Acoustic-optical-electric synergistic welding material fatigue performance testing method and system

Abstract

The invention relates to the technical field of welding material performance testing, and discloses an acoustic-optical-electric synergistic welding material fatigue performance testing method and system, and the method comprises the steps: synchronously collecting a first multi-mode signal data set of a welding material in a fatigue testing process; carrying out fusion processing to obtain a fusion feature vector; performing time sequence feature extraction to obtain time sequence comprehensive representation; constructing and training an initial fatigue performance multi-task learning model comprising a shared feature extractor and a task-specific sub-network based on time sequence comprehensive representation; implementing domain adaptation and transfer learning on the fatigue performance multi-task learning model to obtain a target fatigue performance multi-task learning model; acquiring a second multi-mode signal data set in real time and inputting the second multi-mode signal data set into the target fatigue performance multi-task learning model to obtain a welding material fatigue performance test result, so that acoustic, optical and electric multi-mode signals can be effectively integrated, and high-precision, multi-task and migratable welding material performance test is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of welding material performance testing, and in particular to an acoustic, optical and electrical coordinated welding material fatigue performance testing method and system. Background Art

[0002] Traditional fatigue performance testing methods mainly rely on single-mode measurement data, such as stress-strain curves or acoustic emission signals, which are difficult to fully capture the complex changes in the material fatigue process. With the advancement of sensing technology and signal processing methods, multi-modal signal fusion has become an effective means to improve the accuracy and reliability of fatigue performance testing.

[0003] However, the high dimensionality, nonlinearity, and time-varying characteristics of multimodal signal data pose great challenges to data processing and feature extraction. Existing feature extraction methods often have difficulty effectively capturing the correlation and complementary information between different modal signals, resulting in poor fusion effects. In addition, traditional machine learning models often face the problem of insufficient generalization ability and difficulty in adapting to different materials and test conditions when dealing with high-dimensional, multi-task fatigue performance prediction problems. Summary of the invention

[0004] The present invention provides an acoustic, optical and electrical coordinated welding material fatigue performance testing method and system, which can effectively integrate acoustic, optical and electrical multimodal signals and realize high-precision, multi-task and transferable welding material performance testing.

[0005] In a first aspect, the present invention provides a method for testing fatigue properties of welding materials using acoustic, optical and electrical synergy. The method comprises: Synchronously collecting a first multimodal signal data set of the welding material during fatigue testing, wherein the first multimodal signal data set includes sound feedback signal data, optical feedback signal data, and electrical feedback signal data; fusing the sound feedback signal data, the light feedback signal data and the electrical feedback signal data to obtain a fused feature vector; Inputting the fused feature vector into a multivariate time series representation learning framework with a hybrid contextual attention mechanism to extract time series features, thereby obtaining a comprehensive representation of the time series; constructing and training an initial fatigue performance multi-task learning model comprising a shared feature extractor and task-specific sub-networks based on the comprehensive representation of the time series; Implementing domain adaptation and transfer learning on the fatigue performance multi-task learning model to obtain a target fatigue performance multi-task learning model; The second multimodal signal data set is collected in real time and input into the target fatigue performance multi-task learning model to perform fatigue life prediction, crack growth rate estimation, residual strength evaluation, material damage degree classification and anomaly detection to obtain the fatigue performance test results of the welding material.

[0006] In a second aspect, the present invention provides an acoustic, optical and electrical coordinated welding material fatigue performance testing system, the acoustic, optical and electrical coordinated welding material fatigue performance testing system comprising: An acquisition module, used for synchronously acquiring a first multimodal signal data set of the welding material during the fatigue test, wherein the first multimodal signal data set includes sound feedback signal data, optical feedback signal data and electrical feedback signal data; A fusion module, used for fusing the sound feedback signal data, the light feedback signal data and the electrical feedback signal data to obtain a fusion feature vector; An extraction module, used for inputting the fused feature vector into a multivariate time series representation learning framework with a hybrid context attention mechanism to extract time series features and obtain a comprehensive representation of the time series; A building module for building and training an initial fatigue performance multi-task learning model comprising a shared feature extractor and task-specific sub-networks based on the time series comprehensive representation; A migration module, used to implement domain adaptation and migration learning on the fatigue performance multi-task learning model to obtain a target fatigue performance multi-task learning model; The test module is used to collect the second multimodal signal data set in real time and input it into the target fatigue performance multi-task learning model to perform fatigue life prediction, crack growth rate estimation, residual strength evaluation, material damage degree classification and anomaly detection to obtain the fatigue performance test results of the welding material.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned acoustic, optical and electrical coordinated welding material fatigue performance testing method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned acoustic, optical and electrical coordinated welding material fatigue performance testing method.

[0009] In the technical solution provided by the present invention, efficient fusion of three modal signals of sound, light and electricity is achieved, and a more comprehensive and robust feature representation is extracted through the physical information scattering transformation method, thereby improving the accuracy and reliability of fatigue performance testing. The multivariate time series representation learning framework using a hybrid contextual attention mechanism effectively captures the dependencies of multimodal signals in time and space dimensions, and enhances the model's learning ability for long-term and short-term features. A multi-task learning model including a shared feature extractor and a task-specific subnetwork is constructed, which realizes the simultaneous optimization of multiple tasks such as fatigue life prediction, crack growth rate estimation, residual strength assessment, material damage classification and anomaly detection, thereby improving the overall performance of the model. Through the domain adversarial neural network structure and the progressive domain adaptation strategy, the generalization ability and transfer learning ability of the model are enhanced, so that it can adapt to different welding materials and test conditions. The present invention realizes real-time acquisition and processing of multimodal signal data, and quickly obtains comprehensive fatigue performance evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 A schematic diagram of the steps of a method for testing fatigue properties of welding materials using acoustic, optical and electrical synergy in an embodiment of the present invention; Figure 2 It is a structural schematic diagram of the welding material fatigue performance testing system with acoustic, optical and electrical coordination in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The embodiment of the present invention provides a method and system for testing fatigue properties of welding materials in an acoustic, optical and electrical coordinated manner. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the method for testing fatigue properties of welding materials by acoustic-optical-electrical coordination in an embodiment of the present invention includes: Step S1, synchronously collecting a first multimodal signal data set of the welding material during the fatigue test, wherein the first multimodal signal data set includes sound feedback signal data, optical feedback signal data, and electrical feedback signal data; It is understandable that the execution subject of the present invention may be an acoustic, optical and electrical coordinated welding material fatigue performance testing system, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as an execution subject as an example.

[0014] Specifically, the welding material to be tested is fixed on the fatigue testing machine to ensure that the material can withstand the applied cyclic load during the test and truly simulate the stress state of the material under actual working conditions. An array of sound sensors is arranged around the welding material under stress to capture the sound signal generated by the material during the fatigue test. In order to improve the quality of the signal, the sampling frequency of the sound sensor array is reasonably set according to the material characteristics and test requirements to ensure that the original sound signal with high resolution can be obtained. In order to reduce the impact of environmental noise on the data, the original sound signal is subjected to wavelet transform denoising. Wavelet transform is a signal processing technology that analyzes signals at different frequency scales. The sound signal obtained after denoising can better reflect the real acoustic characteristics of the welding material. Fast Fourier transform is performed on the denoised sound signal to convert the time domain signal into the frequency domain signal, extract the main frequency, frequency distribution and energy distribution characteristics of the sound signal, and obtain the sound feedback signal data. Install a spectrum analyzer, aim at the surface of the welding material under stress, collect spectral data covering the visible light to near infrared spectrum range, and obtain the original spectral data. Perform spectral correction on the original spectral data to eliminate the influence of instrument errors and environmental factors, ensure the accuracy of the optical feedback signal data, and obtain high-quality optical feedback signal data. At the same time, install resistance strain gauges and potential sensors on the welding materials under stress to measure the resistance change and potential distribution of the materials during fatigue and obtain electrical feedback signal data. By monitoring the resistance change of the material under different loads, the resistance change rate is calculated, and the potential distribution gradient can reflect the stress distribution inside the material. The sound feedback signal data, optical feedback signal data, and electrical feedback signal data are used as the first multimodal signal data set to reflect the various performances of the welding material during the fatigue test.

[0015] Step S2, fusing the sound feedback signal data, the light feedback signal data and the electrical feedback signal data to obtain a fused feature vector; Specifically, the sound feedback signal data is subjected to wavelet convolution transformation, and the sound signal is converted into a time-frequency representation matrix to capture the changing characteristics of the signal in the time and frequency domains. Wavelet convolution transformation can provide information with high time resolution and high frequency resolution, so that the system can deeply analyze the instantaneous characteristics of the sound signal. The optical feedback signal data is also subjected to wavelet convolution transformation to obtain the time-frequency representation matrix of the spectral signal. The time-frequency representation of the spectral signal can reflect the response characteristics of the material under different spectral conditions. The electrical feedback signal data is subjected to wavelet convolution transformation to obtain the time-frequency representation matrix of the electrical signal. Nonlinear modulus operations are performed on the time-frequency representation matrix of the sound signal, the time-frequency representation matrix of the spectral signal, and the time-frequency representation matrix of the electrical signal, respectively, to extract the nonlinear feature matrix of the sound signal, the nonlinear feature matrix of the spectral signal, and the nonlinear feature matrix of the electrical signal, which contain important nonlinear relationships in the signal and reveal the complex behavior patterns of the welding material during fatigue testing. The nonlinear feature matrix is ​​input into a two-dimensional Gaussian low-pass filter for processing to remove unnecessary high-frequency information, and the filtered sound signal feature matrix, the filtered spectral signal feature matrix, and the filtered electrical signal feature matrix are obtained. A three-layer fully connected neural network is constructed as a feature fusion network. The number of nodes in the input layer of the network is equal to the total dimension of the filtered sound signal feature matrix, the filtered spectral signal feature matrix, and the filtered electrical signal feature matrix, so as to ensure that the network receives all important information. The number of nodes in the hidden layer is set to half of the input layer, which effectively reduces the computational complexity while maintaining the network's expressive power. The number of nodes in the output layer is set to 100, aiming to compress the features and extract the key features. In the entire neural network, the Sigmoid activation function is used to effectively introduce nonlinear characteristics, so that the network can learn complex feature relationships. Through forward propagation, the data of the input layer is passed through the hidden layer to the output layer. The calculated fusion feature vector is a high-dimensional feature representation that integrates the feedback information of the three signals, reflecting the comprehensive behavioral characteristics of the welding material during fatigue testing.

[0016] Perform a complex transformation on the time-frequency representation matrix of the sound signal to obtain the complex matrix of the sound signal. The phase and amplitude information of the signal are completely retained through the complex transformation. Perform the same complex transformation on the time-frequency representation matrix of the spectral signal and the time-frequency representation matrix of the electrical signal to obtain the complex matrix of the spectral signal and the complex matrix of the electrical signal. Perform square operations on the real and imaginary parts of the complex matrix of the sound signal respectively to obtain the real square matrix and the imaginary square matrix of the sound signal. Perform the same steps on the spectral signal and the electrical signal to calculate their real square matrix and imaginary square matrix. The square matrix can reflect the energy characteristics of the signal. Perform element-by-element addition operations to add the real square matrix and the imaginary square matrix of the sound signal to obtain the modulus square matrix of the sound signal. Perform the same operations on the spectral signal and the electrical signal to obtain the modulus square matrix of the spectral signal and the modulus square matrix of the electrical signal. The modulus square matrix reflects the overall energy distribution of the signal. Perform a square root operation on the modulus square matrix of the sound signal to obtain the modulus value matrix of the sound signal. Similarly, square root operations are performed on the modulus square matrix of the spectral signal and the modulus square matrix of the electrical signal to obtain the corresponding modulus value matrix. The modulus value matrix reflects the amplitude characteristics of the signal. The modulus value matrix of the sound signal is nonlinearly transformed to obtain the preliminary nonlinear feature matrix of the sound signal. Nonlinear functions, such as exponential functions or power functions, are applied to emphasize important features and trends in the signal. The modulus value matrix of the spectral signal and the modulus value matrix of the electrical signal are nonlinearly transformed to obtain the corresponding preliminary nonlinear feature matrix of the spectral signal and the preliminary nonlinear feature matrix of the electrical signal. The L2 norm of the preliminary nonlinear feature matrix of the sound signal, that is, the square root of the sum of squares of all elements, is calculated to effectively compress the numerical range of the feature matrix to a uniform scale, avoid the unbalanced influence caused by the difference in eigenvalues ​​in subsequent analysis, and obtain the normalized nonlinear feature matrix of the sound signal. The preliminary nonlinear feature matrix of the spectral signal and the preliminary nonlinear feature matrix of the electrical signal are respectively normalized by the L2 norm to obtain the normalized nonlinear feature matrix of the spectral signal and the normalized nonlinear feature matrix of the electrical signal. Multi-scale and multi-directional feature extraction is performed on the normalized nonlinear feature matrix of the sound signal. Gabor transform is applied to capture the frequency and time information of the signal at the same time. Through Gabor filters of different scales and directions, the detailed features in the sound signal are extracted to obtain the Gabor feature matrix of the sound signal. Gabor feature extraction is performed on the normalized nonlinear feature matrix of the spectral signal and the electrical signal to obtain the Gabor feature matrix of the spectral signal and the Gabor feature matrix of the electrical signal. Local response normalization is performed on the Gabor feature matrix of the sound signal to effectively eliminate the response differences between different regions in the feature matrix and ensure that the features of the signal in different parts are consistent. Local response normalization compares each eigenvalue with the eigenvalue in its neighborhood to enhance the expression of significant features and suppress the influence of noise and unimportant features to obtain the nonlinear feature matrix of the sound signal.Similarly, the Gabor characteristic matrix of the spectral signal and the Gabor characteristic matrix of the electrical signal are locally normalized to obtain the nonlinear characteristic matrices of the spectral signal and the electrical signal.

[0017] Step S3, inputting the fused feature vector into a multivariate time series representation learning framework with a hybrid contextual attention mechanism to extract time series features and obtain a comprehensive representation of the time series; Specifically, the fused feature vector is input into the local attention module, and the short-term dependencies in the time series are captured by the sliding window method. The local attention module can focus on the features of adjacent moments in the time series and extract local attention features. At the same time, the fused feature vector is input into the global attention module, and the long-term dependencies are learned through the self-attention mechanism, which effectively identifies the patterns and connections of longer time spans in the time series and obtains the global attention features. The local attention features and the global attention features are weighted fused to obtain the mixed context features. By setting the weights, the advantages of local and global features are combined to enhance the ability of feature representation. The mixed context features are input into the autoencoder structure, and the compressed representation of the time series is learned by minimizing the reconstruction error. In this process, the autoencoder can effectively compress the data and extract the most representative features in the signal to obtain the reconstructed features. Based on the mixed context features, positive sample pairs and negative sample pairs are constructed. Through the contrastive learning method, the representations of similar contexts are made closer in the feature space to obtain the context-consistent features. In order to ensure the continuity of the representations adjacent in time, the mixed context features are temporally smoothed to eliminate the noise caused by signal fluctuations, ensure the smooth transition between adjacent features, and obtain the temporally consistent features. The reconstruction features, context-consistent features, and time-consistent features are input into the multivariate time series representation learning framework. Through the joint optimization method, the three objective functions of the reconstruction regression objective, the context-consistent comparative learning objective, and the time-consistent comparative learning objective are integrated to obtain the multi-objective optimization features. The multi-objective optimization features are subjected to attention pooling, and the learned time series representation is compressed into a vector of fixed dimension to obtain the comprehensive representation of the time series.

[0018] Step S4, constructing and training an initial fatigue performance multi-task learning model including a shared feature extractor and a task-specific sub-network based on the comprehensive representation of the time series; Specifically, the time series comprehensive representation is input into the shared feature extractor, which adopts a multi-layer perceptron structure and contains three fully connected layers. In each fully connected layer, the ReLU activation function is used to introduce nonlinearity, and Dropout regularization is applied to prevent overfitting. The shared feature extractor can extract key features from the original time series data to form a robust training shared feature representation. Construct subnetworks for different tasks. Construct a fatigue life prediction subnetwork, which consists of two fully connected layers. Each layer uses the Tanh activation function and outputs a single value representing the predicted fatigue life cycle. The model can directly learn key information related to fatigue life from shared features, thereby improving the accuracy of prediction. Construct a crack growth rate estimation subnetwork, which includes two fully connected layers and one LSTM layer. The LSTM layer can effectively capture the timing characteristics of the crack growth process and help evaluate the crack growth rate. Through the Sigmoid activation function, the final output result is the normalized crack growth rate, ensuring that the output value is in the range of 0 to 1. Construct a residual strength evaluation subnetwork. The subnetwork design contains three fully connected layers. Each layer uses the LeakyReLU activation function, which can retain a certain gradient when the input is negative, making the model more stable during training. The last layer uses a linear activation function to output the predicted residual strength value. The material damage classification subnetwork is constructed. Combining the characteristics of the convolutional neural network, it contains two fully connected layers and a 1D convolution layer. The introduction of the 1D convolution layer is intended to extract local features in the time series data, so that the model can better capture the subtle information of the damage change. The probability distribution of multiple damage categories is output through the Softmax activation function to provide a multi-dimensional evaluation of the material damage. In order to achieve anomaly detection, a special anomaly detection subnetwork is constructed, which adopts the structure of a variational autoencoder. The variational autoencoder consists of an encoder and a decoder. The encoder part contains three fully connected layers, which is intended to compress the input data into a latent variable representation, while the decoder also contains three fully connected layers to reconstruct the input data from the latent variables. In this process, the reparameterization technique is used to generate latent variables to improve the effect of anomaly detection. When each sub-network is constructed, the shared feature representation obtained through training will be input into the fatigue life prediction sub-network, crack growth rate estimation sub-network, residual strength assessment sub-network, material damage classification sub-network and anomaly detection sub-network respectively to obtain the initial output results of each task. In order to improve the performance of the model, a multi-task loss function is constructed. The fatigue life prediction and residual strength assessment tasks use mean square error loss, the crack growth rate estimation task uses mean absolute error loss, the material damage classification task uses cross entropy loss, and the anomaly detection task uses a combination loss of reconstruction error and KL divergence. Through the multi-task loss function, the performance of each task is optimized at the same time, so that the model has stronger generalization ability.The Adam optimizer is used to jointly optimize the shared feature extractor and each task-specific subnetwork. In order to achieve more efficient training, a dynamic weight balancing strategy is used to adaptively adjust the weight of each task loss. This strategy can adjust the weight in real time according to the training status of each task, so that the model can find the best balance between different tasks, and finally obtain an efficient initial fatigue performance multi-task learning model.

[0019] Step S5, performing domain adaptation and transfer learning on the fatigue performance multi-task learning model to obtain a target fatigue performance multi-task learning model; Specifically, the source domain data and the target domain data are input into the fatigue performance multi-task learning model to extract the source domain features and the target domain features. A domain discriminator network is constructed, which contains three fully connected layers. Each layer uses the LeakyReLU activation function to improve the expressiveness of the model. The last layer uses the Sigmoid activation function to output the predicted probability of the domain label and clearly distinguish the features of the source domain and target domain data. The source domain features and the target domain features are input into the domain discriminator network, and the parameters of the domain discriminator are optimized by back propagation to calculate the domain discrimination loss. This loss is used to guide the domain discriminator to effectively distinguish the features of the source domain and the target domain in the feature space, and also provides important feedback for subsequent training. In order to improve the effect of feature extraction, a gradient reversal layer is introduced into the shared feature extractor of the initial fatigue performance multi-task learning model. This layer is used to train the feature extractor by maximizing the domain discrimination loss to generate domain invariant features, which helps to improve the adaptability of the model in the target domain and finally obtain the domain adversarial training loss. A progressive domain adaptation strategy is constructed to divide the target domain data into multiple subsets, which are arranged in descending order of similarity to the source domain to form a subset sequence of the target domain data. In this sequence, domain adversarial training is performed on each subset in turn, and the proportion of target domain data in the training process is gradually increased. This enables the model to maintain a good match with the source domain features while gradually adapting to the target domain features, forming a staged domain adaptation model. Based on the staged domain adaptation model, a consistency regularization term is introduced to process the task-specific subnetwork to minimize the difference between the prediction results of the source domain and the target domain. This enables the model to establish consistency between the source domain and the target domain, improving the stability and reliability of the overall prediction. At the same time, the domain adversarial training loss, domain consistency loss and multi-task learning loss are weightedly combined for end-to-end training to improve the performance of the model in the target domain and maintain the performance in the source domain, thus obtaining a target fatigue performance multi-task learning model.

[0020] Step S6, real-time acquisition of the second multimodal signal data set and input into the target fatigue performance multi-task learning model to perform fatigue life prediction, crack growth rate estimation, residual strength evaluation, material damage classification and anomaly detection to obtain the fatigue performance test results of the welding material.

[0021] Specifically, fatigue tests are performed on the welding materials to be tested, and feedback signals from different modes, including sound feedback signals, optical feedback signals, and electrical feedback signals, are collected synchronously to form a second multimodal signal data set. The second multimodal signal data set is preprocessed, including denoising, standardization, and normalization, to ensure the quality and consistency of the signal. Feature fusion is performed, and the data features of different modes are integrated together through an effective algorithm to obtain a comprehensive representation of the target sequence. The comprehensive representation of the target sequence is input into the shared feature extractor of the target fatigue performance multi-task learning model for feature extraction. The shared feature representation extracted by the feature extractor through the multi-layer structure is used as the input of subsequent sub-networks. The real-time shared feature representation is input into the fatigue life prediction sub-network, in which the predicted number of fatigue life cycles is calculated through two fully connected layers combined with the Tanh activation function, which directly reflects the durability of the welding material in actual use. At the same time, the same real-time shared feature representation is input into the crack growth rate estimation subnetwork, which is designed to contain two fully connected layers and one LSTM layer. The introduction of the LSTM layer enables the model to effectively capture the temporal characteristics of the signal. Through the processing of the Sigmoid activation function, the normalized crack growth rate is output, which helps to evaluate the crack growth rate of the welding material under fatigue state. The real-time shared feature representation is input into the residual strength evaluation subnetwork. In this network, three fully connected layers are used and the LeakyReLU activation function is adopted to enhance the nonlinear fitting ability of the model. The last layer uses a linear activation function to output the predicted residual strength value, which reflects the remaining bearing capacity of the material after fatigue. The real-time shared feature representation is input into the material damage classification subnetwork. In this network, local features are extracted through two fully connected layers and a 1D convolutional layer, and the probability distribution of multiple damage categories is output using the Softmax activation function. The real-time shared feature representation is input into the anomaly detection subnetwork, which is encoded and decoded through the variational autoencoder structure, and the reconstruction error and KL divergence are calculated to determine whether there is anomaly. This process can not only identify potential anomalies, but also provide important information about data quality. The predicted fatigue life cycles, normalized crack growth rate, predicted residual strength value, probability distribution of multiple damage degree categories and abnormal detection results are comprehensively analyzed to form the evaluation results of the fatigue performance of the welding material.

[0022] The real-time shared feature representation is input into the encoder part of the anomaly detection subnetwork. The encoder contains three fully connected layers. These layers convert the input data into a more compact representation through feature compression. The compressed features retain the main information of the input and reduce redundancy. The mean and variance of the compressed features are calculated to obtain the distribution parameters of the latent space. The variational autoencoder defines the probability distribution of the latent variables. Reparameterized sampling is performed based on the distribution parameters to obtain the latent variables, so that the model can perform effective back propagation during the training process and maintain the stability and feasibility of the training. The latent variables are input into the decoder. The decoder consists of three fully connected layers to reconstruct the latent variables back to the feature space. Through the reconstruction process, a set of reconstructed features are obtained, which should be as close as possible to the original real-time shared feature representation. In order to evaluate the quality of the reconstruction, the mean square error between the reconstructed features and the original shared feature representation is calculated. The error value is an important indicator for measuring the performance of the model and reflects the effectiveness of the model in the reconstruction process. At the same time, based on the distribution parameters, the KL divergence between the distribution of the latent variables and the standard normal distribution is calculated. It reflects the distribution of latent variables and reveals the learning effect of the model. The smaller the value of KL divergence, the closer the distribution of latent variables is to the standard normal distribution, thereby improving the model's ability to generate new samples. After weighted summation of the reconstruction error and the KL divergence value, a comprehensive anomaly score is obtained. This score comprehensively considers the accuracy of reconstruction and the rationality of latent variables, and can effectively judge the degree of abnormality of the input sample. For the calculated anomaly score, a suitable threshold is set for judgment. If the anomaly score is higher than this threshold, the sample is judged to be abnormal, otherwise it is considered normal. In this way, the data quality in the fatigue performance test of welding materials is monitored in real time and potential abnormalities are identified.

[0023] In the embodiment of the present invention, the efficient fusion of three modal signals of sound, light and electricity is achieved, and a more comprehensive and robust feature representation is extracted through the physical information scattering transformation method, thereby improving the accuracy and reliability of fatigue performance testing. The multivariate time series representation learning framework using a hybrid contextual attention mechanism effectively captures the dependencies of multimodal signals in time and space dimensions, and enhances the model's learning ability for long-term and short-term features. A multi-task learning model including a shared feature extractor and a task-specific subnetwork is constructed, which realizes the simultaneous optimization of multiple tasks such as fatigue life prediction, crack growth rate estimation, residual strength assessment, material damage classification and anomaly detection, thereby improving the overall performance of the model. Through the domain adversarial neural network structure and the progressive domain adaptation strategy, the generalization ability and transfer learning ability of the model are enhanced, so that it can adapt to different welding materials and test conditions. The present invention realizes real-time acquisition and processing of multimodal signal data, and quickly obtains comprehensive fatigue performance evaluation results.

[0024] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Fixing the welding material on a fatigue testing machine, applying a cyclic load to the welding material, and obtaining the welding material in a stress state; Arrange a sound sensor array around the welding material in a stressed state, set a sampling frequency for the sound sensor array to obtain an original sound signal, and perform wavelet transform on the original sound signal to reduce noise to obtain a noise-reduced sound signal; Perform fast Fourier transform on the denoised sound signal, convert the time domain signal into a frequency domain signal, extract the main frequency, frequency distribution and energy distribution characteristics, and obtain the sound feedback signal data; Install a spectrum analyzer, aim it at the surface of the welding material under stress, collect spectrum data covering the visible light to near infrared spectrum range, obtain original spectrum data, and perform spectrum correction on the original spectrum data to obtain optical feedback signal data; Install resistance strain gauges and potential sensors on the welding material under stress, measure the resistance change and potential distribution of the material during fatigue, calculate the resistance change rate and potential distribution gradient, and obtain electrical feedback signal data; The sound feedback signal data, the light feedback signal data and the electrical feedback signal data are used as the first multimodal signal data set.

[0025] Specifically, the welding material to be tested is fixed on a fatigue testing machine, and a cyclic load is applied to it to simulate the stress state of the material in actual application. After the cyclic load is applied, the welding material enters a stress state. An array of sound sensors is arranged around the welding material in the stress state to monitor the acoustic signals generated during the fatigue process of the material. When arranging the sound sensor array, a suitable sampling frequency is set to ensure that the details of the sound signal are captured to obtain the original sound signal. The original sound signal is subjected to wavelet transform noise reduction to effectively remove unnecessary background noise, improve the quality of the signal, and obtain the noise-reduced sound signal. The noise-reduced sound signal is subjected to fast Fourier transform to convert the time domain signal into a frequency domain signal, and extract the main frequency components, frequency distribution, and energy distribution characteristics of the sound signal. For example, the frequency characteristics of the sound signal are calculated by fast Fourier transform, and the specific formula is: ; in, represents the frequency domain signal, is the time domain signal, is the frequency, is an imaginary unit. The sound feedback signal data is obtained by fast Fourier transform processing of the sound signal after noise reduction. At the same time, in order to obtain the optical feedback signal of the welding material, a spectrometer is installed and data is collected on the surface of the welding material under stress. The spectrometer can collect spectral data covering the spectrum range from visible light to near-infrared. The original spectral data is spectrally corrected to eliminate the influence caused by equipment errors or environmental factors, so as to obtain more accurate optical feedback signal data. Resistance strain gauges and potential sensors are installed on the welding material under stress to measure the change in resistance and the distribution of potential of the material during fatigue. The resistance change rate of the material is obtained by measuring the resistance strain gauge, and the calculation formula is expressed as: ; in, is the resistance of the material after being subjected to force. is the resistance of the material when no force is applied, is strain, which indicates the ratio of the length change of the material after being stressed to its original length. The potential sensor is used to measure the potential distribution and calculate the potential distribution gradient, and these data are finally synthesized into electrical feedback signal data. By comprehensively processing the sound feedback signal data, the optical feedback signal data and the electrical feedback signal data, the first multimodal signal data set is formed, including different types of signals and different response characteristics of the material under stress.

[0026] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Performing wavelet convolution transformation on the sound feedback signal data to obtain a time-frequency representation matrix of the sound signal, performing wavelet convolution transformation on the light feedback signal data to obtain a time-frequency representation matrix of the spectrum signal, and performing wavelet convolution transformation on the electrical feedback signal data to obtain a time-frequency representation matrix of the electrical signal; Performing nonlinear modulus operations on the time-frequency representation matrix of the sound signal, the time-frequency representation matrix of the spectrum signal, and the time-frequency representation matrix of the electrical signal, respectively, to obtain a nonlinear feature matrix of the sound signal, a nonlinear feature matrix of the spectrum signal, and a nonlinear feature matrix of the electrical signal; The nonlinear characteristic matrix of the sound signal, the nonlinear characteristic matrix of the spectral signal and the nonlinear characteristic matrix of the electric signal are respectively input into a two-dimensional Gaussian low-pass filter to filter high-frequency information, so as to obtain a filtered sound signal characteristic matrix, a filtered spectral signal characteristic matrix and a filtered electric signal characteristic matrix; A three-layer fully connected neural network is constructed as the feature fusion network. The number of nodes in the input layer is the total dimension of the filtered sound signal feature matrix, the filtered spectral signal feature matrix, and the filtered electrical signal feature matrix. The number of nodes in the hidden layer is 1 / 2 of the input layer, and the number of nodes in the output layer is 100. The Sigmoid activation function is used to obtain the fused feature vector through forward propagation calculation.

[0027] Specifically, wavelet convolution transform is applied to the sound feedback signal data to obtain the time-frequency representation matrix of the sound signal. This matrix contains detailed information about time and frequency, and reveals the spectral characteristics of the signal in different time periods. At the same time, wavelet convolution transform is performed on the optical feedback signal data and the electrical feedback signal data to obtain the time-frequency representation matrices of the spectral signal and the electrical signal, respectively. These time-frequency representation matrices can reflect the frequency components of each signal and their changes over time. Nonlinear modulus operations are performed on the time-frequency representation matrices to extract the nonlinear characteristics of the signal. Nonlinear modulus operations are performed through a series of mathematical operations, such as performing nonlinear function transformations on each element. Assume that the time-frequency representation matrix of the signal is expressed by Represented as follows, the nonlinear characteristic matrix It is expressed as: ; in, is the nonlinear characteristic matrix, is the time-frequency representation matrix, It is a nonlinear function applied to each element of the matrix, such as using the ReLU function, Sigmoid function, etc., to enhance the important features in the signal, making the subsequent analysis more effective. The nonlinear feature matrix of the sound signal, the nonlinear feature matrix of the spectral signal, and the nonlinear feature matrix of the electrical signal are respectively input into the two-dimensional Gaussian low-pass filter for high-frequency information filtering. The Gaussian low-pass filter effectively removes high-frequency noise, retains the low-frequency signal components, and obtains a cleaner feature matrix. The two-dimensional Gaussian low-pass filter is expressed as: ; in, is the filter response function, and are the coordinates in the frequency domain, is the standard deviation, which controls the bandwidth of the filter. After filtering, the filtered sound signal feature matrix, the filtered spectral signal feature matrix, and the filtered electrical signal feature matrix are obtained. A three-layer fully connected neural network is constructed as a feature fusion network. In this network, the number of nodes in the input layer is equal to the total dimension of the filtered feature matrix, while the number of nodes in the hidden layer is set to 1 / 2 of the number of nodes in the input layer, and the number of nodes in the output layer is fixed to 100. This design enables the network to effectively extract and fuse features from different modalities and improve the performance of the model. The activation function of the neural network uses the Sigmoid function, which is mathematically expressed as: ; in, is the output of the Sigmoid activation function, is the input signal. Through the process of forward propagation, the network calculates the output of each layer and finally obtains a fused feature vector that integrates the characteristic information of sound, spectrum and electrical signals.

[0028] In a specific embodiment, the execution step performs nonlinear modulus operations on the time-frequency representation matrix of the sound signal, the time-frequency representation matrix of the spectral signal, and the time-frequency representation matrix of the electrical signal, respectively, to obtain the nonlinear feature matrix of the sound signal, the nonlinear feature matrix of the spectral signal, and the nonlinear feature matrix of the electrical signal, which may specifically include the following steps: Performing a complex transformation on the time-frequency representation matrix of the sound signal to obtain a complex matrix of the sound signal, performing a complex transformation on the time-frequency representation matrix of the spectral signal to obtain a complex matrix of the spectral signal, and performing a complex transformation on the time-frequency representation matrix of the electrical signal to obtain a complex matrix of the electrical signal; Performing square operations on the real and imaginary parts of the complex matrix of the sound signal, respectively, to obtain a square matrix of the real part of the sound signal and a square matrix of the imaginary part of the sound signal; performing square operations on the real and imaginary parts of the complex matrix of the spectral signal, respectively, to obtain a square matrix of the real part of the spectral signal and a square matrix of the imaginary part of the spectral signal; performing square operations on the real and imaginary parts of the complex matrix of the electric signal, respectively, to obtain a square matrix of the real part of the electric signal and a square matrix of the imaginary part of the electric signal; Performing element-by-element addition operation on the real part square matrix of the sound signal and the imaginary part square matrix of the sound signal to obtain the sound signal modulus square matrix, performing element-by-element addition operation on the real part square matrix of the spectral signal and the imaginary part square matrix of the spectral signal to obtain the spectral signal modulus square matrix, performing element-by-element addition operation on the real part square matrix of the electric signal and the imaginary part square matrix of the electric signal to obtain the electric signal modulus square matrix; Performing a square root operation on the sound signal modulus square matrix to obtain a sound signal modulus value matrix, performing a square root operation on the spectral signal modulus square matrix to obtain a spectral signal modulus value matrix, and performing a square root operation on the electric signal modulus square matrix to obtain an electric signal modulus value matrix; Performing a nonlinear transformation on the sound signal module value matrix to obtain a preliminary nonlinear characteristic matrix of the sound signal, performing a nonlinear transformation on the spectrum signal module value matrix to obtain a preliminary nonlinear characteristic matrix of the spectrum signal, and performing a nonlinear transformation on the electrical signal module value matrix to obtain a preliminary nonlinear characteristic matrix of the electrical signal; Performing L2 norm normalization on the preliminary nonlinear feature matrix of the sound signal to obtain a normalized nonlinear feature matrix of the sound signal, performing L2 norm normalization on the preliminary nonlinear feature matrix of the spectral signal to obtain a normalized nonlinear feature matrix of the spectral signal, and performing L2 norm normalization on the preliminary nonlinear feature matrix of the electric signal to obtain a normalized nonlinear feature matrix of the electric signal; Perform multi-scale and multi-directional feature extraction on the normalized nonlinear feature matrix of the sound signal to obtain the Gabor feature matrix of the sound signal, perform multi-scale and multi-directional feature extraction on the normalized nonlinear feature matrix of the spectral signal to obtain the Gabor feature matrix of the spectral signal, and perform multi-scale and multi-directional feature extraction on the normalized nonlinear feature matrix of the electrical signal to obtain the Gabor feature matrix of the electrical signal; The Gabor feature matrix of the sound signal is locally normalized to obtain the nonlinear feature matrix of the sound signal. The Gabor feature matrix of the spectral signal is locally normalized to obtain the nonlinear feature matrix of the spectral signal. The Gabor feature matrix of the electrical signal is locally normalized to obtain the nonlinear feature matrix of the electrical signal.

[0029] Specifically, the time-frequency representation matrix of the sound signal is transformed into a complex number to obtain the complex number matrix of the sound signal. The complex number transformation captures the phase and amplitude information of the signal by combining the real part and the imaginary part. The time-frequency representation matrix of the spectral signal and the electrical signal is also transformed into a complex number to obtain the complex number matrix of the spectral signal and the complex number matrix of the electrical signal, respectively. The real part and the imaginary part of the complex number matrix are squared. Taking the complex number matrix of the sound signal as an example, if its complex number matrix is ,in is the real part, is the imaginary part, then the real square matrix of the sound signal is and the imaginary square matrix It is expressed as: ; The real and imaginary parts of the complex matrices of the spectral signal and the electrical signal are squared to obtain the real square matrices of the spectral signal and the imaginary square matrix , and the real part square matrix of the electrical signal and the imaginary square matrix After the square operation is completed, the real part square matrix and the imaginary part square matrix of the sound signal are added element by element to obtain the modulus square matrix of the sound signal. The same operation is performed on the spectral signal and the electrical signal to obtain the spectral signal modulus square matrix Sum signal modulus square matrix Perform square root operation on the modulus square matrix of each signal to obtain the sound signal modulus value matrix , spectral signal modulus matrix And the electrical signal modulus matrix . Perform nonlinear transformation on the modulus matrix. Nonlinear transformation takes many forms, for example, by performing logarithmic transformation or power transformation on the modulus to enhance the expressiveness of the feature. Assume that the modulus matrix Perform nonlinear transformation to obtain the preliminary nonlinear feature matrix of the sound signal , and the preliminary nonlinear characteristic matrices of the spectral signal and the electrical signal are recorded as and In order to unify the feature scale, the L2 norm normalization method is used to process each feature matrix. The formula for L2 norm normalization is:

[0030] in, is the original feature matrix, is a matrix The L2 norm of is calculated as . Through this formula, the values ​​of each feature matrix are normalized to the same range to avoid adverse effects caused by inconsistent feature scales in subsequent analysis. Multi-scale and multi-directional feature extraction is performed on the normalized nonlinear feature matrices of sound signals, spectral signals, and electrical signals to obtain the Gabor feature matrix. Gabor feature extraction can capture the local frequency characteristics of the signal and is suitable for processing the time-frequency characteristics of the signal. By applying a series of Gabor filters of different scales and directions, the Gabor feature matrix of the sound signal is obtained. , Gabor characteristic matrix of spectral signal Gabor characteristic matrix of the electrical signal . The Gabor feature matrix is ​​subjected to local response normalization to highlight the significance of the feature and reduce the influence of unnecessary noise. The purpose of local response normalization is to make the response of each feature in its local neighborhood relatively balanced, so as to highlight important features. After completing local response normalization, the final nonlinear feature matrix is ​​obtained, which is the nonlinear feature matrix of the sound signal, the nonlinear feature matrix of the spectral signal, and the nonlinear feature matrix of the electrical signal.

[0031] In a specific embodiment, the process of executing step S3 may specifically include the following steps: The fused feature vector is input into the local attention module, and the short-term dependency in the time series is captured by the sliding window method to obtain the local attention feature. The fused feature vector is input into the global attention module, and the long-term dependency is learned by the self-attention mechanism to obtain the global attention feature. The local attention features and the global attention features are weightedly fused to obtain the mixed context features, which are then input into the autoencoder structure. The compressed representation of the time series is learned by minimizing the reconstruction error to obtain the reconstructed features. Positive and negative sample pairs are constructed based on the mixed context features. The representations of similar contexts are made closer in the feature space through contrastive learning methods to obtain the context-consistent features. Temporal smoothing is performed on the mixed context features to ensure the continuity of temporally adjacent representations and obtain temporally consistent features. The reconstruction features, context-consistent features and time-consistent features are input into the multivariate time series representation learning framework, and the multi-objective optimization features are obtained by jointly optimizing the three objective functions of the reconstruction-based regression objective, the context-consistent comparative learning objective and the time-consistent comparative learning objective. Attention pooling is performed on the multi-objective optimization features to compress the learned time series representation into a vector of fixed dimension to obtain a comprehensive representation of the time series.

[0032] Specifically, the fused feature vector is input into the local attention module, and the sliding window method is used to capture the short-term dependencies in the time series. The sliding window extracts local features by moving a fixed-size window in the time series. The data in the window is used to generate attention scores to highlight important features. By calculating the local attention features, valuable information is extracted from the short-term fluctuations of the time series. At the same time, the fused feature vector is input into the global attention module, and the self-attention mechanism is used to learn long-term dependencies. In the self-attention mechanism, the feature vector of each time step interacts with other time steps, and the attention weight is obtained by calculating the similarity. It can be expressed as: ; in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key vector. This mechanism can effectively capture global information, ensure that relevant features are learned throughout the time series, and enhance the performance of the model. The local attention features and the global attention features are weighted and fused to obtain the mixed context features. The mixed context features are input into the autoencoder structure to learn the compressed representation of the time series by minimizing the reconstruction error. The goal of the autoencoder is to compress the input data to a low-dimensional representation and then reconstruct it. The reconstruction error can be expressed by the mean square error, as shown below: ; in, is the original feature, is the reconstructed feature, is the total number of samples. By optimizing the reconstruction process, an effective time series representation is obtained. Based on the mixed context features, positive sample pairs and negative sample pairs are constructed, and the contrastive learning method is used to make the representations of similar contexts closer in the feature space. In this way, the sensitivity of the model to different contexts in the time series is improved, so that similar contexts can be more closely clustered in the feature space. For positive sample pairs, it is defined as a pair of similar contexts, while negative sample pairs are two dissimilar contexts. Through the construction of sample pairs, the model can learn more effective feature expressions during the training process. In order to ensure the continuity of temporally adjacent representations, the mixed context features are temporally smoothed to ensure that the representations in the time series are consistent in continuity. Temporally consistent features help provide smoother input in subsequent tasks and reduce the impact of data noise or mutations. After completing the temporal smoothing process, the reconstruction features, context-consistent features, and temporally consistent features are input into the multivariate time series representation learning framework. In this framework, multiple objectives are jointly optimized, including the reconstruction-based regression objective, the context-consistent comparative learning objective, and the temporally consistent comparative learning objective. The form of multi-objective optimization enables the model to be adjusted in many aspects to improve the quality of feature representation. By jointly optimizing the objective function, the model can simultaneously focus on the accuracy of reconstruction, the similarity between contexts, and the consistency in time, thus improving the overall performance. Attention pooling is performed on the multi-objective optimization features to compress the learned time series representation into a vector of fixed dimension to obtain a comprehensive representation of the time series. Attention pooling can assign different weights to different features according to the importance of each feature in a specific task, achieve efficient information compression and extraction, and obtain a high-dimensional vector containing rich contextual information and time series characteristics.

[0033] In a specific embodiment, the process of executing step S4 may specifically include the following steps: The time series comprehensive representation is input into the shared feature extractor. The shared feature extractor adopts a multi-layer perceptron structure, which contains 3 fully connected layers. Each layer uses the ReLU activation function and Dropout regularization to obtain the training shared feature representation. Construct a fatigue life prediction subnetwork, which includes two fully connected layers. Each layer uses the Tanh activation function, and the last layer outputs a single value, which represents the predicted number of fatigue life cycles. Construct a crack growth rate estimation subnetwork, which includes two fully connected layers and one LSTM layer. The LSTM layer is used to capture the timing characteristics of crack growth. The last layer uses the Sigmoid activation function to output the normalized crack growth rate. Construct a residual strength evaluation subnetwork, which includes 3 fully connected layers. Each layer uses the LeakyReLU activation function. The last layer uses a linear activation function to output the predicted residual strength value. Construct a material damage classification subnetwork, which includes two fully connected layers and one 1D convolutional layer. The 1D convolutional layer is used to extract local features. The last layer uses the Softmax activation function to output the probability distribution of multiple damage categories. Construct anomaly detection subnetwork, using variational autoencoder structure, including encoder and decoder. The encoder contains 3 fully connected layers, and the decoder contains 3 fully connected layers. The middle layer uses reparameterization technique to generate latent variables. The training shared feature representation is input into the fatigue life prediction subnetwork, crack growth rate estimation subnetwork, residual strength assessment subnetwork, material damage classification subnetwork and anomaly detection subnetwork respectively to obtain the initial output results of each task; Construct a multi-task loss function. The fatigue life prediction and residual strength assessment tasks use mean square error loss, the crack growth rate estimation task uses mean absolute error loss, the material damage classification task uses cross entropy loss, and the anomaly detection task uses a combination loss of reconstruction error and KL divergence. The Adam optimizer is used to jointly optimize the shared feature extractor and each task-specific sub-network, and a dynamic weight balancing strategy is used to adaptively adjust the weight of each task loss to obtain the initial fatigue performance multi-task learning model.

[0034] Specifically, the time series comprehensive representation is input into the shared feature extractor. The shared feature extractor adopts a multi-layer perceptron structure, which contains three fully connected layers. Each layer uses the ReLU activation function and Dropout regularization to prevent overfitting. Through this process, the trained shared feature representation is obtained, which effectively captures the key information in the input data. The output of the shared feature extractor is expressed as:

[0035] in, is a shared feature representation, is the comprehensive representation of the input time series, and is the weight matrix, and is the bias term. After this process, the feature extractor can effectively extract features that are helpful for subsequent tasks. A fatigue life prediction subnetwork is constructed. The network is designed to contain two fully connected layers, each of which uses the Tanh activation function and outputs a single value in the last layer, representing the predicted number of fatigue life cycles. Through this subnetwork, the extracted shared features are mapped to the actual fatigue life prediction, enhancing the practicality of the model. The crack growth rate estimation subnetwork contains two fully connected layers and one LSTM layer. The introduction of the LSTM layer is intended to capture the timing characteristics of crack growth, so that the model can make full use of the dynamic changes in time series data. In the last layer of this subnetwork, the Sigmoid activation function is used to output the normalized crack growth rate to ensure that the output value is between 0 and 1, which is helpful for subsequent damage assessment. At the same time, the residual strength assessment subnetwork is designed to contain three fully connected layers, each of which uses the LeakyReLU activation function, and the last layer uses a linear activation function to output the predicted residual strength value. This design makes the model more expressive when dealing with nonlinear problems, which helps to accurately evaluate the strength of welding materials. The material damage classification subnetwork uses two fully connected layers and one 1D convolutional layer. The function of the 1D convolutional layer is to extract local features of the input features, so as to obtain a more discriminative feature representation when classifying the damage degree. In the last layer, the Softmax activation function is used to output the probability distribution of multiple damage degree categories in order to classify the damage state of the material. For the anomaly detection subnetwork, a variational autoencoder structure is used, which includes an encoder and a decoder. The encoder consists of three fully connected layers, and the decoder consists of three fully connected layers. The intermediate layer generates latent variables through reparameterization techniques, so that the model can learn the latent distribution characteristics of the input data. Latent variables The generation of can be expressed as: ; in, is the mean of the encoder output, is the standard deviation of the output, and is the noise sampled from the standard normal distribution. This process enables the variational autoencoder to generate meaningful representations in the feature space. After completing the construction of each subnetwork, the trained shared feature representations are respectively input into the fatigue life prediction subnetwork, the crack growth rate estimation subnetwork, the residual strength assessment subnetwork, the material damage classification subnetwork and the anomaly detection subnetwork to obtain the initial output results of each task. Through these outputs, the fatigue performance of the material is comprehensively evaluated. In order to optimize the overall performance of the model, a multi-task loss function is constructed. The fatigue life prediction and residual strength assessment tasks use the mean square error loss, while the crack growth rate estimation task uses the mean absolute error loss. The material damage classification task uses the cross entropy loss, and the anomaly detection task uses a combination loss of reconstruction error and KL divergence. In this way, the model can balance learning between multiple tasks to ensure that the loss of each task is not too biased towards one aspect. The Adam optimizer is used to jointly optimize the shared feature extractor and each task-specific subnetwork, and a dynamic weight balancing strategy is used to adaptively adjust the weight of each task loss. This method can not only improve the efficiency of the training process, but also ensure that the importance of each task in the model is reasonably reflected. Through the joint optimization strategy, the initial fatigue performance multi-task learning model can converge quickly and achieve higher performance.

[0036] In a specific embodiment, the process of executing step S5 may specifically include the following steps: The source domain data and the target domain data are input into the fatigue performance multi-task learning model to obtain the source domain features and the target domain features, and a domain discriminator network is constructed, which includes 3 fully connected layers. Each layer uses the LeakyReLU activation function. The last layer uses the Sigmoid activation function and outputs the domain label prediction probability. Input the source domain features and target domain features into the domain discriminator network, optimize the domain discriminator parameters through back propagation, and obtain the domain discriminant loss; A gradient reversal layer is introduced into the shared feature extractor of the initial fatigue performance multi-task learning model. The feature extractor is trained to generate domain-invariant features by maximizing the domain discriminative loss, and the domain adversarial training loss is obtained. Construct a progressive domain adaptation strategy, divide the target domain data into multiple subsets, arrange them in order of similarity with the source domain from high to low, and obtain a sequence of target domain data subsets. Perform domain adversarial training on each subset in the sequence of target domain data subsets, gradually increase the proportion of target domain data in the training process, and obtain a phased domain adaptation model. A consistency regularization term is introduced into the task-specific sub-network of the staged domain adaptation model to minimize the difference between the prediction results of the source domain and the target domain, and the domain consistency loss is obtained. The domain adversarial training loss, domain consistency loss and multi-task learning loss are weightedly combined and end-to-end training is performed to obtain the target fatigue performance multi-task learning model.

[0037] Specifically, source domain data (such as fatigue test data from a certain welding material) and target domain data (such as data from another welding material) are collected and prepared, and input into the multi-task learning model to obtain source domain features. and target domain features . Build a domain discriminator network, which contains three fully connected layers, uses LeakyReLU activation function, and the last layer uses Sigmoid activation function to output the predicted probability of domain labels ,in Represents the domain label, is the input feature. The main goal of the domain discriminator is to accurately distinguish the source domain and target domain data. and target domain features After inputting the domain discriminator, the parameters of the domain discriminator are optimized through the back propagation algorithm to obtain the domain discrimination loss , the loss function is defined as: ; in, is the total number of samples, is the domain label of the sample. In order to make the shared feature extractor generate domain-invariant features, a gradient reversal layer is introduced in the initial fatigue performance multi-task learning model. The function of this layer is to multiply the loss by negative one during back propagation, thereby maximizing the domain discrimination loss. , we get the domain adversarial training loss In the process of gradually achieving domain adaptation, a progressive domain adaptation strategy is adopted to divide the target domain data into multiple subsets, sort them from high to low according to the similarity with the source domain, and obtain a sequence of target domain data subsets. This strategy allows the model to gradually introduce target domain data to perform domain adversarial training at each stage, thereby more effectively learning the characteristics of the target domain. For example, in the In the stage, the proportion of target domain data is defined as ,in along with The consistency regularization term is introduced into the task-specific sub-network in the staged domain adaptation model to minimize the source domain prediction result. And the target domain prediction results This difference is expressed as the following loss: ; in represents the L2 norm, ensuring that the prediction results of the two domains are consistent in the feature space. , domain consistency loss and multi-task learning loss (such as defined by mean square error) to obtain the total training loss . Finally, end-to-end training is performed to obtain the target fatigue performance multi-task learning model. For example, in the fatigue performance test of welding materials, the source domain data is the fatigue test results of a certain type of welding joint, and the target domain data is the test results of another type of welding joint. Through the above-mentioned domain adaptation strategy, the learning effect of the target domain features is improved, and the distribution differences between different domains can be effectively reduced, so that the generalization ability of the model can be enhanced, thereby improving the fatigue performance prediction ability of new materials.

[0038] In a specific embodiment, the process of executing step S6 may specifically include the following steps: Perform fatigue testing on the welding material to be tested, synchronously collect sound feedback signals, optical feedback signals and electrical feedback signals to obtain a second multimodal signal data set, and perform preprocessing and feature fusion on the second multimodal signal data set to obtain a comprehensive representation of the target sequence; Inputting the comprehensive representation of the target sequence into the shared feature extractor of the target fatigue performance multi-task learning model for feature extraction to obtain a real-time shared feature representation; The real-time shared feature representation is input into the fatigue life prediction sub-network, and the predicted number of fatigue life cycles is calculated through two fully connected layers and Tanh activation function; The real-time shared feature representation is input into the crack growth rate estimation subnetwork, which passes through two fully connected layers and one LSTM layer to capture the timing characteristics, and uses the Sigmoid activation function to output the normalized crack growth rate; The real-time shared feature representation is input into the residual strength evaluation sub-network, through three fully connected layers using the LeakyReLU activation function, and the last layer uses a linear activation function to output the predicted residual strength value; The real-time shared feature representation is input into the material damage classification sub-network, and local features are extracted through two fully connected layers and a 1D convolutional layer. Finally, the Softmax activation function is used to output the probability distribution of multiple damage categories. The real-time shared feature representation is input into the anomaly detection subnetwork, encoded and decoded through the variational autoencoder structure, the reconstruction error and KL divergence are calculated, and the presence of anomalies is determined to obtain anomaly detection results; The predicted fatigue life cycles, normalized crack growth rate, predicted residual strength value, probability distribution of multiple damage degree categories and abnormal detection results are comprehensively analyzed to obtain the fatigue performance evaluation results of welding materials.

[0039] Specifically, fatigue test is performed on the welding material to be tested, and sound feedback signal, optical feedback signal and electrical feedback signal are collected synchronously to obtain a second multimodal signal data set, which is expressed as ,in It's a sound signal. It is a light signal. It is an electrical signal. In order to effectively analyze these signals, the data is preprocessed, including noise reduction, wavelet transform and normalization. The second multimodal signal data set is feature fused to extract its important features and construct a comprehensive representation of the target sequence. Feature fusion is achieved in a variety of ways, such as using weighted average or concatenation techniques to extract features from each signal and combine them. The target sequence comprehensive representation is input into the shared feature extractor of the target fatigue performance multi-task learning model to extract features to generate a real-time shared feature representation. The shared feature extractor uses a multi-layer perceptron structure and is processed through several fully connected layers. The obtained feature representation provides important information for each sub-network. Input to the fatigue life prediction subnetwork, which contains two fully connected layers and is processed using the Tanh activation function to output the predicted number of fatigue life cycles , which is calculated as: ; in, and are the weight matrix and the bias vector respectively. The real-time shared feature representation Input the crack growth rate estimation subnetwork, which consists of two fully connected layers and one LSTM layer. The introduction of the LSTM layer enables the model to capture time series characteristics and output the normalized crack growth rate through the Sigmoid activation function.

[0040] ; Similarly, the real-time shared feature representation Input the residual strength evaluation subnetwork, which contains three fully connected layers using the LeakyReLU activation function, and the last layer uses a linear activation function to output the predicted residual strength value

[0041] ; The linear activation function is suitable for regression tasks because its output range is not limited. The real-time shared feature representation is input into the material damage classification subnetwork, and local features are extracted through two fully connected layers and a 1D convolutional layer. The Softmax activation function is used to output the probability distribution of multiple damage categories. , which is of the form: ; The real-time shared feature representation is input into the anomaly detection subnetwork, using a variational autoencoder structure, which consists of an encoder and a decoder. The encoder encodes the input features into latent variables , the calculation process is: ; in and are the mean and standard deviation of the encoder output, respectively. is a random variable sampled from a standard normal distribution. The decoder reconstructs the input features and calculates the reconstruction error and KL divergence , used to evaluate the quality of reconstruction and the distribution of latent variables, and finally output the anomaly detection results. In summary, the output results of all sub-networks include the predicted number of fatigue life cycles , normalized crack growth rate , predicted residual strength value , probability distribution of damage severity categories As well as abnormal detection results, this information is comprehensively analyzed to form the fatigue performance evaluation results of welding materials. Construct a comprehensive loss function , combined with the losses of each subtask:

[0042] Each loss function , , They correspond to fatigue life prediction, residual strength assessment and material damage classification respectively, and and is the loss used for anomaly detection.

[0043] In a specific embodiment, the execution step inputs the real-time shared feature representation into the anomaly detection subnetwork, encodes and decodes through the variational autoencoder structure, calculates the reconstruction error and KL divergence, determines whether there is an anomaly, and obtains the anomaly detection result. The process may specifically include the following steps: The real-time shared feature representation is input into the encoder of the anomaly detection sub-network, and the feature is compressed through three fully connected layers to obtain compressed features; The mean and variance of the compressed features are calculated to obtain the distribution parameters of the latent space, and reparameterized sampling is performed based on the distribution parameters to obtain the latent variables; The latent variables are input into the decoder, and the features are reconstructed through three fully connected layers to obtain the reconstructed features. The mean square error between the reconstructed features and the real-time shared feature representation is calculated to obtain the reconstruction error. The KL divergence between the latent variable distribution and the standard normal distribution is calculated based on the distribution parameters to obtain the KL divergence value, and the reconstruction error and the KL divergence value are weighted and summed to obtain the anomaly score; A threshold is set for the abnormal score to make a judgment. If it is higher than the threshold, it is judged as abnormal, and if it is lower than the threshold, it is judged as normal, and the abnormal detection result is obtained.

[0044] Specifically, we build an effective deep learning framework that can extract and compress features through multiple fully connected layers to facilitate subsequent anomaly detection. Real-time shared feature representation of input Processed through three fully connected layers, denoted as , and , each layer uses the ReLU activation function to increase nonlinearity, and the compressed feature is expressed as: ; in, , and is the weight matrix corresponding to the three fully connected layers, and , and are their bias vectors. After processing, the compressed features obtained It helps to reduce the dimensionality of the data while retaining meaningful information. Calculate the mean and variance to get the distribution parameters of the latent space. Define the mean and variance

[0045] ; in represents the number of samples. These parameters are used to reparameterize the sampling to generate the latent variables By introducing a noise in a standard normal distribution , the calculation formula of latent variables is: ; Through the reparameterization technique, back propagation can be optimized through the distribution of latent variables. Input decoder, the decoder also consists of three fully connected layers, generating new feature representations through feature reconstruction , the formula of the reconstruction process is: ; in, , and is the weight matrix of the fully connected layer in the decoder, , and is its bias vector. Through the decoder, the generated reconstructed features Real-time sharing of feature representations with input Compare to calculate the reconstruction error , usually measured using mean square error, namely: ; In this formula, and Represents the first elements, the reconstruction error reflects the performance of the model in the feature reconstruction process. At the same time, based on the distribution parameters and , calculate the latent variable The KL divergence between the distribution of and the standard normal distribution , which is used to quantify the difference between two distributions and is calculated as: ; in, is the dimension of the latent space, and is the mean and variance of the corresponding dimension of the latent variable. The smaller the KL divergence, the closer the distribution of the latent variable is to the standard normal distribution, which helps the training stability of the model. In order to comprehensively consider the reconstruction error and KL divergence, the two losses are weighted and summed to obtain the anomaly score : ; in, is a weighted coefficient used to balance the impact of the two losses. This anomaly score is used to determine whether the input feature is abnormal. When determining anomalies, a threshold is set. If the anomaly score L is higher than the threshold, it is considered abnormal; otherwise, it is considered normal. This design can effectively distinguish normal data from abnormal data, so that the anomaly detection subnetwork can still maintain efficient detection capabilities in complex application scenarios and obtain anomaly detection results.

[0046] The above describes the welding material fatigue performance testing method of the acoustic-optical-electrical coordinated embodiment of the present invention. The following describes the welding material fatigue performance testing system of the acoustic-optical-electrical coordinated embodiment of the present invention. Figure 2 In one embodiment of the present invention, an acoustic-optical-electrical coordinated welding material fatigue performance testing system includes: An acquisition module, used for synchronously acquiring a first multimodal signal data set of the welding material during the fatigue test, wherein the first multimodal signal data set includes sound feedback signal data, optical feedback signal data and electrical feedback signal data; A fusion module is used to fuse the sound feedback signal data, the light feedback signal data and the electrical feedback signal data to obtain a fusion feature vector; An extraction module is used to input the fused feature vector into a multivariate time series representation learning framework with a hybrid contextual attention mechanism to extract time series features and obtain a comprehensive representation of the time series; A building block for constructing and training an initial fatigue performance multi-task learning model consisting of a shared feature extractor and task-specific subnetworks based on a comprehensive representation of time series; A transfer module is used to implement domain adaptation and transfer learning on the fatigue performance multi-task learning model to obtain the target fatigue performance multi-task learning model; The test module is used to collect the second multimodal signal data set in real time and input the target fatigue performance multi-task learning model to perform fatigue life prediction, crack growth rate estimation, residual strength evaluation, material damage degree classification and anomaly detection to obtain the fatigue performance test results of the welding material.

[0047] Through the collaborative cooperation of the above-mentioned components, the efficient fusion of the three modal signals of sound, light and electricity is achieved, and a more comprehensive and robust feature representation is extracted through the physical information scattering transformation method, which improves the accuracy and reliability of fatigue performance testing. The multivariate time series representation learning framework with a hybrid contextual attention mechanism effectively captures the dependencies of multimodal signals in time and space dimensions, and enhances the model's learning ability for long-term and short-term features. A multi-task learning model containing a shared feature extractor and a task-specific subnetwork is constructed, which realizes the simultaneous optimization of multiple tasks such as fatigue life prediction, crack growth rate estimation, residual strength assessment, material damage classification and anomaly detection, and improves the comprehensive performance of the model. Through the domain adversarial neural network structure and the progressive domain adaptation strategy, the generalization ability and transfer learning ability of the model are enhanced, so that it can adapt to different welding materials and test conditions. The present invention realizes real-time acquisition and processing of multimodal signal data, and quickly obtains comprehensive fatigue performance evaluation results.

[0048] The present invention also provides a computer device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the acoustic, optical and electrical coordinated welding material fatigue performance testing method in the above-mentioned embodiments.

[0049] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the acoustic-optical-coordinated welding material fatigue performance testing method.

[0050] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0051] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0052] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for testing fatigue properties of welding materials using acoustic, optical and electrical synergy, characterized in that: The method comprises: Synchronously collecting a first multimodal signal data set of the welding material during fatigue testing, wherein the first multimodal signal data set includes sound feedback signal data, optical feedback signal data, and electrical feedback signal data; fusing the sound feedback signal data, the light feedback signal data and the electrical feedback signal data to obtain a fused feature vector; Inputting the fused feature vector into a multivariate time series representation learning framework with a hybrid contextual attention mechanism to extract time series features, thereby obtaining a comprehensive representation of the time series; constructing and training an initial fatigue performance multi-task learning model comprising a shared feature extractor and task-specific sub-networks based on the comprehensive representation of the time series; Implementing domain adaptation and transfer learning on the fatigue performance multi-task learning model to obtain a target fatigue performance multi-task learning model; The second multimodal signal data set is collected in real time and input into the target fatigue performance multi-task learning model to perform fatigue life prediction, crack growth rate estimation, residual strength evaluation, material damage degree classification and anomaly detection to obtain the fatigue performance test results of the welding material.

2. The method for testing fatigue properties of welding materials by acoustic, optical and electrical synergy according to claim 1, characterized in that: The synchronous collection of a first multimodal signal data set during fatigue testing of welding materials, wherein the first multimodal signal data set includes sound feedback signal data, optical feedback signal data, and electrical feedback signal data, includes: Fixing the welding material on a fatigue testing machine, applying a cyclic load to the welding material, and obtaining the welding material in a stress state; Arranging a sound sensor array around the welding material in the stressed state, setting a sampling frequency for the sound sensor array to obtain an original sound signal, and performing wavelet transform noise reduction on the original sound signal to obtain a noise-reduced sound signal; Performing a fast Fourier transform on the noise-reduced sound signal, converting the time domain signal into a frequency domain signal, extracting the main frequency, frequency distribution and energy distribution characteristics, and obtaining sound feedback signal data; Installing a spectrum analyzer, aiming it at the surface of the welding material in the stressed state, collecting spectrum data covering the visible light to near infrared spectrum range, obtaining original spectrum data, and performing spectrum correction on the original spectrum data to obtain optical feedback signal data; Installing a resistance strain gauge and a potential sensor on the welding material in the stressed state, measuring the resistance change and potential distribution of the material during the fatigue process, calculating the resistance change rate and the potential distribution gradient, and obtaining electrical feedback signal data; The sound feedback signal data, the light feedback signal data and the electrical feedback signal data are used as a first multimodal signal data set.

3. The method for testing fatigue properties of welding materials by acoustic, optical and electrical synergy according to claim 1, characterized in that: The fusing the sound feedback signal data, the light feedback signal data and the electrical feedback signal data to obtain a fused feature vector includes: Performing a wavelet convolution transformation on the sound feedback signal data to obtain a time-frequency representation matrix of the sound signal, performing a wavelet convolution transformation on the light feedback signal data to obtain a time-frequency representation matrix of the spectrum signal, and performing a wavelet convolution transformation on the electrical feedback signal data to obtain a time-frequency representation matrix of the electrical signal; Performing nonlinear modulus operations on the time-frequency representation matrix of the sound signal, the time-frequency representation matrix of the spectrum signal, and the time-frequency representation matrix of the electrical signal, respectively, to obtain a nonlinear feature matrix of the sound signal, a nonlinear feature matrix of the spectrum signal, and a nonlinear feature matrix of the electrical signal; Inputting the nonlinear characteristic matrix of the sound signal, the nonlinear characteristic matrix of the spectral signal and the nonlinear characteristic matrix of the electrical signal into a two-dimensional Gaussian low-pass filter to filter high-frequency information, respectively, to obtain a filtered sound signal characteristic matrix, a filtered spectral signal characteristic matrix and a filtered electrical signal characteristic matrix; A three-layer fully connected neural network is constructed as a feature fusion network. The number of input layer nodes is the total dimension of the filtered sound signal feature matrix, the filtered spectral signal feature matrix and the filtered electrical signal feature matrix. The number of hidden layer nodes is 1 / 2 of the input layer. The number of output layer nodes is 100. The Sigmoid activation function is used to obtain the fused feature vector through forward propagation calculation.

4. The method for testing fatigue properties of welding materials by acoustic-optical-electrical coordination according to claim 3 is characterized in that: The nonlinear modulus operation is performed on the time-frequency representation matrix of the sound signal, the time-frequency representation matrix of the spectrum signal and the time-frequency representation matrix of the electrical signal to obtain a nonlinear feature matrix of the sound signal, a nonlinear feature matrix of the spectrum signal and a nonlinear feature matrix of the electrical signal, respectively, including: Performing a complex transformation on the time-frequency representation matrix of the sound signal to obtain a complex matrix of the sound signal, performing a complex transformation on the time-frequency representation matrix of the spectral signal to obtain a complex matrix of the spectral signal, and performing a complex transformation on the time-frequency representation matrix of the electrical signal to obtain a complex matrix of the electrical signal; Performing square operations on the real and imaginary parts of the complex matrix of the sound signal, respectively, to obtain a square matrix of the real part of the sound signal and a square matrix of the imaginary part of the sound signal; performing square operations on the real and imaginary parts of the complex matrix of the spectral signal, respectively, to obtain a square matrix of the real part of the spectral signal and a square matrix of the imaginary part of the spectral signal; performing square operations on the real and imaginary parts of the complex matrix of the electric signal, respectively, to obtain a square matrix of the real part of the electric signal and a square matrix of the imaginary part of the electric signal; Performing element-by-element addition operation on the real part square matrix of the sound signal and the imaginary part square matrix of the sound signal to obtain a sound signal modulus square matrix, performing element-by-element addition operation on the real part square matrix of the spectral signal and the imaginary part square matrix of the spectral signal to obtain a spectral signal modulus square matrix, and performing element-by-element addition operation on the real part square matrix of the electric signal and the imaginary part square matrix of the electric signal to obtain an electric signal modulus square matrix; Performing a square root operation on the sound signal modulus square matrix to obtain a sound signal modulus value matrix, performing a square root operation on the spectral signal modulus square matrix to obtain a spectral signal modulus value matrix, and performing a square root operation on the electrical signal modulus square matrix to obtain an electrical signal modulus value matrix; Performing a nonlinear transformation on the sound signal modulus matrix to obtain a preliminary nonlinear characteristic matrix of the sound signal, performing a nonlinear transformation on the spectrum signal modulus matrix to obtain a preliminary nonlinear characteristic matrix of the spectrum signal, and performing a nonlinear transformation on the electrical signal modulus matrix to obtain a preliminary nonlinear characteristic matrix of the electrical signal; Performing L2 norm normalization on the preliminary nonlinear characteristic matrix of the sound signal to obtain a normalized nonlinear characteristic matrix of the sound signal, performing L2 norm normalization on the preliminary nonlinear characteristic matrix of the spectral signal to obtain a normalized nonlinear characteristic matrix of the spectral signal, and performing L2 norm normalization on the preliminary nonlinear characteristic matrix of the electric signal to obtain a normalized nonlinear characteristic matrix of the electric signal; Performing multi-scale and multi-directional feature extraction on the normalized nonlinear feature matrix of the sound signal to obtain a Gabor feature matrix of the sound signal, performing multi-scale and multi-directional feature extraction on the normalized nonlinear feature matrix of the spectral signal to obtain a Gabor feature matrix of the spectral signal, and performing multi-scale and multi-directional feature extraction on the normalized nonlinear feature matrix of the electrical signal to obtain a Gabor feature matrix of the electrical signal; The sound signal Gabor feature matrix is ​​locally normalized to obtain a nonlinear feature matrix of the sound signal, the spectral signal Gabor feature matrix is ​​locally normalized to obtain a nonlinear feature matrix of the spectral signal, and the electrical signal Gabor feature matrix is ​​locally normalized to obtain a nonlinear feature matrix of the electrical signal.

5. The method for testing fatigue properties of welding materials by acoustic, optical and electrical synergy according to claim 1, characterized in that: The step of inputting the fused feature vector into a multivariate time series representation learning framework with a hybrid context attention mechanism to extract time series features and obtain a comprehensive representation of the time series includes: Input the fused feature vector into the local attention module, capture the short-term dependency in the time series through the sliding window method, and obtain the local attention feature; input the fused feature vector into the global attention module, learn the long-term dependency through the self-attention mechanism, and obtain the global attention feature; The local attention feature and the global attention feature are weightedly fused to obtain a mixed context feature, and the mixed context feature is input into an autoencoder structure, a compressed representation of the time series is learned by minimizing the reconstruction error to obtain a reconstructed feature, positive sample pairs and negative sample pairs are constructed based on the mixed context feature, and representations of similar contexts are made closer in feature space by a contrastive learning method to obtain a context-consistent feature; Performing temporal smoothing on the mixed context features to ensure continuity of temporally adjacent representations and obtain temporally consistent features; Inputting the reconstruction feature, the context-consistent feature and the time-consistent feature into a multivariate time series representation learning framework, and obtaining a multi-objective optimization feature by jointly optimizing three objective functions of a reconstruction-based regression objective, a context-consistent comparative learning objective and a time-consistent comparative learning objective; Attention pooling is performed on the multi-objective optimization features, and the learned time series representation is compressed into a vector of fixed dimension to obtain a comprehensive representation of the time series.

6. The method for testing fatigue properties of welding materials by acoustic, optical and electrical synergy according to claim 1, characterized in that: The method of constructing and training an initial fatigue performance multi-task learning model including a shared feature extractor and a task-specific sub-network based on the comprehensive representation of the time series comprises: Inputting the time series comprehensive representation into a shared feature extractor, wherein the shared feature extractor adopts a multi-layer perceptron structure and includes three fully connected layers, each layer uses a ReLU activation function and Dropout regularization, to obtain a training shared feature representation; Construct a fatigue life prediction subnetwork, which includes two fully connected layers. Each layer uses the Tanh activation function, and the last layer outputs a single value, which represents the predicted number of fatigue life cycles. Construct a crack growth rate estimation subnetwork, which includes two fully connected layers and one LSTM layer. The LSTM layer is used to capture the timing characteristics of crack growth. The last layer uses the Sigmoid activation function to output the normalized crack growth rate. Construct a residual strength evaluation subnetwork, which includes 3 fully connected layers. Each layer uses the LeakyReLU activation function. The last layer uses a linear activation function to output the predicted residual strength value. Construct a material damage classification subnetwork, which includes two fully connected layers and one 1D convolutional layer. The 1D convolutional layer is used to extract local features. The last layer uses the Softmax activation function to output the probability distribution of multiple damage categories. Construct anomaly detection subnetwork, using variational autoencoder structure, including encoder and decoder. The encoder contains 3 fully connected layers, and the decoder contains 3 fully connected layers. The middle layer uses reparameterization technique to generate latent variables. Inputting the training shared feature representation into the fatigue life prediction subnetwork, crack growth rate estimation subnetwork, residual strength assessment subnetwork, material damage classification subnetwork and anomaly detection subnetwork respectively to obtain initial output results of each task; Construct a multi-task loss function. The fatigue life prediction and residual strength assessment tasks use mean square error loss, the crack growth rate estimation task uses mean absolute error loss, the material damage classification task uses cross entropy loss, and the anomaly detection task uses a combination loss of reconstruction error and KL divergence. The Adam optimizer is used to jointly optimize the shared feature extractor and each task-specific sub-network, and a dynamic weight balancing strategy is adopted to adaptively adjust the weight of each task loss to obtain an initial fatigue performance multi-task learning model.

7. The method for testing fatigue properties of welding materials by acoustic, optical and electrical synergy according to claim 6, characterized in that: The performing domain adaptation and transfer learning on the fatigue performance multi-task learning model to obtain a target fatigue performance multi-task learning model includes: Input the source domain data and the target domain data into the fatigue performance multi-task learning model to obtain the source domain features and the target domain features, and construct a domain discriminator network, which includes 3 fully connected layers, each layer uses the LeakyReLU activation function, and the last layer uses the Sigmoid activation function to output the domain label prediction probability; Inputting the source domain features and the target domain features into the domain discriminator network, optimizing the domain discriminator parameters by back propagation, and obtaining a domain discriminant loss; A gradient reversal layer is introduced into the shared feature extractor of the initial fatigue performance multi-task learning model, and the feature extractor is trained to generate domain-invariant features by maximizing the domain discrimination loss to obtain a domain adversarial training loss; Construct a progressive domain adaptation strategy, divide the target domain data into multiple subsets, arrange them in descending order of similarity with the source domain, obtain a target domain data subset sequence, and perform domain adversarial training on each subset in the target domain data subset sequence in turn, gradually increase the proportion of target domain data in the training process, and obtain a phased domain adaptation model; A consistency regularization term is introduced into the task-specific subnetwork of the staged domain adaptation model to minimize the difference between the prediction results of the source domain and the target domain, thereby obtaining a domain consistency loss. The domain adversarial training loss, the domain consistency loss and the multi-task learning loss are weightedly combined, and end-to-end training is performed to obtain a target fatigue performance multi-task learning model.

8. The method for testing fatigue properties of welding materials by acoustic, optical and electrical synergy according to claim 7, characterized in that: The real-time acquisition of the second multimodal signal data set and inputting it into the target fatigue performance multi-task learning model to perform fatigue life prediction, crack growth rate estimation, residual strength evaluation, material damage degree classification and anomaly detection to obtain the fatigue performance test results of the welding material, including: Performing fatigue testing on the welding material to be tested, synchronously collecting sound feedback signals, optical feedback signals and electrical feedback signals to obtain a second multimodal signal data set, and performing preprocessing and feature fusion on the second multimodal signal data set to obtain a comprehensive representation of the target sequence; Inputting the target sequence comprehensive representation into the shared feature extractor of the target fatigue performance multi-task learning model for feature extraction to obtain a real-time shared feature representation; Inputting the real-time shared feature representation into the fatigue life prediction subnetwork, and calculating the predicted number of fatigue life cycles through two fully connected layers and a Tanh activation function; The real-time shared feature representation is input into the crack growth rate estimation subnetwork, passes through two fully connected layers and one LSTM layer to capture the timing characteristics, and uses the Sigmoid activation function to output the normalized crack growth rate; The real-time shared feature representation is input into the residual strength evaluation subnetwork, passes through three fully connected layers using the LeakyReLU activation function, and the last layer uses a linear activation function to output the predicted residual strength value; The real-time shared feature representation is input into the material damage classification subnetwork, local features are extracted through two fully connected layers and a 1D convolutional layer, and finally a Softmax activation function is used to output the probability distribution of multiple damage categories; The real-time shared feature representation is input into the anomaly detection subnetwork, encoded and decoded through the variational autoencoder structure, the reconstruction error and KL divergence are calculated, whether an anomaly exists is determined, and an anomaly detection result is obtained; The predicted number of fatigue life cycles, the normalized crack growth rate, the predicted residual strength value, the probability distribution of the multiple damage degree categories and the abnormal detection result are comprehensively analyzed to obtain the fatigue performance evaluation result of the welding material.

9. The method for testing fatigue properties of welding materials by acoustic, optical and electrical synergy according to claim 8, characterized in that: The real-time shared feature representation is input into the anomaly detection subnetwork, encoded and decoded through the variational autoencoder structure, the reconstruction error and KL divergence are calculated, and whether there is an anomaly is determined to obtain an anomaly detection result, including: Input the real-time shared feature representation into the encoder of the anomaly detection subnetwork, and perform feature compression through three fully connected layers to obtain compressed features; Calculating the mean and variance of the compressed features to obtain distribution parameters of the latent space, and performing reparameterized sampling based on the distribution parameters to obtain latent variables; Input the latent variables into a decoder, perform feature reconstruction through three fully connected layers to obtain a reconstructed feature, and perform mean square error calculation between the reconstructed feature and the real-time shared feature representation to obtain a reconstruction error; Calculating the KL divergence between the latent variable distribution and the standard normal distribution based on the distribution parameters to obtain a KL divergence value, and performing weighted summation of the reconstruction error and the KL divergence value to obtain an anomaly score; A threshold is set for the abnormal score to be judged. If it is higher than the threshold, it is judged as abnormal, and if it is lower than the threshold, it is judged as normal, thereby obtaining an abnormal detection result.

10. An acoustic, optical and electrical coordinated welding material fatigue performance testing system, characterized in that: The method for testing the fatigue performance of welding materials by acoustic-optical-electrical coordination according to any one of claims 1 to 9, wherein the system comprises: An acquisition module, used for synchronously acquiring a first multimodal signal data set of the welding material during the fatigue test, wherein the first multimodal signal data set includes sound feedback signal data, optical feedback signal data and electrical feedback signal data; A fusion module, used for fusing the sound feedback signal data, the light feedback signal data and the electrical feedback signal data to obtain a fusion feature vector; An extraction module, used for inputting the fused feature vector into a multivariate time series representation learning framework with a hybrid context attention mechanism to extract time series features and obtain a comprehensive representation of the time series; A building module for building and training an initial fatigue performance multi-task learning model comprising a shared feature extractor and task-specific sub-networks based on the time series comprehensive representation; A migration module, used for performing domain adaptation and migration learning on the fatigue performance multi-task learning model to obtain a target fatigue performance multi-task learning model; The test module is used to collect the second multimodal signal data set in real time and input it into the target fatigue performance multi-task learning model to perform fatigue life prediction, crack growth rate estimation, residual strength evaluation, material damage degree classification and anomaly detection to obtain the fatigue performance test results of the welding material.

Images