Acoustic emission source positioning method based on lightweight convolution and attention mechanism

Abstract

The present application relates to the technical field of acoustic emission damage detection, in particular to an acoustic emission source positioning method based on light-weight convolution and attention mechanism, wherein a four-channel continuous wavelet image (CWT) is taken as input, a DSConv module decomposes standard convolution into channel-by-channel convolution and point-by-point convolution, effectively reducing network parameter quantity and calculation overhead; an attention module adaptively adjusts feature weight through joint channel and spatial attention, highlights key time-frequency features and suppresses redundant information; four-branch features integrate multi-channel complementary information through adaptive weighted fusion, improving signal representation capability; residual connection improves gradient propagation and enhances deep feature learning efficiency, thereby realizing light-weight, high-precision acoustic emission source positioning.

Description

Technical Field

[0001] This invention relates to the field of acoustic emission damage detection technology, and in particular to an acoustic emission source localization method based on lightweight convolution and attention mechanisms. Background Technology

[0002] In fields such as power equipment, cable partial discharge monitoring, and structural damage detection, traditional location methods rely on manual inspection or simple instrument measurements, which are inefficient and difficult to accurately pinpoint the damage source. With the development of acoustic emission technology, damage location methods based on acoustic emission signals have become an effective means. This technology captures stress waves generated within the material due to damage and uses a sensor array to receive these waves to calculate the coordinates of the acoustic emission source. However, existing acoustic emission source location methods have limitations in terms of noise interference, computational complexity, feature extraction, and generalization ability. For example, the time delay estimation (TDOA) method described in "Acoustic Emission Source Location Technology" relies on the time difference of signal arrival at different sensors, but it is susceptible to interference in high-noise environments and requires precise calibration of sensor clock synchronization; traditional methods typically rely on manual feature extraction, which is difficult to adapt to complex environments, resulting in limited location accuracy.

[0003] In recent years, deep learning technology has been increasingly introduced into acoustic emission source localization tasks. In particular, the successful application of convolutional neural networks (CNNs) in image and signal processing has provided new possibilities for feature extraction of acoustic emission signals. However, existing methods still suffer from limited feature extraction capabilities, insufficient noise resistance, and weak generalization ability. For example, traditional CNNs use fixed-size convolutional kernels, making it difficult to flexibly adapt to the non-stationary characteristics of acoustic emission signals, resulting in insufficient feature extraction. 2D-CNN models (see "Single-Sensor Acoustic Emission Source Localization in Plate-Like Structures Using Deep Learning") utilize only single-channel signals, failing to fully leverage multi-sensor information, leading to ineffective learning of spatial features. The AESLNET model (see "Research On The Acoustic Emission Source Localization Methodology In Composite Materials Based On Artificial Intelligence") performs well in composite materials, but this model is primarily trained for specific materials, exhibiting weak generalization ability on metallic materials, resulting in decreased localization accuracy.

[0004] Furthermore, the physical properties of acoustic emission signals share commonalities across different materials: stress waves generated by crack propagation in metallic materials and shock waves from delamination defects in composite materials both exhibit high-frequency transient characteristics (frequency range 100kHz-1MHz) and short-term decay patterns (millisecond-level duration). This commonality provides theoretical feasibility for deep learning methods based on signal time-frequency feature modeling.

[0005] Currently, mainstream methods for locating acoustic emission sources include time delay estimation-based methods, pattern matching-based methods, and data-driven deep learning methods. Among them, time delay estimation-based methods rely on the time difference of signals arriving at different sensors, but are susceptible to interference in high-noise environments; pattern matching-based methods require building a large database of matching samples, resulting in significant computational costs; while traditional CNNs, due to their lack of spatial information modeling capabilities, still have limited localization accuracy in complex signal environments. Summary of the Invention

[0006] To address the limitations of existing acoustic emission source localization methods in high-noise environments and the lack of effective modeling of complex signal spatiotemporal characteristics in traditional convolutional neural networks, which restricts localization accuracy, this invention proposes an acoustic emission source localization method based on lightweight convolution and attention mechanisms. Utilizing an improved acoustic emission source localization network (AE-DSCBRNet), based on a four-branch convolutional structure, and combining depthwise separable convolutional (DSConv) blocks, residual convolutional blocks, and robust channel and spatial attention mechanisms, this method is used for the precise localization of acoustic emission sources in plate-like structures.

[0007] This invention is achieved through the following technical solution:

[0008] A method for locating acoustic emission sources based on lightweight convolution and attention mechanisms includes the following steps:

[0009] Step 1: Acoustic emission signals are collected at sampling grid points on the surface of the steel plate using an acoustic emission piezoelectric ceramic sensor;

[0010] Step 2: Perform continuous wavelet transform processing on the acoustic emission signal to convert the acoustic emission time series signal into time-frequency domain image data, and normalize the image.

[0011] Step 3: Integrate the image data and the corresponding acoustic emission source coordinates into a dataset, divide it into a training set and a test set, build an improved four-branch CNN network (AE-DSCBRNet) model and train it;

[0012] Step 4: Input the test set into the trained network model to obtain the coordinates of the acoustic emission source location, thus realizing the localization of the acoustic emission source.

[0013] Furthermore, in step 1, a sampling network is divided on the surface of the steel plate, and acoustic emission sensors are arranged at the four vertices of the sampling network to form a sensor network; a signal generator is used at the grid points of the sampling to simulate an acoustic emission source and collect acoustic emission signals.

[0014] Furthermore, in step 2, the acoustic emission signal undergoes continuous wavelet transform processing, wherein the wavelet function selected is the Morse wavelet, and the formula for continuous wavelet transform processing is:

[0015]

[0016] in, It is a continuous complex function in the time and frequency domains, called the mother wavelet; a is the scaling factor, which controls the scaling of the mother wavelet, and b is the shift factor, which controls the position of the wavelet on the time axis.

[0017] Furthermore, in step 2, in order to eliminate the amplitude differences of signals acquired by different sensors and improve the numerical stability of the CWT time-frequency image, the generated continuous wavelet transform (CWT) time-frequency image was normalized. The normalization formula is as follows:

[0018]

[0019] in, This is the original image. min and max represent the minimum and maximum pixel values, respectively. Through normalization, we can ensure that the numerical range of all time-frequency features is consistent, thus avoiding the impact of uneven data distribution on network training.

[0020] Furthermore, the AE-DSCBRNet model includes a data input module, a feature extraction module, a feature fusion module, and an output module;

[0021] The data input module contains four input branches, each receiving a continuous wavelet image data;

[0022] The feature extraction module consists of two consecutive depthwise separable convolutional (DSConv) blocks, a residual convolutional block, and a max pooling layer. The DSConv layer extracts local features from the input time-frequency image, the depthwise convolution captures local patterns in the spatial dimension, and the pointwise convolution then linearly combines the features of each channel, thereby significantly reducing the number of parameters and computational cost while preserving the main time-frequency features of the input image. The residual convolutional block further learns the extracted features.

[0023] The residual convolutional block consists of two consecutive DSConv convolutional layers, and the input and output are added through skip connections to achieve the fusion of low-level details and high-level semantics. After each convolution operation, batch normalization and ReLU activation function are used to make the feature distribution more stable and enhance the non-linear expressive ability of the network.

[0024] The mathematical expression for the connection is as follows:

[0025] ;

[0026] Where x is the input, F(x) is the nonlinear transformation of the convolutional layer, and y is the output.

[0027] The goal of residual convolutional blocks is to enhance deep feature learning while preventing information attenuation during deep mapping. After the residual convolutional blocks, the feature maps are then spatially compressed using max pooling layers, thereby preserving key structural information and reducing redundancy.

[0028] The feature fusion module first concatenates the feature vectors output by each branch along the feature dimension to form an overall feature representation containing multi-source time-frequency information. Then, it performs non-linear mapping on the concatenated features through a fully connected layer to learn the importance weights of each branch feature. Finally, it uses Softmax normalization to generate weighting coefficients. Each branch feature vector is multiplied by its corresponding weight and then summed to obtain the final comprehensive feature vector.

[0029] In the output module, the fused comprehensive feature vector is input into a two-layer fully connected network. Each layer performs weighted combination and nonlinear mapping on the input feature vector, progressively compressing the high-dimensional features into a low-dimensional representation related to the sound source coordinates. Batch normalization is added after each layer to ensure the consistency of image feature distribution across different batches. To prevent information from gradually decaying in deeper mapping layers, a residual connection is set after the second fully connected layer, superimposing the initial input image features with the processed features, and obtaining the final feature representation through the ReLU activation function, which is used to predict the coordinates of the acoustic emission source.

[0030] Furthermore, to enhance the model's ability to focus on spatial information, a CBAM attention mechanism is introduced after the pooling layer, including channel attention and spatial attention modules. First, the channel attention module is used to assign weights to different channels to highlight features related to the sound source. Then, the spatial attention module is used to generate the importance distribution of feature regions in the image, guiding the network to focus on key time-frequency regions in the image, thereby forming a more discriminative feature representation and improving the localization accuracy.

[0031] Furthermore, the formula for calculating channel attention is as follows:

[0032] ;

[0033] ;

[0034] The formula for calculating spatial attention is as follows:

[0035] ;

[0036] ;

[0037] Where F is the input feature map, and MaxPool(F) and AvgPool(F) represent global average pooling and max pooling, respectively. This represents a transformation of a 7×7 convolution kernel. Represents the Sigmoid function. This represents element-wise multiplication. This is the final output feature map.

[0038] Furthermore, during the training of the AE-DSCBRNet model, the mean squared error (MSE) is used as the loss function, the mean absolute error (MAE) as the evaluation metric, and RMSprop as the optimizer. The formula for calculating the mean squared error (MSE) is as follows:

[0039]

[0040]

[0041] Where n is the number of samples, It is the actual value of the i-th sample. It is the predicted value of the i-th sample;

[0042] The model performance was evaluated by comparing the difference between the predicted coordinates and the actual coordinates using the mean absolute error (MAE), root mean square error (RMSE), and relative distance error between the two coordinates on the test set.

[0043]

[0044]

[0045] Where n is the number of samples, , It is the actual value of the i-th sample. , It is the predicted value of the i-th sample. It represents the absolute difference between the actual value and the predicted value.

[0046] Beneficial technical effects of the present invention:

[0047] (1) This method uses a four-channel continuous wavelet image (CWT) as input. The DSConv module decomposes the standard convolution into channel-wise convolution and point-wise convolution, effectively reducing the number of network parameters and computational overhead. The attention module adaptively adjusts the feature weights through channel and spatial attention, highlighting key time-frequency features and suppressing redundant information. The four-branch features integrate multi-channel complementary information through adaptive weighted fusion, improving the signal representation capability. The residual connection improves gradient propagation and enhances the efficiency of deep feature learning, thereby achieving lightweight and high-precision acoustic emission source localization.

[0048] (2) This invention collects acoustic emission signals from the surface of a steel plate and performs continuous wavelet transform processing to convert the time-series signal into time-frequency domain image data. Simultaneously, the image is normalized to obtain a preprocessed time-frequency domain image dataset. Based on this, this invention constructs an improved convolutional neural network model AE-DSCBRNet, which includes a data input module, a feature extraction module, a feature fusion module, and an output module. Combining lightweight convolution, channel and spatial attention mechanisms, and adaptive multi-branch fusion design, the model can acquire more complex features from the acoustic emission signal, thereby achieving precise localization. Although this experiment was conducted on a steel plate, from the perspective of wave propagation theory, the time-frequency characteristics of acoustic emission signals have commonalities in different materials. Continuous wavelet transform (CWT) can be used for signal analysis of different materials, and this model can automatically extract key features; therefore, the applicability of this method to different materials has a theoretical basis. Future work will further verify the adaptability of this method in composite materials and heterogeneous structures. Attached Figure Description

[0049] Figure 1 This is a flowchart of the acoustic emission source localization method for steel plate structures described in this invention, illustrating the entire process from signal acquisition, preprocessing, model training to localization prediction;

[0050] Figure 2 The improved four-branch convolutional neural network structure diagram of the present invention includes four input branches, a depthwise separable convolutional feature extraction module, a residual convolutional block, a robust channel and spatial attention module, an adaptive multi-branch feature fusion module, and an output module.

[0051] Figure 3 This is a schematic diagram of the sensor arrangement and grid on the steel plate surface, showing the arrangement of four acoustic emission sensors (S1-S4) on the steel plate surface and the division of the sampling grid;

[0052] Figure 4 This is a schematic diagram of the experimental setup;

[0053] Figure 5 It is a comparison chart of positioning results, with error bars indicating the range of deviation between the predicted and actual values. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0055] like Figure 1 As shown, this embodiment provides a method for locating acoustic emission sources in steel plate structures based on lightweight convolution and attention mechanisms, including the following steps:

[0056] Step 1: Acoustic emission signals are collected at sampling grid points on the surface of the steel plate using an acoustic emission piezoelectric ceramic sensor;

[0057] Specifically, a PZT piezoelectric ceramic sensor (model PXR15, sensitivity >67dB, frequency bandwidth 100kHz-400kHz, resonant frequency 150kHz) was selected as the detection device for acoustic emission signals. The experimental material was a steel plate with dimensions of 650mm × 600mm × 3mm, the surface of which was divided into a monitoring network of 500mm × 500mm (spaced 25mm apart). Figure 3-4 Four acoustic emission sensors are placed at the four vertices of a square using a coupling agent, and numbered S1, S2, S3, and S4 respectively. A coordinate system is established with the center point as the origin; the coordinates of the sensors are then S1(25cm, 25cm), S2(-25cm, 25cm), S3(-25cm, -25cm), and S4(25cm, -25cm).

[0058] Specifically, a step load is applied to the grid points using a signal generator to simulate the signal of a real damage source. The same number of coordinate points are collected in each quadrant, and to avoid randomness in the localization results, each coordinate point is collected multiple times.

[0059] Step 2: Perform continuous wavelet transform processing on the acoustic emission signal to convert the acoustic emission time series signal into time-frequency domain image data, and normalize the image.

[0060] Specifically, the acquired acoustic emission signal is processed by continuous wavelet transform, using the following formula:

[0061]

[0062] in, It is a continuous complex function in both the time and frequency domains, called the mother wavelet. 'a' is a scaling factor that controls the scaling of the mother wavelet, and 'b' is a shift factor that controls the position of the wavelet on the time axis.

[0063] Specifically, the image is then normalized using the following formula:

[0064]

[0065] Specifically, This is the original image, where min and max represent the minimum and maximum pixel values, respectively.

[0066] Step 3: Integrate the image data and the corresponding acoustic emission source coordinates into a dataset, divide it into a training set and a test set, and build an improved convolutional network with multi-channel input for training;

[0067] Specifically, the continuous wavelet transform (CWT) images are integrated with their corresponding coordinate labels to form an acoustic emission signal localization dataset, containing 960 samples. The dataset is divided into a training set and a test set in an 8:2 ratio. During model training, the training and validation sets are input into the multi-layer convolutional network described in this invention. Using in-training set testing, the feature extraction capability and localization accuracy of the network are optimized by adjusting the training batch size, learning rate, and network structure parameters to obtain a training model capable of accurately predicting the coordinates of acoustic emission sources.

[0068] Specifically, this invention constructs an improved four-branch CNN network (AE-DSCBRNet) based on the TensorFlow deep learning framework to achieve high-precision localization of acoustic emission sources in steel plate structures. The acoustic emission signal is first subjected to continuous wavelet transform to obtain a time-frequency image, which is then input to four parallel branches. The feature extraction part of each branch consists of a depthwise separable convolutional (DSConv) block, a residual convolutional block, and a max-pooling layer. DSConv extracts local features from the input time-frequency image, the depthwise convolution captures local patterns in the spatial dimension, and pointwise convolution then linearly combines the features of each channel, thus significantly reducing the number of parameters and computational cost while preserving the main time-frequency features of the input image. Subsequently, the residual convolutional block further learns the extracted features. This module consists of two consecutive DSConv convolutional layers, and the input and output are added through skip connections to achieve the fusion of low-level details and high-level semantics. Each convolution operation is followed by batch normalization and a ReLU activation function to stabilize the feature distribution and enhance the network's nonlinear expressive power.

[0069] The goal of this module is to enhance deep feature learning while avoiding information attenuation during deep mapping. After the residual block, the feature map undergoes spatial compression through a max pooling layer to retain key structural information and reduce redundancy. Following the pooling layer, a robust CBAM attention mechanism is introduced, including channel attention and spatial attention modules. The channel attention module first assigns weights to different channels to highlight features related to the sound source. Then, the spatial attention module generates an importance distribution of feature regions in the image, guiding the network to focus on key time-frequency regions, thereby forming a more discriminative feature representation and improving localization accuracy.

[0070] The formula for calculating channel attention is as follows:

[0071] ;

[0072] ;

[0073] The formula for calculating spatial attention is as follows:

[0074] ;

[0075] ;

[0076] Where F is the input feature map, and MaxPool(F) and AvgPool(F) represent global average pooling and max pooling, respectively. This represents a transformation of a 7×7 convolution kernel. Represents the Sigmoid function. This represents element-wise multiplication. This is the final output feature map.

[0077] Specifically, the four branches perform convolution and pooling processes on time-frequency images from different sensors to obtain feature vectors representing the local patterns and frequency information of the input time-frequency images. These feature vectors are then integrated by an adaptive multi-branch fusion module. This module first concatenates the feature vectors output from each branch along their feature dimensions to form a holistic feature representation containing multi-source time-frequency information. Next, a fully connected layer performs a non-linear mapping on the concatenated features to learn the importance weights of each branch's features. Softmax normalization is then used to generate weighted coefficients. Each branch's feature vector is multiplied by its corresponding weight and summed to obtain the final integrated feature vector. This module enables dynamic adjustment of the contributions from different sensor inputs and weighted filtering and enhancement of multi-source time-frequency image features, thereby more effectively responding to key features and improving the accuracy and robustness of acoustic emission source localization.

[0078] In the output module, the fused comprehensive feature vector is input into a two-layer fully connected network. Each layer performs weighted combination and nonlinear mapping on the input feature vector, progressively compressing high-dimensional features into a low-dimensional representation related to the sound source coordinates. Batch normalization is added after each layer to ensure the consistency of image feature distribution across different batches. To prevent information from gradually decaying in deeper mappings, a residual connection is set after the second fully connected layer to superimpose the initial input image features with the processed features, and the final feature representation is obtained through the ReLU activation function, used to predict the coordinates of the acoustic emission source. During model training, the input image size is uniformly set to (224, 224, 3), the training batch size is set to 64, the maximum number of epochs is set to 900, and the dropout rate is set to 0.3. The optimizer is RMSprop, the loss function is mean squared error (MSE), and the mean absolute error (MAE) is used as the evaluation metric. Furthermore, an early stopping strategy and an adaptive learning rate adjustment algorithm are employed to improve the training efficiency and generalization ability of the model, and the best-performing model during training is saved.

[0079] A schematic diagram of the AE-DSCBRNet network structure is shown below. Figure 2 As shown.

[0080] The formulas for calculating the mean square error (MSE) and mean absolute error (MAE) are as follows:

[0081]

[0082]

[0083] Where n is the number of samples, It is the actual value of the i-th sample. It is the predicted value of the i-th sample.

[0084] Step 4: Input the test set into the trained network model to obtain the coordinates of the acoustic emission source location, thus realizing the localization of the acoustic emission source.

[0085] The performance of the model was evaluated by calculating the error between the predicted coordinates and the actual coordinates. The positioning accuracy of the model was quantified using the mean absolute error (MAE), root mean square error (RMSE), and relative error indices. The results are shown in Table 1.

[0086]

[0087]

[0088] Where n is the number of samples, , It is the actual value of the i-th sample. , It is the predicted value of the i-th sample. It represents the absolute difference between the actual value and the predicted value.

[0089] Predicted coordinates and actual coordinates are as follows Figure 5 As shown, the distance error between the predicted coordinates and the actual coordinates is less than 12mm, which is less than half of the 25mm grid spacing in this study. This method achieves a resolution of 25mm.

[0090] like Figure 5 As shown in the figure, the difference between the predicted coordinates and the actual coordinates is illustrated. The results demonstrate that this invention can achieve relatively accurate acoustic emission source localization.

[0091] Table 1

[0092]

[0093] To further verify the effectiveness of this invention, comparative experiments were conducted on the same dataset with the hyperbolic localization method based on time delay estimation (from "Acoustic Emission Source Localization Technology") and different CNN methods (2DCNN model from "Single-Sensor Acoustic Emission Source Localization in Plate-Like Structures Using Deep Learning"; AESLNET model from "Research On The Acoustic Emission Source Localization Methodology In Composite Materials Based On Artificial Intelligence"). The results are shown in Table 2.

[0094] Table 2

[0095]

[0096] Experimental results show that the depthwise separable convolution (DSConv) blocks, residual convolution blocks, and robust channel and spatial attention mechanisms employed in this invention can effectively improve the localization accuracy of acoustic emission sources. Through the adaptive multi-branch fusion module, the model can fully integrate complementary features from multiple sensor inputs, improving the robustness of localization. Since this method relies on data-driven feature learning rather than physical modeling of specific materials, it can be extended to acoustic emission source localization tasks for steel structures, composite materials, and other engineering structures after appropriate data transfer and fine-tuning, while maintaining a lightweight design and low computational cost.

Claims

1. A method for locating acoustic emission sources based on lightweight convolution and attention mechanisms, characterized in that: Includes the following steps: Step 1: Acoustic emission signals are collected at sampling grid points on the surface of the steel plate using an acoustic emission piezoelectric ceramic sensor; Step 2: Perform continuous wavelet transform processing on the acoustic emission signal to convert the acoustic emission time series signal into time-frequency domain image data, and normalize the image. Step 3: Integrate the image data and the corresponding acoustic emission source coordinates into a dataset, divide it into a training set and a test set, build an improved four-branch CNN network AE-DSCBRNet model and train it; Step 4: Input the test set into the trained network model to obtain the coordinates of the acoustic emission source location, thus realizing the localization of the acoustic emission source; In step 3, the AE-DSCBRNet model includes a data input module, a feature extraction module, a feature fusion module, and an output module; The data input module contains four input branches, each receiving a continuous wavelet image data; The feature extraction module consists of two consecutive depthwise separable convolutional (DSConv) blocks, a residual convolutional block, and a max-pooling layer. The DSConv block extracts local features from the input time-frequency image. Depthwise convolution captures local patterns in the spatial dimension, and pointwise convolution then linearly combines the features from each channel, significantly reducing the number of parameters and computation while preserving the main time-frequency features of the input image. The residual convolutional block further learns the extracted features. It consists of two consecutive DSConv convolutional layers, with skip connections adding the input and output to fuse low-level details with high-level semantics. Each convolution operation is followed by batch normalization and ReLU activation to stabilize the feature distribution and enhance the network's non-linear expressive power. After the residual convolutional block, the feature map is spatially compressed using a max-pooling layer, preserving key structural information and reducing redundancy. The feature fusion module first concatenates the feature vectors output by each branch along the feature dimension to form an overall feature representation containing multi-source time-frequency information. Then, it performs non-linear mapping on the concatenated features through a fully connected layer to learn the importance weights of each branch feature. Finally, it uses Softmax normalization to generate weighting coefficients. Each branch feature vector is multiplied by its corresponding weight and then summed to obtain the final comprehensive feature vector. In the output module, the fused integrated feature vector is input into a two-layer fully connected network. Each layer performs weighted combination and nonlinear mapping on the input feature vector, gradually compressing the high-dimensional features into a low-dimensional representation related to the sound source coordinates. Batch normalization is added after each layer to ensure the consistency of image feature distribution across different batches. To prevent information from gradually decaying in deep mapping, a residual connection is set after the second fully connected layer to superimpose the initial input image features with the processed features and obtain the final feature representation through the ReLU activation function, which is used to predict the coordinates of the acoustic emission source.

2. The acoustic emission source localization method based on lightweight convolution and attention mechanism according to claim 1, characterized in that: In step 1, a sampling network is divided on the surface of the steel plate, and acoustic emission sensors are arranged at the four vertices of the sampling network to form a sensor network. A signal generator is used at the grid points of the sampling to simulate an acoustic emission source and collect acoustic emission signals.

3. The acoustic emission source localization method based on lightweight convolution and attention mechanism according to claim 1, characterized in that: In step 2, the acoustic emission signal is subjected to continuous wavelet transform processing, wherein the wavelet function selected is the Morse wavelet, and the formula for continuous wavelet transform processing is: ； in, It is a continuous complex function in the time and frequency domains, called the mother wavelet; a is the scaling factor, which controls the scaling of the mother wavelet, and b is the shift factor, which controls the position of the wavelet on the time axis.

4. The acoustic emission source localization method based on lightweight convolution and attention mechanism according to claim 1, characterized in that: In step 2, the generated continuous wavelet transform (CWT) time-frequency image was normalized to ensure that the numerical range of all time-frequency features was consistent, thus avoiding the impact of uneven data distribution on network training. The normalization formula is as follows: ； in, This is the original image, where min and max represent the minimum and maximum pixel values, respectively.

5. The acoustic emission source localization method based on lightweight convolution and attention mechanism according to claim 1, characterized in that: In step 3, in order to enhance the model's ability to focus on spatial information, a CBAM attention mechanism is introduced after the pooling layer, including channel attention and spatial attention modules. First, the channel attention module is used to assign weights to different channels to highlight features related to the sound source. Then, the spatial attention module is used to generate the importance distribution of feature regions in the image, guiding the network to focus on key time-frequency regions in the image, thereby forming a more discriminative feature representation and improving the localization accuracy. The formula for calculating channel attention is as follows: ; ； The formula for calculating spatial attention is as follows: ; ； Where F is the input feature map, and MaxPool(F) and AvgPool(F) represent global average pooling and max pooling, respectively. This represents a transformation of a 7×7 convolution kernel. Represents the Sigmoid function. This represents element-wise multiplication. This is the final output feature map.

6. The acoustic emission source localization method based on lightweight convolution and attention mechanism according to claim 1, characterized in that: In step 3, during the training of the AE-DSCBRNet model, the mean squared error (MSE) is used as the loss function, the mean absolute error (MAE) as the evaluation metric, and RMSprop as the optimizer. The formulas for calculating MSE and MAE are as follows: ； ； Where n is the number of samples, It is the actual value of the i-th sample. It is the predicted value of the i-th sample. It represents the absolute difference between the actual value and the predicted value.

7. The acoustic emission source localization method based on lightweight convolution and attention mechanism according to claim 6, characterized in that: In step 4, the difference between the predicted coordinates and the actual coordinates is compared, and the model performance is evaluated using the mean absolute error (MAE), root mean square error (RMSE), and relative distance error between the two coordinates on the test set. ； ； Where n is the number of samples, , It is the actual value of the i-th sample. , It is the predicted value of the i-th sample.

Images