Gas turbine acoustic fault diagnosis method based on lightweight multi-branch multi-dimensional attention convolutional network

Abstract

The invention discloses a gas turbine acoustic fault diagnosis method based on a lightweight multi-branch multi-dimensional attention convolutional network. The method comprises the following steps: collecting acoustic signals to be classified; the CWT module converts the acoustic signal into a two-dimensional time-frequency diagram through wavelet transform; the multi-branch parallel processing module obtains a multi-scale processing feature map through multi-branch parallel processing; the multi-dimensional dynamic convolution module extracts a multi-dimensional attention weight and processes a feature map through a convolution sub-module, a feature extraction sub-module and a splicing sub-module, and performs feature fusion and iterative enhancement to obtain a feature map after feature enhancement; and the KAN classifier obtains a fault diagnosis result through classification after fusion of the main path and the residual path. According to the method, the feature extraction capability is enhanced, the fault diagnosis accuracy is improved, the parameter quantity and the calculation complexity are reduced, and efficient and lightweight fault diagnosis is realized.

Description

Technical Field

[0001] This invention belongs to the field of sound signal technology, and in particular relates to a method for diagnosing acoustic faults in gas turbines based on a lightweight multi-branch multidimensional attention convolutional network. Background Technology

[0002] Sound is the most natural carrier of information, serving as a crucial tool for communication between humans and machines. The information conveyed by sound can be extracted and analyzed using signal processing methods to enhance environmental perception and understanding. Mechanical fault diagnosis involves collecting, analyzing, and judging the operating status signals of mechanical equipment to identify any abnormalities, determine the location and type of fault, and assess the severity of the fault. Mechanical fault diagnosis can detect early-stage faults, prevent major damage or downtime, and provide a basis for maintenance decisions, demonstrating promising real-world applications. Deep learning, due to its superior performance in processing large amounts of features, is widely used in acoustic fault diagnosis classification tasks. For example, Convolutional Neural Networks (CNNs), Recurrent Convolutional Neural Networks, and Deeply Separable Convolutional Neural Networks have all been widely applied in the field of acoustic fault diagnosis classification.

[0003] To date, scholars both domestically and internationally have proposed various acoustic fault diagnosis and classification methods. Hou et al., for high-speed train wheel fault diagnosis, established a fingerprint feature recognition method combining dynamic thresholds and acoustic emission signals, achieving high fault diagnosis accuracy. Bai et al., addressing the class imbalance problem in gas turbine fault diagnosis, proposed a novel deep neural network class balancing method based on a combination of focus loss and random oversampling. To address the problem of limited labeled data, Liu et al. proposed a passive robust domain adaptive method based on pseudo-label uncertainty estimation. Shuai et al. designed a processing flow combining spatial and time-frequency domains, which significantly enhanced fault features in acoustic signals and improved classification and detection accuracy. Li et al. designed the lightweight network MS-Ghost Net V3, employing a reparameterization training strategy during convolutional network training. This approach improved model performance while reducing the number of parameters in the model. While feature enhancement methods can significantly improve fault diagnosis accuracy, they often increase model complexity. This makes them difficult to deploy in industrial environments with limited computing and storage resources. Conversely, lightweight strategies effectively reduce computational burden, but their simple network structure may lead to insufficient fault feature extraction, affecting diagnostic accuracy. Achieving an effective balance between feature enhancement and model lightweighting remains a major challenge in current fault diagnosis research. To address this issue, a gas turbine acoustic fault diagnosis method based on a lightweight multi-branch multidimensional attention convolutional network is proposed. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a gas turbine acoustic fault diagnosis method based on a lightweight multi-branch multidimensional attention convolutional network, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, this invention provides a gas turbine acoustic fault diagnosis method based on a lightweight multi-branch multidimensional attention convolutional network, comprising: Acoustic signals to be classified are collected; The acoustic signal to be classified is input into the acoustic fault diagnosis classification model to obtain the classification result; wherein, the acoustic fault diagnosis classification model is constructed based on the CWT module, the multi-branch parallel processing module, the multi-dimensional dynamic convolution module and the KAN classifier; The acoustic signal to be classified is input into the CWT module and subjected to wavelet transform to obtain a two-dimensional time-frequency diagram. The two-dimensional time-frequency graph is input into a multi-branch parallel processing module for multi-scale parallel processing to obtain a multi-scale processed feature map. The multi-scale processed feature map is input into a multi-dimensional dynamic convolution module for feature enhancement and extraction to obtain a feature map with enhanced features. The enhanced feature map is input into the KAN classifier for classification processing to obtain the fault diagnosis classification result.

[0006] Optionally, the process of inputting the acoustic signal to be classified into the continuous wavelet transform (CWT) module for wavelet transform to obtain a two-dimensional time-frequency diagram includes: Select a mother wavelet function; construct wavelet basis functions based on the mother wavelet function by changing the scaling and translation parameters; perform convolution operation between the acoustic signal to be classified and the wavelet basis functions to obtain the wavelet coefficient matrix; convert the wavelet coefficient matrix into a two-dimensional time-frequency graph.

[0007] Optionally, the multi-branch parallel processing module includes convolutional layers, depthwise convolutional layers, vertical stripe kernel convolutional layers, horizontal stripe kernel convolutional layers, skip layers, feature concatenation layers, normalization layers, fully connected layers, and result concatenation layers; the process of inputting the two-dimensional time-frequency map into the multi-branch parallel processing module for multi-scale parallel processing to obtain multi-scale processed feature maps includes: The two-dimensional time-frequency map is channel-expanded through the convolutional layer to obtain the input feature map; The input feature map is evenly divided into four sub-feature maps according to the number of channels; The four sub-feature maps are respectively input into the depth convolutional layer, the vertical strip kernel convolutional layer, the horizontal strip kernel convolutional layer, and the skip layer for processing to obtain the corresponding output results; The four output results are concatenated along the channel dimension through the feature concatenation layer to obtain a multi-branch concatenated feature map; The multi-branch spliced ​​feature map is normalized using the normalization layer. The number of channels in the normalized feature map is recovered through the fully connected layer to obtain the feature map after the fully connected layer; The feature map after the fully connected layer is stitched together with the two-dimensional time-frequency map in the channel dimension through the result stitching layer to obtain the multi-scale processed feature map.

[0008] Optionally, the multidimensional dynamic convolution module includes a convolution sub-module, a feature extraction sub-module, and a concatenation sub-module connected in sequence; the process of inputting the multi-scale processed feature map into the multidimensional dynamic convolution module for feature enhancement and extraction to obtain the feature-enhanced feature map includes: The multi-scale feature map is processed by the convolutional submodule to expand the number of channels and obtain an expanded feature map; the expanded feature map is then processed by an activation function to obtain an activated expanded feature map. The feature extraction submodule extracts the multidimensional attention of the activated extended feature map, and processes the activated extended feature map using the extracted multidimensional attention to obtain the output feature map; the attention feature map is then reduced in dimensionality through convolution to obtain the reduced feature map. The multi-scale processed feature map is compared with the output feature map. If the sizes are the same, feature fusion is performed through the splicing submodule and the result is used as the output. If the sizes are different, the dimensionality-reduced output feature map is used as the output. The convolution submodule, feature extraction submodule, and splicing submodule are repeatedly processed a preset number of times to iteratively enhance the output result and obtain the enhanced feature map.

[0009] Optionally, the process of processing the multi-scale feature map and extracting multi-dimensional attention weights through the feature extraction submodule includes: The activated extended feature map is subjected to global pooling; the pooled features are then subjected to dimensionality reduction; the dimensionality-reduced features are processed through four independent attention branches to obtain spatial attention, input channel attention, output channel attention, and convolutional kernel attention, respectively.

[0010] Optionally, the KAN classifier includes a main path module and a residual path module; the enhanced feature map is input into the KAN classifier for classification processing to obtain the fault diagnosis classification result, including: The enhanced feature maps are input into the main path module and the residual path module respectively to obtain the main path output and the residual path output; the main path output and the residual path output are fused; the fused result is input into the fully connected layer for classification to obtain the fault diagnosis classification result.

[0011] Optionally, in the main path module, a B-spline transformation method is used to perform a nonlinear transformation to obtain the main path output. The B-spline transformation method uses a piecewise polynomial function to smoothly connect and fit the nonlinear features at the nodes, and achieves adaptive changes in weights by dynamically adjusting the parameters of the spline function.

[0012] Optionally, in the residual path module, the input features are processed using the Sigmoid activation function, and the residual path output is obtained by calculating through a linear layer with weights and biases.

[0013] Optionally, the main path output and the residual path output can be merged by weighted summation.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: The proposed network framework of this invention includes a CWT module, a multi-branch parallel processing module, a multi-dimensional dynamic convolution module, and a KAN classifier. Specifically, the CWT module converts acoustic signals into high-resolution time-frequency maps, providing richer input for fault diagnosis; the multi-branch parallel processing module and the KAN classifier reduce the number of parameters, improving the efficiency of fault diagnosis; and the multi-dimensional dynamic convolution module captures multi-dimensional features, significantly enhancing the network's feature extraction capabilities and improving the classification accuracy of moving mechanical parts under fault conditions. This invention successfully balances accuracy and efficiency, achieving a unified high accuracy, fast inference speed, and low resource consumption, making it highly valuable for industrial deployment. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the fault diagnosis network architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the multi-branch parallel processing module according to an embodiment of the present invention; Figure 3 This is a schematic diagram of multi-dimensional attention in the multi-dimensional dynamic convolution module of this invention; wherein, (a) is a schematic diagram of spatial attention, (b) is a schematic diagram of input channel attention, (c) is a schematic diagram of output channel attention, and (d) is a schematic diagram of convolution kernel attention; Figure 4This is a schematic diagram of the KAN classifier according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the confusion matrix according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 like Figure 1 As shown, this embodiment provides a gas turbine acoustic fault diagnosis method based on a lightweight multi-branch multidimensional attention convolutional network, including: Acoustic signals to be classified are collected; The acoustic signal to be classified is input into the acoustic fault diagnosis classification model to obtain the classification result.

[0019] The acoustic fault diagnosis classification model processes the acoustic signal to be classified as follows: Step 1: Input the acoustic signal to be classified into the CWT module to obtain a two-dimensional time-frequency graph. Step 2: Input the two-dimensional time-frequency graph into a multi-branch parallel processing module to obtain a feature map X. Step 3: Input the feature map X into a multi-dimensional dynamic convolution module to obtain a feature map Z. Step 4: Input the feature map Z into a KAN classifier to obtain the classification result.

[0020] Specifically: Step 1: Input the acoustic signal to be classified into the CWT module to obtain a two-dimensional time-frequency diagram.

[0021] Step 1 specifically involves: Step 1.1: Obtain the raw audio signals required for the fault diagnosis task.

[0022] Step 1.2: In the CWT module, select a suitable transformation method to convert one-dimensional sound signals into two-dimensional images. Since continuous wavelet transform is suitable for processing non-stationary signals, and the sound signals generated in the fault are non-stationary signals, continuous wavelet transform is selected as the transformation method to convert one-dimensional sound signals into two-dimensional images.

[0023] Step 1.3: Configure the detailed parameters for the continuous wavelet transform in the CWT module, selecting the Morlet wavelet as the mother wavelet function used in the continuous wavelet transform. The formula for the mother wavelet function is: (1) in, The scale parameter representing the stretching or compressing of the wavelet, and This represents the translation parameter, which is changed by... and The value of continuous wavelet transform can be used to analyze signals at different scales and locations, thereby extracting features at different scales.

[0024] Step 1.4: Process the original acoustic signal using the continuous wavelet transform in the CWT module. The formula for the continuous wavelet function is: (2) in, It is the input signal. It is the mother wavelet function after scaling and translation. Indicates time, The scale parameter representing the stretching or compressing of the wavelet, and This represents the translation parameter, which controls the position of the wavelet in the signal.

[0025] Step 1.5: Set the scale range. Setting an appropriate scale range is a crucial step in ensuring time-frequency resolution and computational efficiency. Different scale ranges determine the frequency range covered by CWT. Select an appropriate scale range based on the input audio signal.

[0026] Step 1.6: Perform continuous wavelet transform processing on the input sound signal, that is, for each preset scale parameter, perform convolution operation on the input sound signal and the wavelet basis function that has been translated and scaled to obtain the transform results at different scales.

[0027] Step 1.7: Based on the convolution results at each scale and time position, calculate the corresponding two-dimensional wavelet coefficient matrix. The two-dimensional wavelet coefficient matrix is ​​used to characterize the energy distribution characteristics of the sound signal in the time domain and frequency domain.

[0028] Step 1.8: Visualize the two-dimensional wavelet coefficient matrix to generate a corresponding two-dimensional time-frequency graph, so as to intuitively show the energy changes of the sound signal at different times and frequencies.

[0029] Step 2: Input the two-dimensional time-frequency graph into the multi-branch parallel processing module to obtain the multi-scale processed feature map X. The structure of the multi-branch parallel processing module is as follows: Figure 2 As shown.

[0030] Step 2 specifically involves: Step 2.1: Input the two-dimensional time-frequency image into the convolutional layer. After convolution calculation, expand the number of channels of the two-dimensional time-frequency image to obtain the input feature map. The size of the input feature map is C×W×H; where C refers to the number of channels of the feature map, W refers to the width of the feature map, and H refers to the height of the feature map.

[0031] Step 2.2: Divide the input feature map into four sub-feature maps evenly according to the channel dimension. The size of each sub-feature map is C / 4×W×H. Input the four sub-feature maps into the depth convolutional layer, the vertical strip kernel convolutional layer, the horizontal strip kernel convolutional layer, and the skip layer respectively to obtain the corresponding output results.

[0032] Step 2.3: Input the output results of the depth convolutional layer, the vertical strip kernel convolutional layer, the horizontal strip kernel convolutional layer and the skip layer into the feature splicing layer, and splice the output results in the channel dimension to obtain a multi-branch spliced ​​feature map. The size of the spliced ​​feature map is C×W×H.

[0033] Step 2.4: Input the multi-branch concatenated feature map into the normalization layer. The normalization layer uses BatchNormalization (BN) to process the input. The formula for BN is as follows: (3) in, This is the result after normalization. It is the mean. It is variance. A very small constant is used to prevent the denominator from being zero. This yields the normalized feature map C×W×H. Step 2.5: Input the normalized feature map into the fully connected layer. The fully connected layer restores the number of channels of the normalized feature map to the number of channels of the two-dimensional time-frequency map, thus obtaining the feature map after the fully connected layer. Step 2.6: The feature map after the fully connected layer is combined with the two-dimensional time-frequency map input result in a concatenation layer to complete the concatenation in the channel dimension, thereby obtaining a multi-scale processed feature map X.

[0034] Step 3: Input the multi-scale processed feature map X into the multi-dimensional dynamic convolution module for feature enhancement to obtain the enhanced feature map Z.

[0035] Step 3 specifically involves: Step 3.1: Use a 1×1 convolutional layer to expand the dimensions of the multi-scale processing feature map X to obtain an expanded feature map.

[0036] Step 3.2: Process the expanded feature map using Batch Normalization (BN) and Hardswish activation functions to obtain the activated expanded feature map. The Hardswish activation function is as follows: (4) in, Indicates time, This involves restricting the input to the range of 0 to 6. Hardswish avoids complex exponential operations, thereby reducing the computational cost that contributes to gradient propagation and improving model performance.

[0037] Step 3.3: Focus on the feature distribution of the activated feature map through the feature extraction submodule to obtain the following multi-dimensional attention: spatial attention, input channel attention, output channel attention, and convolutional kernel attention. The process of processing the multi-scale feature map and extracting multi-dimensional attention weights through the feature extraction submodule includes: performing global pooling on the activated expanded feature map; performing dimensionality reduction on the pooled features; and processing the dimensionality-reduced features through four independent attention branches to obtain spatial attention, input channel attention, output channel attention, and convolutional kernel attention, respectively. The multi-dimensional attention in the multi-dimensional dynamic convolution module is as follows: Figure 3 As shown, where, Figure 3 (a) in the diagram is a schematic of spatial attention. Figure 3 (b) in the diagram is a schematic diagram of the input channel attention. Figure 3 (c) in the diagram is a schematic of the output channel attention. Figure 3 (d) in the diagram is a schematic of the convolution kernel attention.

[0038] Step 3.4: Perform feature enhancement processing on the activated feature map using the four-dimensional attention. The processing details are as follows: (5) in, Indicates the features after activation. This represents the output feature map. Represents the convolution kernel Note the scalar, This represents attention along the spatial dimension. This represents attention along the input channel dimension. Attention along the output channel dimension, This represents attention along the spatial dimension of the convolution kernel.

[0039] Step 3.5: Process the output feature map using a 1×1 convolutional layer. Dimensionality reduction is performed to obtain the dimensionality-reduced feature map.

[0040] Step 3.6: Compare the multi-scale processed feature map X with the dimensionality-reduced feature map. If the two feature maps are the same size, input them into the concatenation submodule to concatenate the features along the channel dimension to obtain a new output result. If the two feature maps are different sizes, use the dimensionality-reduced feature map as the output result.

[0041] Step 3.7: Repeat the processing of the feature extraction submodule, convolution submodule, and concatenation submodule on the output result 11 times to achieve more complete feature extraction and obtain the feature map Z after feature enhancement extracted by the multidimensional dynamic convolution module.

[0042] Step 4: Input the feature map Z into the KAN classifier to obtain the classification result.

[0043] Step 4 specifically involves: Step 4.1: Input the feature map Z into the main path and branch path of the KAN classifier respectively.

[0044] Step 4.2: For the KAN classifier branch path, process the feature map Z using the Sigmoid activation function to obtain the result. .

[0045] Step 4.3: For the branch paths of the KAN classifier, use weights. and bias For the above Processing is performed to obtain the branch path results of the KAN classifier. The formula is as follows: (6) Step 4.4: For the main path of the KAN classifier, the main path processes the feature map Z using a spline function, which achieves a dynamic nonlinear mapping. The spline function is a piecewise polynomial function used for function approximation. Each polynomial piece is defined within a specified node and smoothly connected at the nodes to ensure the continuity of the derivative. The spline function can be mathematically represented as: (7) in, This represents the mapping from the spline function to the input, while It is a piecewise polynomial, defined on the interval of nodes. Internally, for the feature map Z, after spline transformation processing, the following is obtained: .

[0046] Step 4.5: In the KAN main path, the weights are dynamic and parameterized by a spline function. For the feature map Z, each weight... It can be represented as: (8) in, The feature map Z is used as the input variable and A spline function with parameters. This indicates the coefficients and the position of the junction for each polynomial segment.

[0047] Step 4.6: Combine the splined input features with dynamic weights through weighted summation. The weighting method can be expressed as: (9) in This represents the parameterized weights of the spline. Then, the weighted sum is passed through a linear layer for dimensionality transformation. The specific calculation can be expressed as follows: (10) in These are the corresponding weights. That is the corresponding bias.

[0048] Step 4.7: Add the main path result and the branch path result of the KAN classifier together, using the following formula: (11) Obtain the output of the KAN classifier. This is the classification result. This combination helps the model capture both global linear relationships and local nonlinear relationships simultaneously, thereby enhancing the model's representational power and generalization performance.

[0049] The detailed parameter settings for the above fault diagnosis classification model are shown in Table 1.

[0050] Table 1

[0051] To evaluate the effectiveness of the proposed method, a dataset was used for verification. The dataset used in this embodiment was derived from acoustic signal data collected from a gas turbine under different operating conditions. Acoustic emission signal data were collected during normal and fault operation of the gas turbine. The acoustic emission of the gas turbine was detected using a Micro-80D acoustic emission sensor. The sensor was connected to a waveguide welded to the turbine exhaust. The sensor was specifically positioned on the waveguide to avoid recording bearing-induced acoustic emissions. In all tests, the raw acoustic emission data acquisition configuration remained consistent under all conditions. Under each operating condition, 20 sets of raw acoustic emission data were recorded, each lasting 0.03 seconds, at a sampling rate of 5MHz. Four different types of operating conditions were tested, including two normal conditions and two fault conditions. To ensure consistency, each test was repeated three times. After removing interfering invalid acoustic signal data, a total of 2578 valid acoustic emission signals were recorded. The available acoustic signals were converted into two-dimensional time-frequency plots, divided into training and test sets. Table 2 shows the detailed settings of the two-dimensional time-frequency plots obtained under different operating conditions.

[0052] Table 2

[0053] The final classification result of the fault diagnosis is represented by a confusion matrix. The confusion matrix of the proposed fault diagnosis method is as follows: Figure 5 As shown, the numbers on the diagonal of the confusion matrix from the top left to the bottom right represent the correct classification accuracy for each category. High classification accuracy was achieved for all categories, demonstrating the effectiveness of the proposed fault diagnosis method.

[0054] To evaluate the practical performance of the proposed fault diagnosis method, four evaluation factors were used: accuracy, precision, recall, and F1 score. These evaluation metrics rely on four basic components: true positives, false positives, false negatives, and true negatives. Specifically, true positives (TP) represent the number of samples correctly classified as belonging to the fault category by the model; false positives (FP) represent the number of samples misclassified as belonging to the fault category; false negatives (FN) correspond to the number of samples that the model failed to identify as belonging to the fault category; and true negatives (TN) represent the number of samples correctly classified as not belonging to the fault category.

[0055] (12) (13) (14) (15) To verify the fault diagnosis performance of the proposed method, comparative experiments were conducted with other fault diagnosis methods based on MobileNetV2, ResNet50, ConvNeXt-Tiny, and DenseNet. The experimental results are shown in Table 3. As can be seen from Table 3, the proposed method achieved an accuracy of 87%. Compared with other fault diagnosis methods, its diagnostic efficiency is improved by approximately 5%. Compared with other fault diagnosis methods, the proposed method requires only 1.88M parameters, which is less than other fault diagnosis methods. Experiments demonstrate that this method significantly achieves lightweight design while maintaining classification capabilities. This method has high computational efficiency and low resource consumption, making it more suitable for resource-constrained scenarios. These comparisons demonstrate the effectiveness and superiority of the proposed method.

[0056] Table 3

[0057] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A gas turbine acoustic fault diagnosis method based on a lightweight multi-branch multi-dimensional attention convolutional network, characterized in that, The method comprises the following steps: Collecting an acoustic signal to be classified; Inputting the acoustic signal to be classified into an acoustic fault diagnosis classification model to obtain a classification result; wherein the acoustic fault diagnosis classification model is constructed based on a CWT module, a multi-branch parallel processing module, a multi-dimensional dynamic convolution module, and a KAN classifier; Inputting the acoustic signal to be classified into the CWT module to perform wavelet transform and obtain a two-dimensional time-frequency graph; Inputting the two-dimensional time-frequency graph into the multi-branch parallel processing module to perform multi-scale parallel processing and obtain a multi-scale processing feature map; Inputting the multi-scale processing feature map into the multi-dimensional dynamic convolution module to perform feature enhancement extraction and obtain a feature map after feature enhancement; Inputting the feature map after feature enhancement into the KAN classifier to perform classification processing and obtain a fault diagnosis classification result.

2. The method of claim 1, wherein, The process of inputting the acoustic signal to be classified into the continuous wavelet transform CWT module to perform wavelet transform and obtain a two-dimensional time-frequency graph comprises: Selecting a mother wavelet function; constructing a wavelet basis function by changing scale parameters and translation parameters based on the mother wavelet function; performing convolution operation on the acoustic signal to be classified and the wavelet basis function to obtain a wavelet coefficient matrix; and converting the wavelet coefficient matrix into a two-dimensional time-frequency graph.

3. The method of claim 1, wherein, The multi-branch parallel processing module comprises a convolution layer, a deep convolution layer, a vertical strip kernel convolution layer, a horizontal strip kernel convolution layer, a jump layer, a feature splicing layer, a normalization layer, a full connection layer, and a result splicing layer; the process of inputting the two-dimensional time-frequency graph into the multi-branch parallel processing module to perform multi-scale parallel processing and obtain a multi-scale processing feature map comprises: Expanding the channel of the two-dimensional time-frequency graph through the convolution layer to obtain an input feature map; Dividing the input feature map into four sub-feature maps according to the number of channels; Inputting the four sub-feature maps into the deep convolution layer, the vertical strip kernel convolution layer, the horizontal strip kernel convolution layer, and the jump layer respectively to obtain corresponding output results; Splicing the four output results in the channel dimension through the feature splicing layer to obtain a multi-branch splicing feature map; Performing normalization processing on the multi-branch splicing feature map through the normalization layer; Restoring the channel number of the normalized feature map through the full connection layer to obtain a feature map after the full connection layer; Splicing the feature map after the full connection layer and the two-dimensional time-frequency graph in the channel dimension through the result splicing layer to obtain the multi-scale processing feature map.

4. The method of claim 1, wherein, The multi-dimensional dynamic convolution module comprises a convolution submodule, a feature extraction submodule, and a splicing submodule connected in sequence; the process of inputting the multi-scale processing feature map into the multi-dimensional dynamic convolution module to perform feature enhancement extraction and obtain a feature map after feature enhancement comprises: Processing the multi-scale processing feature map through the convolution submodule to expand the channel number of the feature map and obtain an expanded feature map; processing the expanded feature map through an activation function to obtain an activated expanded feature map; The feature extraction submodule extracts the multi-dimensional attention of the activated extended feature map, processes the activated extended feature map using the extracted multi-dimensional attention, and obtains an output feature map; the attention feature map is reduced in dimension by convolution, and a reduced feature map is obtained; The multi-scale processing feature map and the output feature map are compared, and if they are of the same size, the feature fusion is performed by the splicing submodule as an output result; if they are not of the same size, the reduced output feature map is taken as an output result; The processing of the convolution submodule, the feature extraction submodule and the splicing submodule is repeated for a preset number of times to perform iterative enhancement on the output result, and a feature-enhanced feature map is obtained.

5. The method of claim 4, wherein, The process of extracting multi-dimensional attention weight by processing the multi-scale processing feature map by the feature extraction submodule includes: The activated extended feature map is globally pooled, and the pooled feature is reduced in dimension; the reduced feature is processed by four independent attention branches to obtain spatial attention, input channel attention, output channel attention and convolution kernel attention, respectively.

6. The method of claim 1, wherein, The KAN classifier includes a main path module and a residual path module; the feature-enhanced feature map is input into the KAN classifier for classification processing to obtain a fault diagnosis classification result, which includes: The feature-enhanced feature map is input into the main path module and the residual path module, respectively, to obtain a main path output and a residual path output; the main path output and the residual path output are fused; the fused result is input into a fully connected layer for classification to obtain a fault diagnosis classification result.

7. The method of claim 6, wherein, In the main path module, a B-spline transformation method is used for nonlinear transformation to obtain a main path output; the B-spline transformation method uses a piecewise polynomial function to smoothly connect and fit nonlinear features at nodes, and realizes adaptive changes of weights by dynamically adjusting parameters of the spline function.

8. The method of claim 6, wherein, In the residual path module, an input feature is processed using a Sigmoid activation function, and then a residual path output is calculated through a linear layer with weights and bias.

9. The method of claim 6, wherein, The main path output and the residual path output are fused by weighted summation.

Images