous phase reflects the frequency modulation characteristics.

[0135] Perform bicubic spline interpolation on the extracted instantaneous amplitude to obtain a smooth spectral envelope. For example, perform bicubic spline interpolation on the instantaneous amplitude of the first sub-band to obtain a smooth spectral envelope curve, calculate the spectral cross-correlation function between adjacent modal components, and obtain the spectral overlap. If the spectral overlap exceeds a preset overlap threshold, such as 0.5, the position of the initial center frequency needs to be adjusted to reduce the degree of spectral aliasing. For example, if the spectral overlap between the first and second sub-bands is 0.6, which exceeds the threshold of 0.5, the center frequencies of the first and second sub-bands need to be adjusted, for example, adjusting the center frequency of the first sub-band to 14 Hz and the center frequency of the second sub-band to 17 Hz.

[0136] Bandwidth constraints are set based on the initial center frequency to control the spectral broadening of the modal component. If the spectral bandwidth of a modal component exceeds the upper and lower bandwidth constraints, the bandwidth is corrected by scaling the smoothed spectral envelope. For example, if the bandwidth of the first sub-band exceeds the upper and lower bandwidth constraints, its smoothed spectral envelope needs to be scaled to bring its bandwidth within the allowed range. The bandwidth constraint weight is determined by the spectral energy contribution of the modal component; modal components with greater energy contributions receive a greater bandwidth constraint weight.

[0137] Calculate the relative change in the center frequency and bandwidth of each modal component. If the relative change falls below a convergence threshold, such as 0.01, the iteration is terminated. The resulting modal components maintain a complete oscillation form in the time domain and a band-limited characteristic in the frequency domain, forming an adaptive modal function group.

[0138] In this embodiment, by iteratively optimizing the center frequency and bandwidth, the modal components in the signal can be more accurately extracted, modal aliasing can be reduced, and the decomposition accuracy can be improved. The center frequency and bandwidth can be adaptively adjusted according to the spectral characteristics of the signal. It is applicable to various types of complex signals. By extracting modal components with physical significance, the intrinsic structure and dynamic characteristics of the signal can be better understood, providing more effective information for subsequent signal analysis and processing.

[0139] Figure 2 This figure shows a comparative analysis of the time-frequency characteristics of the surface crack detection method based on array surface waves according to an embodiment of the present invention, demonstrating a comparison between this technical solution and variational mode decomposition (VMD). The horizontal axis represents time (seconds), and the vertical axis represents frequency (Hz). The figure shows the time-varying trajectory of the center frequency of each of the four sub-bands. The solid line represents this technical solution, and the dashed line represents variational mode decomposition. For the first sub-band (about 15Hz), the frequency deviation of this technical solution is only ±0.8Hz, while the frequency deviation of the variational modal decomposition reaches ±3.5Hz, which is 4.4 times that of this solution; for the second sub-band (about 40Hz), the frequency deviation of this technical solution is ±1.3Hz, while the variational modal decomposition is ±4.8Hz, which is 3.7 times that of this solution; for the third sub-band (about 60Hz), the frequency deviation of this technical solution is ±1.7Hz, while the variational modal decomposition is ±6.3Hz, which is 3.7 times that of this solution; for the fourth sub-band (about 80Hz), the frequency deviation of this technical solution is ±2.2Hz, while the variational modal decomposition is ±8.7Hz, which is 4.0 times that of this solution.

[0140] It can be clearly observed from the figure that the center frequency trajectory of this technical solution is smoother and the fluctuation amplitude is significantly smaller than that of the variational modal decomposition method, indicating that it has better time-frequency stability in the processing of non-stationary signals. The bandpass filter constructed by the Gaussian window function used in this technical solution and the bandwidth correction mechanism based on the smoothed spectrum envelope jointly ensure the stability of the frequency characteristics of the modal component in the time dimension. Due to the lack of effective time-frequency stability constraints, the center frequency of the variational modal decomposition method fluctuates greatly when processing non-stationary signals, especially in the high-frequency region. This may lead to inconsistency in the characteristics of the modal component and a decrease in the accuracy of time domain reconstruction. S3. Construct a hybrid neural network based on the causal attention mechanism, input the compensated sound path value and the multi-dimensional crack feature vector into the hybrid neural network to calculate the initial crack size prediction value and the crack extension trend prediction tensor, input the initial crack size prediction value into the micro-dynamic model to obtain the theoretical extension prediction value, and fuse the crack extension trend prediction tensor and the theoretical extension prediction value through the unscented Kalman filter algorithm to obtain an accurate detection result.

[0141] In an optional embodiment,

[0142] A hybrid neural network is constructed based on a causal attention mechanism. The compensated sound path value and the multi-dimensional crack feature vector are input into the hybrid neural network to calculate an initial crack size prediction value and a crack growth trend prediction tensor. The initial crack size prediction value is input into a microscopic dynamics model to obtain a theoretical growth prediction value. The crack growth trend prediction tensor and the theoretical growth prediction value are fused through an unscented Kalman filter algorithm to obtain an accurate detection result, including:

[0143] The encoder of the hybrid neural network is constructed based on the convolutional layers, batch normalization layers, and rectified linear units with leakage terms in the multi-layer convolution module. Different filter kernels are used to compensate the sound path value and the multi-dimensional crack feature vector to extract features. The batch normalization layer standardizes the feature distribution. The decoder of the hybrid neural network is constructed through a bidirectional long short-term memory network of the forward hidden layer state and the reverse hidden layer state.

[0144] Construct a query matrix, a key matrix, and a value matrix for the input feature sequence. Take the dot product of the query matrix and the key matrix to get the attention score. After softmax normalization, take the weighted sum with the value matrix to get the context vector. Mask the attention scores at future moments.

[0145] Multi-scale feature maps are extracted through convolution kernels of different sizes. The global average pooling response and maximum pooling response of the channel dimension are calculated and input into a shared multi-layer perceptron to obtain the channel attention weight. The spatial attention weight is obtained by spatial dimension pooling and convolution, and multiplied by the multi-scale feature map to obtain the multi-scale fusion weight.

[0146] An external memory matrix is ​​established to store historical feature information. The cosine similarity between the input feature vector and each row of the external memory matrix is ​​calculated. The memory row with the highest similarity is selected and weighted averaged to update the memory matrix. The initial crack size prediction value and the crack growth trend prediction tensor are output.

[0147] The hole nucleation is controlled by the stress state, and a nonlinear damage evolution equation related to the number of cycles, stress amplitude and average stress is established. The initial predicted value of the crack size is substituted into the integral to obtain the theoretical expansion prediction value.

[0148] A symmetric sampling point set prediction is generated to obtain the predicted mean and covariance. The crack growth trend prediction tensor and the theoretical growth prediction value are fused through the observation equation and the Kalman gain matrix to obtain the crack size probability density function. The high-order statistical moments and information entropy of the crack size probability density function are calculated to obtain the reliability assessment result. The prediction interval is corrected to obtain the accurate detection result.

[0149] The rectified linear unit with leakage term is a neural network activation function that enables the neural network to better process complex patterns by performing nonlinear mapping on the input. The leakage term avoids the "dead neuron" problem of negative input. The nucleation process refers to the process in which tiny defects (such as holes and cracks) first appear and begin to expand when the material undergoes phase change or fails. It is usually the early stage of material damage or rupture.

[0150] The collected raw ultrasonic signal undergoes preprocessing, including denoising, filtering, and signal enhancement. For example, a wavelet threshold denoising method is used to remove high-frequency noise from the signal, and a bandpass filter is used to extract frequency band information related to the crack. The preprocessed signal is then used to extract a multidimensional crack feature vector and a compensation path value. Assume that the extracted feature vector contains information such as the amplitude, energy, and arrival time of the crack echo, and the compensation path value is 1.5 mm.

[0151] A hybrid neural network was constructed. The encoder portion of the network consists of multiple layers of convolutional modules, each of which includes a convolutional layer, a batch normalization layer, and a rectified linear unit with a leaky term. The convolutional layers use filter kernels of different sizes (for example, 3x3 and 5x5) to extract features from the compensated sound path values ​​and multidimensional crack feature vectors. The batch normalization layer normalizes the distribution of the extracted features to prevent vanishing or exploding gradients. The decoder portion uses a bidirectional long short-term memory network, which uses forward and backward hidden layer states to capture the temporal dependencies of crack propagation.

[0152] A causal attention mechanism is introduced. The input feature sequence is converted into a query matrix, a key matrix, and a value matrix. An attention score is obtained by calculating the dot product of the query matrix and the key matrix, and then normalized using the softmax function. The normalized attention score is weighted and summed with the value matrix to obtain the context vector. To prevent information leakage, the attention scores at future moments are masked so that they are not included in the calculation of the current moment.

[0153] To fuse feature information at different scales, a cross-scale feature fusion module is constructed. Convolution kernels of varying sizes are used to extract input features, generating multi-scale feature maps. The global average pooling response and maximum pooling response in the channel dimension are calculated for each feature map, and these responses are fed into a shared multi-layer perceptron to generate channel attention weights. Simultaneously, the multi-scale feature maps are pooled in the spatial dimension, and spatial attention weights are calculated through a convolutional layer. The channel attention weights and spatial attention weights are multiplied with the multi-scale feature maps in their corresponding dimensions and then added together to generate the final multi-scale fusion weights.

[0154] To extract long-term dependent features of crack growth, a dynamic memory enhancement module is constructed. This module maintains an external memory matrix to store historical feature information. The cosine similarity between the input feature vector and each row of the external memory matrix is ​​calculated, and the memory rows with the highest similarity are weighted averaged to obtain the memory readout. Based on the degree of similarity between the features and the memory, the contents of the external memory matrix are updated. A time decay factor is introduced to reduce the weight of earlier memory content to prevent outdated information from interfering with the current prediction. The module then outputs an initial crack size prediction (e.g., 2.1 mm) and a crack growth trend prediction tensor.

[0155] To achieve more accurate predictions, a microdynamic model based on microscopic void evolution and fatigue damage accumulation was constructed. This model uses the stress state to control the nucleation process of new voids, calculates the volume change of individual voids under stress, and analyzes the interaction between adjacent voids. A nonlinear damage evolution equation was established that is dependent on the number of cycles, stress amplitude, and mean stress. The initial crack size predictions from the neural network were substituted into this equation, and a theoretical extension prediction (e.g., 2.3 mm) was obtained through numerical integration.

[0156] The unscented Kalman filter algorithm is used to fuse the predictions from the neural network and the microdynamic model. The dimensions of the state and observation vectors are determined, and a set of symmetrical sampling points is generated. These sampling points are substituted into the state equation for a one-step prediction, obtaining the predicted mean and covariance. The observed prediction value is calculated using the observation equation. The crack growth trend prediction tensor and the theoretical growth prediction value are fused using the Kalman gain matrix to obtain the crack size probability density function. The higher-order statistical moments and information entropy of this function are calculated to obtain a reliability assessment result. Based on the reliability assessment results, the prediction interval is corrected to obtain the final accurate detection result (for example, 2.2 mm ± 0.1 mm).

[0157] In this embodiment, the advantages of data-driven and physical models are combined to more accurately predict crack size and expansion trends and improve detection accuracy. The causal attention mechanism and dynamic memory enhancement module can effectively capture the temporal dependency and long-term memory information of crack expansion, enhance the robustness of the model, calculate the probability density function of the crack size, and provide reliability assessment results through high-order statistical moments and information entropy, providing more comprehensive information for engineering decision-making.

[0158] In an optional embodiment,

[0159] The stress state is used to control the nucleation of the cavity. A nonlinear damage evolution equation related to the number of cycles, stress amplitude, and average stress is established. The initial predicted value of the crack size is substituted and integrated to obtain the theoretical expansion prediction value, including:

[0160] Obtain the principal stress components, equivalent stress, and hydrostatic pressure at each location of the material to construct stress state parameters. Characteristic decomposition is used to obtain the principal stresses and principal stress directions. These stresses are then compared with the critical stress for void nucleation, and a time accumulation effect is introduced to control the nucleation of new voids and mark potential nucleation points.

[0161] Extract local stress field information of potential nucleation points, calculate stress gradient to characterize heterogeneity, establish a mapping relationship between stress state and pore deformation rate, introduce shape factor to construct anisotropic growth criterion, and calculate local strain increment based on constitutive relationship to determine pore volume change;

[0162] The interaction between adjacent holes was analyzed based on the hole spacing threshold criterion. The stress field interference was calculated using the stress superposition principle. A stress distribution model in the ligament area was constructed to analyze the stress concentration effect. The stress triaxiality parameter was introduced to establish the hole aggregation criterion, and adjacent holes that met the criterion were merged.

[0163] A nonlinear damage evolution equation related to the number of cycles, stress amplitude, and average stress is established. A damage acceleration factor and damage threshold criterion are introduced. The initial predicted value of the crack size is substituted and high-order numerical integration is used within discrete subintervals to calculate the stress distribution, determine the location of the hole nucleation, update the hole size, determine the aggregation behavior, accumulate fatigue damage, and calculate the state quantity at the next moment.

[0164] The microscopic void evolution and fatigue damage accumulation are converted into macroscopic crack extension to establish a multi-scale quantitative relationship, and the prediction results are analyzed to obtain the theoretical extension prediction value.

[0165] The hydrostatic pressure refers to the pressure exerted on an object when the fluid is at rest, which is usually caused by the weight of the liquid, and the pressure is proportional to the depth. It is widely used in fluid mechanics and structural mechanics. The plastic deformation theory is a theory that describes the irreversible deformation of a material when it exceeds the yield strength under the action of an external force. Rheological theory and yield criterion are usually used to characterize plastic behavior. The anisotropic growth criterion is used to describe the law of different crack expansion rates in different directions in a material. It is usually analyzed in combination with the crystal structure of the material and the external stress field. The stress superposition principle means that under the action of a composite load, the stress of a material can be obtained by superimposing the effects of each individual stress source to obtain the total stress. It is widely used in material mechanics. The damage acceleration factor refers to the acceleration multiple of the damage occurrence rate of a material under conditions such as fatigue or corrosion relative to standard conditions. It is used to predict the life of a material under extreme conditions.

[0166] Obtain stress state information at each material location, including principal stress components, equivalent stress, and hydrostatic pressure. Principal stresses and their corresponding principal stress directions are obtained through eigendecomposition. The calculated principal stress values ​​are compared with the material's critical stress for void nucleation, while also considering the cumulative effect of time to identify potential void nucleation locations. For example, for high-strength steel, when the principal stress exceeds 85% of the material's yield strength and persists for more than 100 hours, the location is marked as a potential nucleation point.

[0167] For the identified potential nucleation points, the local stress field distribution information is extracted. The degree of stress distribution inhomogeneity is characterized by calculating the stress gradient. Based on the theory of plastic deformation, a correspondence between the stress state and the hole deformation rate is established. A shape factor is introduced to describe the geometric characteristics of the hole and an anisotropic growth criterion is constructed. The local strain increment is calculated in combination with the constitutive relationship of the material to determine the volume change of a single hole. Taking metal materials as an example, when the local stress gradient exceeds 50MPa / mm, the volume growth rate of the hole will increase significantly.

[0168] To analyze the interaction between adjacent holes, a hole spacing threshold criterion is established. When the distance between two holes is less than three times the material grain size, mutual influence is considered. The stress superposition principle is used to calculate the stress field interference effect, and a ligament zone stress distribution model is constructed to analyze the influence of stress concentration effects on local material deformation. The stress triaxiality parameter is introduced to characterize the multiaxiality of the stress state, and a hole aggregation criterion is established. When the stress in the ligament zone between adjacent holes exceeds 90% of the material's tensile strength, the holes will aggregate and merge into a single equivalent hole.

[0169] A nonlinear damage evolution equation is established that takes into account the number of cycles, stress amplitude, and mean stress. A damage acceleration factor is introduced at high stress levels to describe the rapid material degradation phenomenon. The time domain is discretized into multiple subintervals. Within each subinterval, the stress distribution of the current state is calculated, the location of new void nucleation is determined, the size of existing voids is updated, the aggregation behavior of voids is determined, and fatigue damage is accumulated. For example, when the cyclic stress amplitude exceeds 75% of the material's fatigue limit, the damage acceleration factor is set to 2.5, accelerating the rate of damage accumulation.

[0170] Microscopic void evolution and fatigue damage accumulation are converted into macroscopic crack expansion, establishing a multi-scale quantitative relationship. The predicted expansion rate, acceleration, and stability are analyzed to obtain a theoretical expansion prediction. In practical applications, when the cumulative damage value reaches 0.8, the material enters a rapid expansion phase.

[0171] In this embodiment, by establishing a correlation mechanism between stress state and pore nucleation and evolution, accurate prediction of the initiation and development process of material damage is achieved, and the accuracy and reliability of the prediction results are improved. The interaction relationship between pores is taken into account, and stress field interference and ligament zone stress distribution models are introduced to more realistically reflect the damage evolution law of the material, overcome the limitation of traditional methods that ignore the influence of microstructure, establish a nonlinear damage evolution equation including the number of cycles, stress amplitude and average stress, and describe the rapid degradation phenomenon under high stress levels through the damage acceleration factor, thereby improving the prediction accuracy under complex service conditions and providing reliable theoretical guidance for practical engineering applications.

[0172] Figure 3 This figure compares the void volume evolution predicted by the model of the surface crack detection method based on array surface waves in an embodiment of the present invention. It shows the comparison between the void volume evolution predicted by this technical solution and the traditional void growth model under different cycle numbers and average stress levels. The horizontal axis represents the cycle number (0-5000), and the vertical axis represents the average stress level (0-300 MPa).

[0173] The contour lines in the figure represent the normalized pore volume (V / V0, the ratio relative to the initial volume). The solid line represents the predictions of our proposed solution, while the dashed line represents the predictions of the traditional pore growth model. Several key features can be observed from the contour distribution. The contour spacing predicted by our proposed solution decreases with increasing cycle number and stress level, indicating a significant acceleration in pore volume growth under high stress and high cycle number conditions. In contrast, the contour spacing predicted by the traditional model varies only slightly, failing to fully reflect this nonlinear acceleration effect. For example, at a stress level of 150 MPa and 5000 cycles, the normalized pore volume predicted by our proposed solution reaches 6.0, while that of the traditional model is only around 3.0, a difference of approximately 100%. Comparing the intersection points of the contour lines of the two models reveals that the predictions of the two models are relatively close in the low stress region (<50 MPa) or low cycle number (<1000), but the difference rapidly increases with increasing stress and cycle number. In particular, in the 150-200 MPa stress region in the upper right corner of the figure, the void volume growth rate predicted by this technical solution is significantly higher than that of traditional models, more consistent with the damage acceleration phenomenon of actual materials under high-stress cyclic loading. The shaded area in the figure highlights the "accelerated growth region," demonstrating that this technical solution, by introducing stress state parameters, stress gradients, and shape factors to construct an anisotropic growth criterion, can more accurately capture the nonlinear acceleration characteristics of void evolution under high stress conditions, which is of great significance for the precise prediction of fatigue life of engineering materials.

[0174] Figure 4 FIG. 1 is a schematic structural diagram of a surface crack detection system based on array surface waves according to an embodiment of the present invention. Figure 4 As shown, the system includes:

[0175] The first unit is used to collect the reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected, probabilistically reconstruct the reflection signal matrix through a variational autoencoder to obtain a latent variable distribution, construct a Gaussian mixture model based on the latent variable distribution to obtain a preliminary denoised signal, input the preliminary denoised signal into a multi-scale densely connected residual network, obtain an enhanced signal matrix through a parallel spatial pyramid pooling module and a channel attention module, and generate an environmental compensation feature tensor through a spatiotemporal graph attention neural network;

[0176] The second unit is configured to perform a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extract a valid signal interval in combination with a region growing algorithm, construct a multi-scale morphological filter bank in the valid signal interval to identify key feature points, extract a first sound path and a second sound path from the key feature points, train a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and perform variational mode decomposition on the enhanced signal matrix to obtain a multidimensional crack feature vector;

[0177] The third unit is used to construct a hybrid neural network based on the causal attention mechanism, input the compensated sound range value and the multi-dimensional crack feature vector into the hybrid neural network to calculate the initial crack size prediction value and the crack extension trend prediction tensor, input the initial crack size prediction value into the microdynamic model to obtain the theoretical extension prediction value, and fuse the crack extension trend prediction tensor and the theoretical extension prediction value through the unscented Kalman filter algorithm to obtain an accurate detection result.

[0178] According to a third aspect of the embodiments of the present invention,

[0179] An electronic device is provided, comprising:

[0180] processor;

[0181] a memory for storing processor-executable instructions;

[0182] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0183] According to a fourth aspect of the embodiments of the present invention,

[0184] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0185] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A surface crack detection method based on array surface waves, characterized in that: include: The reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected are collected, and the reflection signal matrix is ​​probabilistically reconstructed by a variational autoencoder to obtain a latent variable distribution. A Gaussian mixture model is constructed based on the latent variable distribution to obtain a preliminary denoised signal. The preliminary denoised signal is input into a multi-scale densely connected residual network, and an enhanced signal matrix is ​​obtained through a parallel spatial pyramid pooling module and a channel attention module. The environmental compensation feature tensor is generated through a spatiotemporal graph attention neural network. Performing a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extracting an effective signal interval in combination with a region growing algorithm, constructing a multi-scale morphological filter bank in the effective signal interval to identify key feature points, extracting a first sound path and a second sound path from the key feature points, training a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and performing variational mode decomposition on the enhanced signal matrix to obtain a multi-dimensional crack feature vector, including: A multi-level filtering structure is constructed to perform wavelet scattering transform. Directional filters are used to collect features of different scales. The filtered signal is convolved with the mother wavelet function to obtain wavelet coefficients. Modulus and average operations are performed on the coefficients to obtain local feature coefficients. The feature coefficients of each level are superimposed according to energy size to construct a multi-scale invariant feature map. A seed point is selected in the high-energy area of ​​the feature map, and the adjacent area is searched outward from the seed point. The grayscale co-occurrence frequency of pixels in the area is calculated to obtain the grayscale co-occurrence matrix. The contrast, correlation and entropy values ​​are calculated based on the matrix to construct a texture similarity function. The cost function is constructed by combining the local energy density and iteratively optimized until convergence. The optimal segmentation threshold is obtained to extract the effective signal interval. Morphological filter banks are constructed using structural elements of different scales. The dilation-erosion operation is performed to obtain the opening result, and then the erosion-dilation operation is performed to obtain the closing result. The results are cascaded by scale and a dynamic window is set to compare point by point to extract local extreme points. Calculating the forward and backward differences of the signal to obtain first- and second-order difference sequences and curvature change rates, analyzing topological features through a persistent homology graph, identifying key feature points through weighted combination with extreme points, extracting the first sound path corresponding to the two largest peaks, performing normalization preprocessing on the environmental compensation feature tensor, inputting it into a Gaussian process regression model, optimizing kernel function parameters using the maximum likelihood method to construct a compensation function, and calculating the compensated sound path based on the first sound path; The signal is decomposed by the alternating direction multiplier method, the Hilbert transform is calculated to construct the analytical signal, the spectrum envelope and center frequency are updated, and the modal function group is obtained by iterating until convergence under bandwidth constraints; Construct analytical signals to extract instantaneous features and construct Hilbert spectrum, calculate the second-order Rayleigh entropy of signal probability distribution, reconstruct the arrangement pattern to calculate the entropy value, and combine to construct crack feature vector; A hybrid neural network is constructed based on the causal attention mechanism. The compensated sound path value and the multidimensional crack feature vector are input into the hybrid neural network to calculate the initial crack size prediction value and the crack extension trend prediction tensor. The initial crack size prediction value is input into the microdynamic model to obtain the theoretical extension prediction value. The crack extension trend prediction tensor and the theoretical extension prediction value are fused through the unscented Kalman filter algorithm to obtain an accurate detection result.

2. The method according to claim 1, characterized in that The reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected are collected, and the reflection signal matrix is ​​probabilistically reconstructed by a variational autoencoder to obtain a latent variable distribution. A Gaussian mixture model is constructed based on the latent variable distribution to obtain a preliminary denoised signal. The preliminary denoised signal is input into a multi-scale densely connected residual network, and an enhanced signal matrix is ​​obtained through a parallel spatial pyramid pooling module and a channel attention module. The environmental compensation feature tensor is generated through a spatiotemporal graph attention neural network. Collect ultrasonic phased array scanning data from the surface of the workpiece to be inspected to obtain a reflection signal matrix, and collect surface temperature field, stress field and acoustic emission signals to obtain an environmental physical feature matrix group; The reflected signal matrix is ​​input into a variational autoencoder. The encoder consists of a three-layer convolutional neural network. Each convolution layer is connected to a batch normalization layer and a ReLU activation function. The last layer outputs a mean vector and a logarithmic variance vector. Random noise is introduced in the sampling stage and a reparameterization method is used to convert it into a randomly generated network to obtain a latent variable distribution. The decoder uses a symmetric three-layer deconvolution network to reconstruct the signal. The loss function is constructed based on the reconstruction error and the KL divergence for training. A Gaussian mixture model is constructed based on the latent variable distribution. The expectation maximization algorithm is used to optimize the parameters. The posterior probability of the sample is calculated and the component parameters are updated. The abnormal signal is identified and eliminated through conditional probability to obtain a preliminary noise reduction signal. The preliminary denoised signal is input into a multi-scale densely connected residual network, passed through dense block full connection, a feature compression layer is set, and multi-scale features are extracted through spatial pyramid pooling; Global pooling is performed on multi-scale features to obtain channel statistical features, and feature weight coefficients are obtained through multi-layer perceptron and sigmoid function, which are multiplied with the original feature map to obtain the enhanced signal matrix; A spatiotemporal graph structure is constructed based on the environmental physical feature matrix group. The temporal and spatial attention scores are calculated through the spatiotemporal graph attention neural network. Information fusion is performed to obtain the spatiotemporal correlation coefficient, and the node features are weighted and updated to generate the environmental compensation feature tensor.

3. The method according to claim 2, characterized in that A Gaussian mixture model is constructed based on the latent variable distribution. The expectation maximization algorithm is used to optimize the parameters. The posterior probability of the sample is calculated and the component parameters are updated. The initial noise reduction signal is obtained by identifying and eliminating abnormal signals through conditional probability. A Gaussian mixture model is constructed based on the latent variable distribution. Multiple K-means clustering analyses are performed. The silhouette coefficient is calculated through cross-validation. The number of clusters corresponding to the maximum silhouette coefficient is selected as the number of components. The cluster center is set as the initial mean vector. The initial covariance matrix is ​​generated based on the degree of sample dispersion within the cluster. The initial mixing coefficient is set based on the proportion of cluster samples. The expectation maximization algorithm is used to optimize the parameters. In the expectation step, the likelihood probability of the sample point under different Gaussian components is calculated, multiplied by the mixing coefficient to obtain the numerator term, and normalized by the sum of the numerator terms of all components to obtain the posterior probability of the sample belonging to different components; In the maximization step, the mean vector is updated with the posterior probability as the weight, the outer product matrix of the deviation between the sample point and the mean is calculated and the covariance matrix is ​​updated with the weighted average, and the mixing coefficient is updated based on the average value of the posterior probability; Record the parameter changes and stop the iteration when the convergence condition is met or the maximum number of iterations is reached. If the convergence condition is not met and the maximum number of iterations is not exceeded, return to the desired step and continue the iteration. The conditional probability of each component of the new input detection signal is calculated, and the signals with probabilities less than the abnormality judgment threshold are marked and eliminated to obtain the preliminary noise reduction signal.

4. The method according to claim 1, wherein The signal is decomposed by the alternating direction multiplier method, the Hilbert transform is calculated to construct the analytical signal, the spectrum envelope and center frequency are updated, and the modal function group is obtained by iterating until convergence under the bandwidth constraint. Determining the number of sub-bands to be decomposed based on the spectral characteristics of the enhanced signal matrix, dividing the entire spectral range into a plurality of sub-band intervals according to logarithmic intervals, selecting a point with a local spectral energy maximum within the sub-band interval as an initial center frequency of the current sub-band, substituting the initial center frequency into a Gaussian window function and constructing a frequency domain response characteristic of a bandpass filter, setting an initial bandwidth parameter proportional to the initial center frequency and constructing a bandpass filter bank; Filtering the enhanced signal matrix using the bandpass filter bank to obtain an initial mode function corresponding to each sub-band, performing a Hilbert transform on the initial mode function to obtain an analytical signal, and extracting an instantaneous amplitude and an instantaneous phase from the analytical signal, wherein the instantaneous amplitude represents the energy distribution at different moments, and the instantaneous phase represents the frequency modulation characteristics; performing bicubic spline interpolation on the instantaneous amplitude to obtain a smoothed spectrum envelope, calculating a spectrum cross-correlation function between adjacent modal components based on the smoothed spectrum envelope to obtain a spectrum overlap, and when the spectrum overlap exceeds a preset overlap threshold, adjusting the position of the initial center frequency and reducing the degree of spectrum aliasing; Setting upper and lower bandwidth constraints to control the spectral bandwidth of the modal component based on the initial center frequency; when the spectral bandwidth of the modal component exceeds the upper and lower bandwidth constraints, performing bandwidth correction by scaling the smoothed spectrum envelope; and calculating the spectral energy contribution of the modal component to determine a bandwidth constraint weight; The relative changes in the center frequency and bandwidth of each modal component are calculated. When the relative changes are lower than the convergence threshold, the iteration is stopped. The modal components are kept in complete oscillation form and band-limited characteristics in the time domain and frequency domain respectively to obtain an adaptive modal function group.

5. The method according to claim 1, wherein A hybrid neural network is constructed based on a causal attention mechanism. The compensated sound path value and the multi-dimensional crack feature vector are input into the hybrid neural network to calculate an initial crack size prediction value and a crack growth trend prediction tensor. The initial crack size prediction value is input into a microscopic dynamics model to obtain a theoretical growth prediction value. The crack growth trend prediction tensor and the theoretical growth prediction value are fused through an unscented Kalman filter algorithm to obtain an accurate detection result, including: The encoder of the hybrid neural network is constructed based on the convolutional layers, batch normalization layers, and rectified linear units with leakage terms in the multi-layer convolution module. Different filter kernels are used to compensate the sound path value and the multi-dimensional crack feature vector to extract features. The batch normalization layer standardizes the feature distribution. The decoder of the hybrid neural network is constructed through a bidirectional long short-term memory network of the forward hidden layer state and the reverse hidden layer state. Construct a query matrix, a key matrix, and a value matrix for the input feature sequence. Take the dot product of the query matrix and the key matrix to get the attention score. After softmax normalization, take the weighted sum with the value matrix to get the context vector. Mask the attention scores at future moments. Multi-scale feature maps are extracted through convolution kernels of different sizes. The global average pooling response and maximum pooling response of the channel dimension are calculated and input into a shared multi-layer perceptron to obtain the channel attention weight. The spatial attention weight is obtained by pooling and convolution in the spatial dimension and multiplied by the multi-scale feature map to obtain the multi-scale fusion weight. An external memory matrix is ​​established to store historical feature information. The cosine similarity between the input feature vector and each row of the external memory matrix is ​​calculated. The memory row with the highest similarity is selected and weighted averaged to update the memory matrix. The initial crack size prediction value and the crack growth trend prediction tensor are output. The hole nucleation is controlled by the stress state, and a nonlinear damage evolution equation related to the number of cycles, stress amplitude and average stress is established. The initial predicted value of the crack size is substituted into the integral to obtain the theoretical expansion prediction value. A symmetric sampling point set prediction is generated to obtain the predicted mean and covariance. The crack growth trend prediction tensor and the theoretical growth prediction value are fused through the observation equation and the Kalman gain matrix to obtain the crack size probability density function. The high-order statistical moments and information entropy of the crack size probability density function are calculated to obtain the reliability assessment result. The prediction interval is corrected to obtain the accurate detection result.

6. The method according to claim 5, characterized in that The stress state is used to control the nucleation of the cavity. A nonlinear damage evolution equation related to the number of cycles, stress amplitude, and average stress is established. The initial predicted value of the crack size is substituted and integrated to obtain the theoretical expansion prediction value, including: Obtain the principal stress components, equivalent stress, and hydrostatic pressure at each location of the material to construct stress state parameters. Characteristic decomposition is used to obtain the principal stresses and principal stress directions. These stresses are then compared with the critical stress for void nucleation, and a time accumulation effect is introduced to control the nucleation of new voids and mark potential nucleation points. Extract local stress field information of potential nucleation points, calculate stress gradient to characterize heterogeneity, establish a mapping relationship between stress state and pore deformation rate, introduce shape factor to construct anisotropic growth criterion, and calculate local strain increment based on constitutive relationship to determine pore volume change; The interaction between adjacent holes was analyzed based on the hole spacing threshold criterion. The stress field interference was calculated using the stress superposition principle. A stress distribution model in the ligament area was constructed to analyze the stress concentration effect. The stress triaxiality parameter was introduced to establish the hole aggregation criterion, and adjacent holes that met the criterion were merged. A nonlinear damage evolution equation related to the number of cycles, stress amplitude, and average stress is established. A damage acceleration factor and damage threshold criterion are introduced. The initial predicted value of the crack size is substituted and high-order numerical integration is used within discrete subintervals to calculate the stress distribution, determine the location of the hole nucleation, update the hole size, determine the aggregation behavior, accumulate fatigue damage, and calculate the state quantity at the next moment. The microscopic void evolution and fatigue damage accumulation are converted into macroscopic crack extension to establish a multi-scale quantitative relationship, and the prediction results are analyzed to obtain the theoretical extension prediction value.

7. A surface crack detection system based on array surface waves, for implementing the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to collect the reflection signal matrix and the environmental physical feature matrix group corresponding to the workpiece to be detected, probabilistically reconstruct the reflection signal matrix through a variational autoencoder to obtain a latent variable distribution, construct a Gaussian mixture model based on the latent variable distribution to obtain a preliminary denoised signal, input the preliminary denoised signal into a multi-scale densely connected residual network, obtain an enhanced signal matrix through a parallel spatial pyramid pooling module and a channel attention module, and generate an environmental compensation feature tensor through a spatiotemporal graph attention neural network; The second unit is configured to perform a wavelet scattering transform on the enhanced signal matrix to obtain a multi-scale invariant feature map and extract a valid signal interval in combination with a region growing algorithm, construct a multi-scale morphological filter bank in the valid signal interval to identify key feature points, extract a first sound path and a second sound path from the key feature points, train a Gaussian process regression model based on the environmental compensation feature tensor to obtain a compensated sound path value, and perform variational mode decomposition on the enhanced signal matrix to obtain a multidimensional crack feature vector; The third unit is used to construct a hybrid neural network based on the causal attention mechanism, input the compensated sound range value and the multi-dimensional crack feature vector into the hybrid neural network to calculate the initial crack size prediction value and the crack extension trend prediction tensor, input the initial crack size prediction value into the microdynamic model to obtain the theoretical extension prediction value, and fuse the crack extension trend prediction tensor and the theoretical extension prediction value through the unscented Kalman filter algorithm to obtain an accurate detection result.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images