Ultrasonic guided wave damage positioning method of gated differential attention unit network

Abstract

The invention belongs to the field of ultrasonic guided wave detection, and relates to an ultrasonic guided wave damage positioning method of a gated differential attention unit network, which comprises the following steps: acquiring and preprocessing an ultrasonic guided wave scattering signal xS to obtain an energy density signal xE, and inputting the signals xS and xE into a trained damage positioning model to obtain a damage position; the training process of the damage positioning model comprises the following steps: inputting an ultrasonic guided wave scattering signal xS into a scattering wave feature extraction module to obtain a scattering wave feature fS; inputting the energy density signal xE into an energy density feature extraction module to obtain an energy density feature fE; inputting the features fS and fE into a feature fusion module to obtain a fusion feature; inputting the fused features into a prediction module to obtain a predicted damage position; updating parameters of the damage positioning model according to the predicted damage position until a trained damage positioning model is obtained; the gating differential vector attention unit excavates difference information between two adjacent moments so as to guide updating of a network state, and the model damage positioning capability is improved.

Description

Technical Field

[0001] This invention belongs to the field of ultrasonic guided wave detection and relates to a method for ultrasonic guided wave damage localization using a gated differential attention unit network. Background Technology

[0002] Composite materials are widely used in various complex engineering scenarios due to their advantages such as light weight, high strength, and fatigue resistance. However, under external environmental corrosion and the action of external forces, their internal structure can suffer various hidden damages such as cracking, delamination, and fracture, leading to functional degradation of the material. This can cause property damage and serious personal injury. Therefore, reliable testing techniques must be used to regularly evaluate composite material structures. Ultrasonic guided wave (UGW) based damage detection is a promising method in the existing field of structural damage detection (SHM).

[0003] Existing ultrasonic guided wave-based detection methods can be categorized into physics-based and deep learning-based methods. Traditional physics-based methods typically involve manually extracting damage features from the signal or analyzing damage indices such as correlation coefficients or energy ratios to differentiate the overall difference between the newly recorded signal and the baseline signal. These methods often rely solely on changes in a single damage index to measure the damage location, resulting in the incomplete analysis of the damage information contained within the signal. Consequently, traditional physics-based detection methods exhibit low long-term positioning accuracy and poor robustness when facing noise and clutter interference. Furthermore, echoes reflected from the boundary of the detection area further complicate the use of guided wave signals. In contrast, deep learning-based detection methods avoid manual feature extraction and can adaptively establish a nonlinear high-order mapping relationship between the damage location and the input signal. The network uses these learned mapping relationships to predict the results of unknown label inputs, thereby achieving the prediction of the damage location in composite materials.

[0004] Currently, researchers typically utilize historical data to build statistical models, including Convolutional Neural Networks (CNNs), Graph-in-Graph Convolutional Networks (G-GCNs), Long Short-Term Memory (LSTMs), Dense Convolutional Sparse Coding Networks (DCSCNets), and Multimodal Gated Recurrent Unit Neural Networks. While these methods have been successfully applied to damage localization tasks, they focus primarily on extracting non-temporal features and fail to mine and learn temporal features between different moments, thus limiting the generalization ability and damage localization accuracy of these methods.

[0005] In reality, changes in the amplitude of guided wave signals often reflect the damage status of the area to be detected. In long-distance time series processing, amplitude changes at different times are correlated and dependent. Therefore, the amplitude difference information between each time step is crucial for damage localization tasks. Existing deep learning methods based on ultrasonic guided waves neglect the processing of difference information between two time steps, which leads to problems such as low long-term prediction and localization accuracy, poor network robustness, and limited transferability. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention employs an ultrasonic guided wave damage localization method based on a gated differential attention unit network, comprising: acquiring ultrasonic guided wave scattering signals; preprocessing the ultrasonic guided wave scattering signals to obtain energy density signals; and inputting the ultrasonic guided wave scattering signals and their energy density signals into a trained damage localization model to obtain the predicted damage location. The damage localization model includes: a scattering wave feature extraction module, an energy density feature extraction module, a feature fusion module, and a prediction module. The training process of the damage localization model includes:

[0007] S1. Obtain the ultrasonic guided wave scattering signal dataset, and preprocess the ultrasonic guided wave scattering signal in the ultrasonic guided wave scattering signal dataset to obtain the energy density signal;

[0008] S2. Input the ultrasonic guided wave scattering signal into the scattering wave feature extraction module to obtain the scattering wave features;

[0009] S3. Input the energy density signal into the energy density feature extraction module to obtain the energy density features;

[0010] S4. Input the scattered wave characteristics and energy density characteristics into the feature fusion module to obtain the fused characteristics;

[0011] S5. Input the fused features into the prediction module to obtain the predicted damage location;

[0012] S6. Calculate the loss function value based on the predicted damage location, update the parameters of the damage localization model based on the loss function value, and obtain the trained damage localization model when the loss function value is minimized.

[0013] Beneficial effects:

[0014] 1. The gated differential vector attention unit of this invention effectively mines and learns the difference information between two adjacent time points, generates a differential vector to guide the update of the network state, thereby improving the network's damage localization capability; 2. Based on the gated differential vector attention unit, this invention designs a bidirectional gated differential vector attention unit, which can capture difference information bidirectionally, further improving the network's damage localization capability; 3. This invention combines more damage-characteristic scattered wave signals and energy density signals into the network, which can effectively map the correlation between the damage area and signal features. Attached Figure Description

[0015] Figure 1 A flowchart illustrating an ultrasonic guided wave damage localization method using a gated differential attention unit network, provided in an embodiment of the present invention;

[0016] Figure 2 This is a structural diagram of the scattered wave feature extraction module provided in an embodiment of the present invention;

[0017] Figure 3 This is a structural diagram of the energy density feature extraction module provided in an embodiment of the present invention;

[0018] Figure 4 This is a structural diagram of the gated differential vector attention unit provided in an embodiment of the present invention;

[0019] Figure 5 This is a structural diagram of a bidirectional gated differential vector attention unit provided in an embodiment of the present invention;

[0020] Figure 6 A robustness comparison diagram between different networks on a CFRP board structure provided in an embodiment of the present invention;

[0021] Figure 7 This is a comparison diagram of damage localization tests between different networks on an aluminum plate structure provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this invention employs a gated differential attention unit network-based ultrasonic guided wave damage localization method, comprising: acquiring ultrasonic guided wave scattering signals; preprocessing the ultrasonic guided wave scattering signals to obtain energy density signals; inputting the ultrasonic guided wave scattering signals and their energy density signals into a trained damage localization model to obtain the predicted damage location; the damage localization model includes: a scattering wave feature extraction module, an energy density feature extraction module, a feature fusion module, and a prediction module; the training process of the damage localization model includes:

[0024] S1. Obtain the ultrasonic guided wave scattering signal dataset, preprocess each ultrasonic guided wave scattering signal in the ultrasonic guided wave scattering signal dataset, and obtain the energy density signal of each ultrasonic guided wave scattering signal.

[0025] Because single-mode signals contain limited damage features, the network's ability to represent damage locations is poor. Furthermore, single-mode signals are susceptible to noise. To address this issue, this invention uses scattered wave signals with richer damage information and their corresponding energy density signals as input, combining temporal features and energy information to make the input network's features more diverse and comprehensive.

[0026] Continuous wavelet transform (CWT) can provide correlation damage characteristics of signals in the frequency domain and obtain the local energy time-domain distribution at instantaneous frequency. Therefore, preprocessing of ultrasonic guided wave scattering signals in ultrasonic guided wave scattering signal datasets includes: processing ultrasonic guided wave scattering signals x... S =[x1,x2,...,x T By performing continuous wavelet transform, the wavelet coefficients of the ultrasonic guided wave scattering signal in multiple time intervals are obtained. The square modulus (i.e., energy density) of the wavelet coefficients in each time interval is calculated to obtain the energy density signal x of the ultrasonic guided wave scattering signal. E =[x′1,x′2,...,x′ N ]; where x t Let x′ be the amplitude of the ultrasonic guided wave scattered signal at time t. n Let N be the energy density signal over time period n, T be the length of the ultrasonic guided wave scattering signal, and N be the length of the energy density signal.

[0027] In one embodiment, the ultrasonic guided wave damage signal is subtracted from the healthy signal to obtain the final ultrasonic guided wave scattering signal x. S =[x1,x2,...,x T ].

[0028] The definition of continuous wavelet transform is:

[0029]

[0030] Where C(a,b) represents the continuous wavelet coefficients, which represent the correlation between the signal and the wavelet at different scales, f(t) is the original time-domain signal, and ψ * (t) represents the complex conjugate of the mother wavelet function, a represents the scaling parameter, and b represents the translation parameter.

[0031] Because Morlet wavelets have good smoothness and symmetry, they are suitable for analyzing non-stationary signals. Therefore, the mother wavelet function ψ(t) is a Morlet wavelet. The Morlet wavelet is shown in the following equation.

[0032]

[0033] Among them, f b f represents the frequency bandwidth. c Indicates the center frequency.

[0034] The square modulus of wavelet coefficients represents the distribution of signal energy over different time periods and reflects the signal's energy intensity within a local time frame. Especially in signals passing through damaged regions, the presence of damage in the detection area causes sudden signal fluctuations, leading to changes in local amplitude. This is reflected in the square modulus sequence as a distinct peak. Therefore, this invention uses the square modulus of coefficients obtained from CWT as the energy density signal. The square modulus of wavelet coefficients is defined as follows.

[0035]

[0036] S2. Input the ultrasonic guided wave scattering signal into the scattering wave feature extraction module to obtain the scattering wave features;

[0037] The scattered wave feature extraction module includes a gated differential vector attention unit (GDVU); the gated differential vector attention unit includes a differential operation module, a reset gate, a first update gate, a second update gate, a candidate hidden state calculation module, and a hidden state update module; such as Figure 4 As shown, the gated differential vector attention unit affects the ultrasonic guided wave scattered signal x. S The processing includes:

[0038] S21. Input the amplitude x1 of the ultrasonic guided wave scattering signal at the first time step into the gated differential vector attention unit to obtain the hidden state h1;

[0039] S22. Input the amplitude x2 of the ultrasonic guided wave scattering signal at the second time step and the hidden state h1 into the gated differential vector attention unit to obtain the hidden state h2.

[0040] S23. Input the amplitude x3 of the ultrasonic guided wave scattering signal at the third time step, the hidden state h1, and the hidden state h2 into the gated differential vector attention unit to obtain the hidden state h3.

[0041] S24. The amplitude x of the ultrasonic guided wave scattered signal at the current time step t. t The hidden state h at time step t-1 t-1 And the hidden state h at time step t-2 t-2 By inputting the gated difference vector attention unit, the hidden state h at the current time step t is obtained. t ;

[0042] The amplitude x of the ultrasonic guided wave scattered signal at the current time step t by the gated differential vector attention unit. t The hidden state h at time step t-1 t-1 And the hidden state h at time step t-2 t-2 The processing includes:

[0043] S241, Hide state h t-1 Hidden state h t-2 and amplitude x t Input the difference operation module to obtain the hidden state containing the difference information.

[0044] The difference operation module operates on the hidden state h t-1 Hidden state h t-2 and amplitude x t The processing includes:

[0045] Use the Tanh function to hide the state h t-1 with h t-2 Perform a difference operation to obtain the difference vector DV1; use the Tanh function to convert the hidden state h... t-1 With amplitude x t Perform a difference operation to obtain the difference vector DV2; DV1 represents h. t-1 with h t-2 The degree of difference between them, similarly, DV2 represents h. t-1 With x t The degree of difference between them; the formula is shown below:

[0046] DV1 = tanh(h)t-1 -h t-2 )

[0047] DV2=tanh(x t -h t-1 )

[0048] Interchange vectors DV1 and DV2 with the hidden state h respectively. t-1 h t-2 Multiplying them yields the hidden state containing the difference information. The formula is shown below:

[0049]

[0050] S242, Hide state and amplitude x t Enter the reset button to reveal the hidden state. The retention probability r t The formula is shown below:

[0051]

[0052] Among them, W r U r Let b represent the weight matrix. r Let σ represent the bias matrix, and σ be the Sigmoid function.

[0053] S243, Hide state and amplitude x t Enter the first update gate to get the hidden state. The probability of forgetting z t The formula is shown below:

[0054]

[0055] Among them, W z U z Let b represent the weight matrix. z This represents the bias matrix.

[0056] S244, Hide state and amplitude x t Enter the second update gate to get the hidden state. The probability of forgetting p t The formula is shown below:

[0057]

[0058] Among them, W p U p Let b represent the weight matrix. p This represents the bias matrix.

[0059] S245, Hide state Hidden state The retention probability r t and amplitude x t Input the candidate hidden state into the candidate hidden state calculation module to obtain the candidate hidden state. The formula is shown below:

[0060]

[0061] in, Represents the weight matrix, This represents the bias matrix.

[0062] S246, Hide the candidate state Hidden state Hidden state Hidden state The probability of forgetting z t and hidden state The probability of forgetting p t Input the hidden state update module to obtain the hidden state h at time step t. t The formula is shown below:

[0063]

[0064] S26. Repeat step S25 until the hidden state h at the final time step T is obtained. T That is, the characteristics of scattered waves.

[0065] In one embodiment, such as Figure 2 As shown, the scattered wave feature extraction module also includes: a first batch normalization layer (BN), a second batch normalization layer, and a pooling layer (Max-Pooling); the scattered wave feature extraction module processes the ultrasonic guided wave scattered signal by: inputting the ultrasonic guided wave scattered signal into the first batch normalization layer, inputting the output of the first batch normalization layer into the gated differential vector attention module unit, inputting the output of the gated differential vector attention unit into the second batch normalization layer, and inputting the output of the second batch normalization layer into the pooling layer to obtain the scattered wave features.

[0066] S3. Input the energy density signal into the energy density feature extraction module to obtain the energy density features;

[0067] like Figure 5 As shown, the energy density feature extraction module includes: a bidirectional gated differential vector attention module (BiGDVU); the bidirectional gated differential vector attention module includes: a forward gated differential vector attention unit, a backward gated differential vector attention unit, and an output layer;

[0068] The bidirectional gated differential vector attention module processes the energy density signal including:

[0069] The energy density signal x E Inputting a forward-gated differential vector attention unit yields the energy density signal x. E Forward hidden state at each time step n The formula is shown below:

[0070]

[0071] Where Z represents the forward-gated differential vector attention unit.

[0072] The energy density signal x E The energy density signal x is obtained by inputting the backward-gated differential vector attention unit in reverse order. E Backward hidden state at each time step n The formula is shown below:

[0073]

[0074] Where Z represents the backward-gated differential vector attention unit.

[0075] The energy density signal x E Forward hidden state at each time step and backward hidden state Input-output layer, to obtain energy density signal x E The hidden state h′ at each time step n The state h′ of the last time step N. N As a characteristic of energy density, the formula is shown below:

[0076]

[0077] Among them, w n and v n Let b represent the forward weight matrix and the backward weight matrix, respectively. n This represents the bias matrix.

[0078] In one embodiment, such as Figure 3 As shown, the energy density feature extraction module also includes: a first batch normalization layer (BN), a second batch normalization layer (BN), and a pooling layer (Max-Pooling). The energy density feature extraction module processes the energy density signal by: inputting the energy density signal into the first batch normalization layer, inputting the output of the first batch normalization layer into the bidirectional gated differential vector attention module, inputting the output of the bidirectional gated differential vector attention module into the second batch normalization layer, and inputting the output of the second batch normalization layer into the pooling layer to obtain the energy density features.

[0079] S4. Input the scattered wave characteristics and energy density characteristics into the feature fusion module to obtain the fused characteristics;

[0080] The feature fusion module consists of two linear layers; the feature fusion module processes the scattered wave features and energy density features, including: [processing the scattered wave features f...] S and energy density characteristics f E Inputting the data into two linear layers respectively yields the features. and characteristics Features With the characteristics of scattered waves f S Multiply to obtain the characteristics Features With energy density characteristics f E Multiply to obtain the characteristics Features With features By splicing the data, the fused features are obtained.

[0081] The fusion process is shown in the following formula:

[0082]

[0083]

[0084] Where δ and σ represent the ReLU activation function and the Sigmoid activation function, respectively, W1 and W2 represent the weight matrices of the first and second linear layers, respectively, and H m "+" indicates a multiplication operation, while "Contact" indicates a concatenation operation.

[0085] S5. Input the fused features into the prediction module to obtain the predicted damage location;

[0086] The prediction module consists of three linear layers, and the specific process is shown below.

[0087]

[0088] in, For the predicted damage location, Linear represents a linear layer.

[0089] The specific parameters of the damage localization model are shown in Table 1.

[0090] Table 1. Parameters of the Damage Localization Model

[0091]

[0092] S6. Calculate the loss function value based on the predicted damage location, update the parameters of the damage localization model based on the loss function value, and obtain the trained damage localization model when the loss function value is minimized.

[0093] The loss function is defined as follows:

[0094]

[0095] Among them, y i This represents the actual damage location corresponding to the i-th ultrasonic guided wave scattered signal. Let M be the damage location predicted by the i-th ultrasonic guided wave scattering signal, M represent the total number of samples, i.e. the number of ultrasonic guided wave scattering signals, and MAE be the average error loss function.

[0096] Positioning accuracy is defined as follows:

[0097]

[0098] When the MAE (mean absolute error) between the model's true and predicted values ​​is less than 1.8, the predicted value is considered correct. MAE PR represents the number of samples with correct predictions, and PR represents the prediction accuracy.

[0099] In typical damage detection processes, the acquired signals generally contain clutter and noise. To verify the robustness of this method, five different sets of Gaussian noise (σ∈0, 0.25, 0.125, 0.175, 0.2) were sequentially added to the test set to test the noise resistance of existing deep learning models and the method of this invention under different conditions.

[0100] like Figure 6As shown, when the noise effect is small (σ∈0, 0.025), most networks perform quite well. The accuracy of the method in this invention (Ours), DCSCNet, MGNN (Multi-Graph Neural Network), G-GCN, etc., is around 99%. However, in real-world engineering environments, the impact of noise cannot be ignored. When the influence of noise and clutter suddenly increases (σ∈0.125, 0.175, 0.2), the experimental results of most networks inevitably deteriorate. The classification accuracy of methods such as CNN, FCN, and LSTM inevitably decreases as σ increases; this is because these methods use a single-modal input pattern and cannot extract features from multiple different modalities, resulting in limited noise resistance and poor robustness. The accuracy of DCSCNet and G-GCN also fluctuates to varying degrees with the increase of noise influence; although MGNN has better accuracy than other networks when facing noise interference, MGNN only integrates the ultrasonic guided wave scattering signal received from the sensor and the corresponding continuous wavelet transform signal. Therefore, when σ = 0.2, the accuracy of MGNN can only reach 88.78%. Under the same conditions, the accuracy of the method of this invention can still reach 92.78%. This is because the improved multimodal input mode integrates energy domain features and temporal domain features, so that the interference caused by noise has a limited impact on the feature extraction process. Therefore, the test results are better under harsh noise conditions. The above experiments fully demonstrate that the improved multimodal input mode can significantly improve the robustness and noise resistance of the network. Under various noise conditions, the robustness of the method of this invention surpasses that of existing deep learning methods.

[0101] Table 2 shows the damage location results based on the aluminum plate dataset. The results indicate that methods such as CNN, FCN, and LSTM cannot independently handle the damage location task, despite their short training times. Their mean absolute error and accuracy do not meet the requirements. This is because CNN and FCN networks are poor at processing long-distance sequences, while LSTM ignores the continuity between time points and the changing relationships between adjacent time points, and its inherent noise resistance limits its accuracy. DCSCNet, using an attention mechanism for fusion, achieves good accuracy, but its training time is too long. Multimodal networks (G-GCN, MGNN), even when fusing features from different modalities, still suffer from lower accuracy and generalization ability than the method described in this invention because they neglect the capture and extraction of amplitude difference information between adjacent time points. The amplitude difference information between historical and current time points can effectively characterize the changes in guided waves. Therefore, learning amplitude difference information is indispensable in the damage location task. This invention not only improves the multimodal input mode, but also learns amplitude difference information that other models cannot capture. With this information, the gated differential attention unit can adaptively update the current network parameters, thereby constructing a nonlinear mapping relationship between damage features and damage locations more efficiently.

[0102] Table 2 Comparison of damage localization accuracy among different networks on aluminum plate structures.

[0103]

[0104] Figure 7 The results show the actual location and predicted location of each model under randomly given coordinate labels. It can be seen that, compared with advanced deep learning methods, this invention has better prediction accuracy and superior experimental performance. Furthermore, it achieves excellent experimental results on different plate structures, outperforming existing deep learning methods. This sufficiently demonstrates its excellent transferability and interoperability.

[0105] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for ultrasonic guided wave damage localization using a gated differential attention unit network, characterized in that, include: The ultrasonic guided wave scattering signal is acquired, preprocessed to obtain the energy density signal, and then input into a trained damage localization model to obtain the predicted damage location. The damage localization model includes: a scattered wave feature extraction module, an energy density feature extraction module, a feature fusion module, and a prediction module. The training process of the damage localization model includes: S1. Obtain the ultrasonic guided wave scattering signal dataset, and preprocess the ultrasonic guided wave scattering signal in the ultrasonic guided wave scattering signal dataset to obtain the energy density signal; S2. Input the ultrasonic guided wave scattering signal into the scattering wave feature extraction module to obtain the scattering wave features; S3. Input the energy density signal into the energy density feature extraction module to obtain the energy density features; S4. Input the scattered wave characteristics and energy density characteristics into the feature fusion module to obtain the fused characteristics; S5. Input the fused features into the prediction module to obtain the predicted damage location; S6. Calculate the loss function value based on the predicted damage location, update the parameters of the damage localization model based on the loss function value, and obtain the trained damage localization model when the loss function value is minimized.

2. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 1, characterized in that, Preprocessing of the ultrasonic guided wave scattering signal in the ultrasonic guided wave scattering signal dataset includes: performing continuous wavelet transform on the ultrasonic guided wave scattering signal to obtain the wavelet coefficients of the ultrasonic guided wave scattering signal in multiple time periods, calculating the square modulus of the wavelet coefficients of the ultrasonic guided wave scattering signal in each time period, and obtaining the energy density signal of the ultrasonic guided wave scattering signal.

3. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 1, characterized in that, The scattered wave feature extraction module is a gated differential vector attention unit; the gated differential vector attention unit includes: a differential operation module, a reset gate, a first update gate, a second update gate, a candidate hidden state calculation module, and a hidden state update module; the ultrasonic guided wave scattered signal x S =[x1,x2,...,x T ], x t The ultrasonic guided wave scattered signal x S The amplitude at time step t, where T is the length of the ultrasonic guided wave scattered signal; the gated differential vector attention unit for the ultrasonic guided wave scattered signal x S The processing includes: S21. Input the amplitude x1 of the ultrasonic guided wave scattering signal at the first time step into the gated differential vector attention unit to obtain the hidden state h1; S22. Input the amplitude x2 of the ultrasonic guided wave scattering signal at the second time step and the hidden state h1 into the gated differential vector attention unit to obtain the hidden state h2. S23. Input the amplitude x3 of the ultrasonic guided wave scattering signal at the third time step, the hidden state h1, and the hidden state h2 into the gated differential vector attention unit to obtain the hidden state h3. S24. The amplitude x of the ultrasonic guided wave scattered signal at the current time step t. t The hidden state h at time step t-1 t-1 And the hidden state h at time step t-2 t-2 By inputting the gated difference vector attention unit, the hidden state h at the current time step t is obtained. t ; S25. Repeat step S24 until the hidden state h at the final time step T is obtained. T That is, the characteristics of scattered waves.

4. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 3, characterized in that, The amplitude x of the ultrasonic guided wave scattered signal at the current time step t by the gated differential vector attention unit. t The hidden state h at time step t-1 t-1 And the hidden state h at time step t-2 t-2 The processing includes: S241, Hide state h t-1 Hidden state h t-2 and amplitude x t Input the difference operation module to obtain the hidden state containing the difference information. S242, Hide state and amplitude x t Enter the reset button to reveal the hidden state. The retention probability r t ; S243, Hide state and amplitude x t Enter the first update gate to get the hidden state. The probability of forgetting z t ; S244, Hide state and amplitude x t Enter the second update gate to get the hidden state. The probability of forgetting p t ; S245, Hide state Hidden state The retention probability r t and amplitude x t Input the candidate hidden state into the candidate hidden state calculation module to obtain the candidate hidden state. S246, Hide the candidate state Hidden state Hidden state Hidden state The probability of forgetting z t and hidden state The probability of forgetting p t Input the hidden state update module to obtain the hidden state h at time step t. t .

5. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 4, characterized in that, The difference operation module operates on the hidden state h t-1 Hidden state h t-2 and amplitude x t The processing includes: hiding the state h t-1 with h t-2 Perform a difference operation to obtain the difference vector DV1; then process the hidden state h. t-1 With amplitude x t Perform a difference operation to obtain the difference vector DV2; then combine the difference vector DV1 with the hidden state h. t-1 Multiply to obtain the hidden state containing the difference information. Combine the difference vector DV2 with the hidden state h t-2 Multiply to obtain the hidden state containing the difference information.

6. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 4, characterized in that, The hidden state update module's processing includes:

7. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 1, characterized in that, The bidirectional gated differential vector attention module includes: a forward-gated differential vector attention unit, a backward-gated differential vector attention unit, and an output layer; The bidirectional gated differential vector attention module processes the energy density signal, including: converting the energy density signal x... E Inputting a forward-gated differential vector attention unit yields the energy density signal x. E Forward hidden state at each time step n The energy density signal x E The energy density signal x is obtained by inputting the backward-gated differential vector attention unit in reverse order. E Backward hidden state at each time step n The energy density signal x E Forward hidden state at each time step and backward hidden state Input-output layer, to obtain energy density signal x E The hidden state h at each time step n ′, the state h′ of the last time step N. N As a characteristic of energy density.

8. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 1, characterized in that, The feature fusion module consists of two linear layers; The feature fusion module processes the scattered wave features and energy density features, including: merging the scattered wave features f S and energy density characteristics f E Inputting the data into two linear layers respectively yields the features. and characteristics Features With the characteristics of scattered waves f S Multiply to obtain the characteristics Features With energy density characteristics f E Multiply to obtain the characteristics Features With features By splicing the data, the fused features are obtained.

9. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 8, characterized in that, Two linear layers for scattering wave characteristics f S The processing includes: Where δ and σ represent the ReLU activation function and the Sigmoid activation function, respectively, and W1 and W2 represent the weight matrices of the first and second linear layers, respectively.

10. The ultrasonic guided wave damage localization method using a gated differential attention unit network according to claim 1, characterized in that, The loss function is the average error loss function between the actual damage location and the predicted damage location.

Images