An ultrasonic signal classification and recognition method, medium and device

Abstract

The application discloses an ultrasonic signal classification and identification method, medium and equipment, and belongs to the technical field of signal processing and deep learning; the method comprises the following steps: collecting ultrasonic signals of excitation channels and receiving channels under different working conditions in structural ultrasonic detection, reconstructing channel signals into time sequence samples; inputting the time sequence samples into a double-branch deep learning model, and outputting working condition category probability distribution results; the double-branch deep learning model obtains the working condition category probability distribution results through a time domain feature extraction module, a space domain feature extraction module, a feature fusion module and a classification and prediction module; the application fully considers and utilizes the excitation and receiving signal features in ultrasonic detection, adopts a time and space domain feature fusion mode to fully capture multi-scale feature information and perform deep-level fusion, solves the problem that a single network is difficult to represent ultrasonic signal features from a multi-scale level, has strong engineering practicability, and can be widely applied in the fields of structural nondestructive detection, structural health monitoring and the like.

Description

Technical Field

[0001] This application relates to the fields of signal processing and deep learning technology, and in particular to a method, medium and device for ultrasonic signal classification and recognition. Background Technology

[0002] Ultrasonic signal classification is one of the core technologies for structural nondestructive testing and damage identification. By analyzing and classifying the characteristics of ultrasonic signals under different damage states, the health status of structures can be identified. However, ultrasonic signals have typical temporal characteristics, and the large data scale and randomness of the data increase the difficulty of acquiring and analyzing feature information, resulting in low accuracy of traditional unsupervised and semi-supervised classification and failure to meet the needs of real-time classification. Deep learning-based sample recognition constructs a deep neural network framework to extract feature information from the original input data and perform abstract expression and mapping, thereby achieving the purpose of autonomous learning and classification recognition. However, existing deep learning models are limited in their ability to fully extract damage feature information representing the signal due to the single data input format. Furthermore, existing deep learning models usually require a sufficient number of original samples for training to ensure recognition accuracy, and current testing techniques are insufficient to meet the actual needs of large-scale samples.

[0003] Therefore, the overfitting problem caused by the inability to extract multi-scale feature information from a single data form and fuse features of different scales in the process of ultrasonic signal classification and recognition has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, medium, and device for ultrasonic signal classification and recognition to address the aforementioned technical problems.

[0005] The following technical solution is adopted in this specification: An ultrasound signal classification and recognition method, comprising: The ultrasonic signals from the excitation and receiving channels under different working conditions during structural ultrasonic testing are collected, and the signals from each channel are reconstructed into time-series samples.

[0006] The time series samples are input into a dual-branch deep learning model, which outputs the probability distribution results of the working condition categories.

[0007] The category with the highest probability in the probability distribution of the working condition categories is selected as the working condition classification result.

[0008] The dual-branch deep learning model includes a temporal feature extraction module, a spatial feature extraction module, a feature fusion module, and a classification prediction module. The temporal feature extraction module uses a multi-branch bidirectional long short-term memory (BiLSTM) network structure to extract temporal correlation features of ultrasonic signals at different scales, and concatenates the temporal correlation features output by each branch to obtain multi-scale temporal features. The spatial feature extraction module uses a multi-branch one-dimensional convolutional structure to extract spatial features of ultrasonic signals containing local amplitude and waveform structure, and concatenates the spatial features output by each branch to obtain multi-scale spatial features. The feature fusion module performs cross-correlation and adaptive weighted fusion of multi-scale temporal and multi-scale spatial features through multi-head cross-attention. The fused features are then concatenated with the multi-scale temporal and multi-scale spatial features and subjected to convolutional dimensionality reduction to obtain coupled temporal and spatial core features. For these coupled temporal and spatial core features, the classification prediction module sequentially performs layer normalization, pooling, fully connected spatial mapping, and random discarding operations, followed by probability normalization and category determination operations using a classification layer to obtain the predicted probability distribution of the working condition category.

[0009] Furthermore, before inputting the time series samples into the dual-branch deep learning model, preprocessing of the time series samples is also included, specifically comprising the following steps: The ultrasonic signal from the receiving channel is reconstructed to have the same length as the ultrasonic signal from the excitation channel, resulting in multiple processed signal matrices. These processed signal matrices are then normalized and mapped to a higher dimension.

[0010] Furthermore, the temporal feature extraction module specifically includes multiple BiLSTM sub-branches and a first connection layer, Concat.

[0011] Each BiLSTM sub-branch selects a different hidden size and contains a BiLSTM layer and a sub-connection layer.

[0012] For the feature matrix, temporal correlation feature extraction is performed through a bidirectional long short-term memory network with different scales of BiLSTM sub-branches. Then, the sub-connection layers of the sub-branch are used for hop-through connections to obtain the preliminary temporal features of the scale corresponding to the sub-branch. For multiple preliminary temporal features of different scales, the first connection layer is used for splicing and fusion to obtain multi-scale temporal features.

[0013] Furthermore, the spatial feature extraction module specifically includes multiple convolutional neural network (CNN) sub-branches and second connection layers.

[0014] Each CNN sub-branch includes multiple one-dimensional convolutional layers (Conv1D), and the convolutional layer parameters are different in different CNN sub-branches.

[0015] For the feature matrix, multi-scale local spatial feature extraction is performed using multiple convolutional neural networks with different convolutional parameters to obtain multiple preliminary spatial features at different scales; for the multiple preliminary spatial features at different scales, a splicing and fusion operation is performed using the second connection layer to obtain multi-scale spatial features.

[0016] Furthermore, the feature fusion module specifically includes a cross-self-aggregating attention unit (CSAA), a third connection layer, a one-dimensional convolutional layer, and a multi-head self-attention layer; the CSAA unit specifically includes a multi-head cross-attention mechanism layer and a multi-head self-attention mechanism layer.

[0017] Cross-aggregation attention units perform cross-domain feature association and adaptive weighted focusing operations to obtain fused features.

[0018] For the fused features, multi-scale temporal features, and multi-scale spatial features, a concatenation and recombination operation is performed through a third connection layer to obtain the recombined fused features. For the recombined fused features, a feature dimensionality reduction operation is performed through a one-dimensional convolutional layer to obtain the dimensionality-reduced features. For the dimensionality-reduced features, a global association modeling operation is performed through a multi-head self-attention layer to establish the global dependency relationship of the feature sequence, thereby obtaining the temporal and spatial coupled core features.

[0019] Furthermore, before inputting the time series samples into the dual-branch deep learning model, the training of the dual-branch deep learning model is also included, specifically including the following steps: constructing a sample matrix dataset containing ultrasonic signals from each channel and working condition category labels; inputting the sample matrix dataset into the dual-branch deep learning model to perform spatiotemporal feature learning and mapping, and based on the loss curve, accuracy curve, and confusion matrix, performing adaptive dynamic parameter adjustment and adaptive dynamic weight adjustment until the model's accuracy reaches a first threshold.

[0020] Furthermore, after selecting the category corresponding to the highest probability in the probability distribution of the operating condition categories as the operating condition classification result and outputting it, the method further includes using the operating condition classification result to augment the time series samples, specifically including the following steps: If the working condition classification result exceeds the second threshold, then perform data augmentation on the corresponding ultrasonic signals of each channel and the working condition label, and store them in the time series sample; otherwise, determine that the predicted classification and identification result is unreliable.

[0021] Furthermore, the data augmentation operation includes at least one of: multiple rounds of equally spaced sampling or sliding window interval sampling.

[0022] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0023] This specification provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0024] The beneficial effects of this invention are as follows: Compared with the prior art, this invention utilizes a convolutional neural network in a dual-branch deep learning model to extract local spatial features, a long short-term memory neural network to extract temporal correlation features, an attention mechanism to fuse cross-branch features, and a connection layer to retain feature information at different scales; and performs data augmentation operations on the ultrasonic signals and condition labels corresponding to the probability distribution results of the working condition categories, expands the amount of data, and stores it in a sample matrix dataset; thus enhancing the high-dimensional representation capability of the data and improving the ability to extract multi-scale feature information from a single data form and fuse features at different scales. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating an embodiment of the present invention; Figure 2 This is a schematic diagram of the dual-branch deep learning model architecture of the present invention; Figure 3 This is a schematic diagram of the temporal feature extraction module architecture of the present invention; Figure 4 This is a schematic diagram of the spatial feature extraction module architecture of the present invention; Figure 5 This is a schematic diagram of the feature fusion module architecture of the present invention; Figure 6 This is a schematic diagram of the classification prediction module architecture of the present invention; Figure 7 This is a network architecture diagram of an embodiment of the present invention; Figure 8 This is a loss curve diagram of the present invention; Figure 9 This is an accuracy curve of the present invention; Figure 10 This is the confusion matrix diagram of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0027] This invention provides an ultrasonic signal classification and recognition method, comprising acquiring ultrasonic signals from excitation and receiving channels under different working conditions in structural ultrasonic testing, and reconstructing the signals of each channel into time-series samples. The time-series samples are input into a two-branch deep learning model, which outputs a probability distribution of working condition categories. The category corresponding to the highest probability in the probability distribution is selected as the working condition classification result for output. This method utilizes convolutional neural networks in the two-branch deep learning model to extract local spatial features, long short-term memory neural networks to extract temporal correlation features, attention mechanisms to fuse cross-branch features, and connection layers to retain feature information at different scales. Furthermore, the ultrasonic signals of each channel and the working condition labels corresponding to the probability distribution of working condition categories are augmented to increase the data volume before being stored in a sample matrix dataset. This enhances the high-dimensional representation capability of the data, improves the ability to extract multi-scale feature information from a single data form and fuse features at different scales, and achieves effective augmentation of sample data.

[0028] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0029] This invention proposes an ultrasonic signal classification and recognition method, including, as follows: Figure 1 As shown, the implementation of this method includes the following steps: Step 1: Collect ultrasonic signals from the excitation and receiving channels under different working conditions during structural ultrasonic testing, and reconstruct the signals from each channel into time series samples.

[0030] Step 2: Input the time series samples into the dual-branch deep learning model and output the probability distribution results of the working condition categories.

[0031] refer to Figure 2 and Figure 7The dual-branch deep learning model includes a temporal feature extraction module, a spatial feature extraction module, a feature fusion module, and a classification prediction module. The temporal feature extraction module uses a multi-branch bidirectional long short-term memory (BiLSTM) network structure to extract temporal correlation features of ultrasound signals at different scales, and concatenates the temporal correlation features output from each branch to obtain multi-scale temporal features. The spatial feature extraction module uses a multi-branch one-dimensional convolutional structure to extract spatial features of ultrasound signals containing local amplitude and waveform structure, and concatenates the spatial features output from each branch to obtain multi-scale spatial features. The feature fusion module performs cross-correlation and adaptive weighted fusion of multi-scale temporal and spatial features through multi-head cross-attention, concatenates the fused features with the multi-scale temporal and spatial features, and performs convolutional dimensionality reduction to obtain coupled temporal and spatial core features. (Reference) Figure 6 For the core features of temporal and spatial coupling, the classification prediction module sequentially performs layer normalization, pooling, fully connected spatial mapping and random discarding operations. Then, the classification layer is used to perform probability normalization and category determination operations to obtain the prediction results of the probability distribution of working condition categories.

[0032] Optionally, before inputting the time series samples into the dual-branch deep learning model, preprocessing of the time series samples is also included, specifically including the following steps: The ultrasonic receiving channel signal to be identified is reconstructed to have the same length as the ultrasonic excitation channel signal to be identified, resulting in multiple processed signal matrices. These processed signal matrices are then normalized and mapped to a high-dimensional space to obtain feature matrices representing different feature dimensions using different spatial tensors.

[0033] refer to Figure 3 Optionally, the temporal feature extraction module specifically includes two BiLSTM sub-branches and a first connection layer, Concat.

[0034] Each BiLSTM sub-branch selects a different hidden size, and all contain a BiLSTM layer and a sub-connection layer.

[0035] For the feature matrix, temporal correlation feature extraction is performed through a bidirectional long short-term memory network with different scales of BiLSTM sub-branches. Then, the sub-connection layers of the sub-branch are used for hop-through connections to obtain the preliminary temporal features of the scale corresponding to the sub-branch. For multiple preliminary temporal features of different scales, the first connection layer is used for splicing and fusion to obtain multi-scale temporal features.

[0036] refer to Figure 4 Optionally, the spatial feature extraction module specifically includes two convolutional neural network (CNN) sub-branches and a second connection layer.

[0037] Each CNN sub-branch includes a one-dimensional convolutional layer Conv1D, and the parameters of the convolutional layer are different in different CNN sub-branches.

[0038] For the feature matrix, multi-scale local spatial features are extracted by using multiple convolutional neural networks with different convolutional parameters to obtain multiple preliminary spatial features at different scales. For the multiple preliminary spatial features at different scales, a second connection layer is used to perform a splicing and fusion operation to obtain multi-scale spatial features.

[0039] refer to Figure 5 The feature fusion module specifically includes a cross-self-aggregating attention unit (CSAA), a third connection layer, a one-dimensional convolutional layer, and a multi-head self-attention layer; the CSAA unit specifically includes a multi-head cross-attention mechanism layer and a multi-head self-attention mechanism layer.

[0040] For the fused features, multi-scale temporal features, and multi-scale spatial features, a third connection layer is used to perform a splicing and recombination operation to obtain the recombined fused features. For the recombined fused features, a one-dimensional convolutional layer is used to perform feature dimensionality reduction to obtain dimensionality-reduced features. For the dimensionality-reduced features, a multi-head self-attention layer is used to perform global association modeling to establish global dependencies of feature sequences, thereby obtaining core features coupled in the temporal and spatial domains.

[0041] refer to Figure 8 , Figure 9 and Figure 10 Before inputting time series samples into the dual-branch deep learning model, the training of the dual-branch deep learning model is also included, specifically including the following steps: constructing a sample matrix dataset containing ultrasonic signals from each channel and working condition category labels; inputting the sample matrix dataset into the dual-branch deep learning model to perform spatiotemporal feature learning and mapping, and comparing the difference between the model's prediction results and the actual results based on the loss curve, accuracy curve, and confusion matrix, and performing adaptive dynamic parameter adjustment and adaptive dynamic weight adjustment until the model's accuracy reaches the first threshold.

[0042] Step 3: Select the category with the highest probability in the probability distribution of the working condition categories as the working condition classification result and output it.

[0043] Optionally, after selecting the maximum value from the probability distribution results of the working condition categories as the working condition classification result for output, the method further includes using the working condition classification result to augment the time series samples, thereby achieving effective expansion of the sample data. This specifically includes the following steps: If the working condition classification result exceeds the second threshold, then the corresponding ultrasonic signals of each channel and the working condition label are sampled at multiple equal intervals or at sliding window intervals for data augmentation and then stored in the time series sample; otherwise, the predicted classification and recognition result is deemed unreliable.

[0044] Example The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. In this embodiment, shear studs in steel-concrete composite beams are used as the research object. Ultrasonic signals of studs under different health conditions are collected to train a model and complete the classification task. (Referencing...) Figure 1 Specifically, it includes the following steps: S1. Fabricate and cure steel-concrete composite beam components under different working conditions: no damage, 50% damage, and 100% damage. This experiment adopts a single-transmitter, single-receiver dual-channel method to record the excitation channel signal and the receiving channel signal.

[0045] S2. To enhance the ultrasound signal, the dataset is expanded using a multi-round, equally spaced sampling method, with an enhancement factor of 5. The specific processing method is as follows: A sampling step size of 5 is set. Starting from the 1st, 2nd, 3rd, 4th, and 5th data points of the sequence, a sampling point is selected every 5 positions along the time sequence direction. This method divides the same original sequence into 5 non-overlapping subsequences. The input and received signals undergo the same processing. Each signal is sampled for 2000 time steps, with a length of 4000, resulting in a sample A of the following form: ; in, For the first The stimulus data points at each time step For the first The received data points at each time step.

[0046] The data is classified and labeled to obtain the input samples. The obtained data is divided into training set, validation set and test set. In this embodiment, the number of samples is: 100 original training samples, 35 validation samples and 30 test samples, which is in the category of small samples.

[0047] The sample data is standardized using a standard scaler to eliminate the influence of dimensions and convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0048] S3. Construct a multi-branch classification prediction module The model input is a standardized, discretized wave enhancement signal tensor matrix with class labels: Each sample contains 2000 time steps, and each time step has two feature parameters: excitation and reception.

[0049] The output is a class probability distribution. The model structure includes a temporal feature extraction module, a spatial feature extraction module, a feature fusion module, and a classification prediction module. Considering the dataset size, L2 regularization is added to mitigate overfitting caused by small sample sizes. (See reference...) Figure 7 The structure of each network module used in this embodiment is as follows: The temporal feature extraction module consists of a bidirectional long short-term memory network (BiLSTM) sub-branch and a connection layer (Concat). The BiLSTM sub-branch has two BiLSTM layers, and the outputs of the two BiLSTMs are concatenated. The feature information is then directly fed into the feature fusion module after one max pooling operation.

[0050] The spatial feature extraction module consists of a sub-branch of a convolutional neural network (CNN). The CNN sub-branch has two one-dimensional convolutional layers (Conv1D). After passing through the second Conv1D, the feature information directly enters the feature fusion module.

[0051] The feature fusion module, connected after the temporal feature extraction module and the spatial feature extraction module, consists of a Cross-Self-Aggregation Attention (CSAA) unit, a Concat layer, a Conv1D layer, and a multi-head self-attention layer.

[0052] The classification prediction module consists of a normalization layer (LN), a global average pooling layer, a fully connected layer, a random discard layer, and a classification layer.

[0053] The specific network parameters are shown in Table 1: Table 1 Network Parameter Table

[0054] S4. Model Iterative Training and Testing The model uses the Adam optimizer, reference Figure 8 The loss function is the classification cross-entropy function; during model training, a cosine decay learning rate is used to avoid the later convergence difficulties or training oscillations caused by a fixed learning rate, and the validation set accuracy is used as an indicator to evaluate the model performance.

[0055] Initial learning rate In this embodiment, the batch size is set to 32, and the decay step number T=3000 is set in combination with the amount of data. The training is planned for 200 epochs. An early stopping strategy is set to monitor the loss of the validation set, with a patience value of 30 and a minimum change of 0.001.

[0056] Regarding loss, the model loss decreased rapidly during the first 50 epochs, with training loss and validation loss reaching 0.8 and 0.95 respectively. From epochs 50 to 160, the loss decreased slowly and steadily, stabilizing after epoch 160, with training loss and validation loss stabilizing at 0.5 and 0.73 respectively. Regarding accuracy, the model accuracy increased rapidly during the first 25 epochs, with both training and validation accuracy at 0.7. From epochs 25 to 130, the accuracy increased slowly and steadily, stabilizing after epoch 130, with training accuracy and validation accuracy reaching 0.88 and 0.85 respectively. The model weights with the highest validation set accuracy were preserved.

[0057] The generalization ability of the model was tested using test data, as shown in the confusion matrix diagram. There were 30 test data points in total. Among them, 26 predicted the same class as the true class, 1 case of 50% damage was predicted as no damage, 1 case of 100% damage was predicted as no damage, and 2 cases of 100% damage were predicted as 50% damage. The test accuracy was 86.7%.

[0058] This specification also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0059] This specification provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0060] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0061] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for classifying and recognizing ultrasonic signals, characterized in that, include: Ultrasonic signals from the excitation and receiving channels under different working conditions during structural ultrasonic testing are collected, and the signals from each channel are reconstructed into time-series samples. The time series samples are input into a dual-branch deep learning model, which outputs the probability distribution results of the working condition categories. The category with the highest probability in the probability distribution of the working condition categories is selected as the working condition classification result; The dual-branch deep learning model includes a temporal feature extraction module, a spatial feature extraction module, a feature fusion module, and a classification prediction module. The temporal feature extraction module uses a multi-branch bidirectional long short-term memory (BiLSTM) network structure to extract temporal correlation features of ultrasonic signals at different scales, and concatenates the temporal correlation features output by each branch to obtain multi-scale temporal features. The spatial feature extraction module uses a multi-branch one-dimensional convolutional structure to extract spatial features of ultrasonic signals containing local amplitude and waveform structure, and concatenates the spatial features output by each branch to obtain multi-scale spatial features. The feature fusion module performs cross-correlation and adaptive weighted fusion of multi-scale temporal and multi-scale spatial features through multi-head cross-attention. The fused features are then concatenated with the multi-scale temporal and multi-scale spatial features and subjected to convolutional dimensionality reduction to obtain coupled temporal and spatial core features. For these coupled temporal and spatial core features, the classification prediction module sequentially performs layer normalization, pooling, fully connected spatial mapping, and random discarding operations, followed by probability normalization and category determination operations using a classification layer to obtain the predicted probability distribution of the working condition category.

2. The ultrasonic signal classification and recognition method as described in claim 1, characterized in that, Before inputting the time series samples into the dual-branch deep learning model, preprocessing of the time series samples is also included, specifically the following steps: The ultrasonic signal from the receiving channel is reconstructed to have the same length as the ultrasonic signal from the excitation channel, resulting in multiple processed signal matrices. These processed signal matrices are then normalized and mapped to a higher dimension.

3. The ultrasonic signal classification and recognition method as described in claim 1, characterized in that: The temporal feature extraction module specifically includes multiple BiLSTM sub-branches and a first connection layer, Concat. Each BiLSTM sub-branch selects a different hidden size and contains a BiLSTM layer and a sub-connection layer; For the time series samples, temporal correlation features are extracted using a bidirectional long short-term memory network with different scales of BiLSTM sub-branches. Then, the sub-connection layers of the sub-branches are used for hierarchical connections to obtain the preliminary temporal features corresponding to the scale of the sub-branch. For multiple preliminary temporal features at different scales, the first connection layer is used for splicing and fusion to obtain multi-scale temporal features.

4. The ultrasonic signal classification and recognition method as described in claim 1, characterized in that: The spatial feature extraction module specifically includes multiple convolutional neural network (CNN) sub-branches and second connection layers; Each CNN sub-branch includes multiple one-dimensional convolutional layers (Conv1D), and the convolutional layer parameters are different in different CNN sub-branches. For the time series samples, multi-scale local spatial features are extracted using multiple convolutional neural networks with different convolutional parameters to obtain multiple preliminary spatial features at different scales. For the multiple preliminary spatial features at different scales, a splicing and fusion operation is performed using the second connection layer to obtain multi-scale spatial features.

5. The ultrasonic signal classification and recognition method as described in claim 1, characterized in that: The feature fusion module specifically includes a cross-self-aggregating attention unit (CSAA), a third connection layer, a one-dimensional convolutional layer, and a multi-head self-attention layer; the CSAA unit specifically includes a multi-head cross-attention mechanism layer and a multi-head self-attention mechanism layer. Cross-aggregation attention units perform cross-domain feature association and adaptive weighted focusing operations to obtain fused features; For the fused features, multi-scale temporal features, and multi-scale spatial features, a concatenation and recombination operation is performed through a third connection layer to obtain the recombined fused features. For the recombined fused features, a feature dimensionality reduction operation is performed through a one-dimensional convolutional layer to obtain the dimensionality-reduced features. For the dimensionality-reduced features, a global association modeling operation is performed through a multi-head self-attention layer to establish the global dependency relationship of the feature sequence, thereby obtaining the temporal and spatial coupled core features.

6. The ultrasonic signal classification and recognition method as described in claim 1, characterized in that, Before inputting the time series samples into the dual-branch deep learning model, the training of the dual-branch deep learning model is also included, specifically including the following steps: constructing a sample matrix dataset containing ultrasonic signals from each channel and working condition category labels; inputting the sample matrix dataset into the dual-branch deep learning model to perform spatiotemporal feature learning and mapping, and based on the loss curve, accuracy curve, and confusion matrix, performing adaptive dynamic parameter adjustment and adaptive dynamic weight adjustment until the model's accuracy reaches a first threshold.

7. The ultrasonic signal classification and recognition method as described in claim 1, characterized in that, After selecting the category with the highest probability from the probability distribution of the working condition categories as the working condition classification result for output, the method further includes using the working condition classification result to augment the time series sample data, specifically including the following steps: If the working condition classification result exceeds the second threshold, then perform data augmentation on the corresponding ultrasonic signals of each channel and the working condition label, and store them in the time series sample; otherwise, determine that the predicted classification and identification result is unreliable.

8. The ultrasonic signal classification and recognition method as described in claim 7, characterized in that, The data augmentation operation includes at least one of: multiple rounds of equally spaced sampling or sliding window interval sampling.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 8.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 8.

Images