Mechanical fault diagnosis method and device based on multi-branch fusion

Abstract

The invention relates to a mechanical fault diagnosis method and device based on multi-branch fusion. The method comprises the following steps: preprocessing an original vibration signal to obtain a fault feature map; processing the fault features through an AMF branch by using a signal denoising branch and a signal ViT branch which are executed in parallel; wherein the signal denoising branch is used for extracting fault-related frequency domain features in a fault feature map to obtain local map structure features, and the signal ViT branch is used for capturing a global long-range dependency relationship in the fault feature map to obtain global features; performing alignment on the local graph structure features and the global features through AMF branches, and performing adaptive fusion based on multi-scale grouping convolution and an attention mechanism to obtain fusion features; the fusion features are processed through global average pooling and a full connection layer, and a fault classification result is output; according to the technical scheme, the diagnosis accuracy and the robustness under different noise interferences are improved, the parameter quantity is reduced to optimize the calculation efficiency, and the model reasoning speed is increased.

Description

Technical Field

[0001] This application relates to the field of machine equipment fault diagnosis technology, and in particular to a mechanical fault diagnosis method and device based on multi-branch fusion. Background Technology

[0002] As a core component of rotating machinery, the reliable operation of rolling bearings is crucial for industrial applications. Traditional condition monitoring methods mainly rely on extracting discriminative features reflecting the health status of bearings from vibration signals. However, under complex operating conditions, the fault information of rolling bearings is often masked by noise and other interference, making it difficult for existing methods to effectively extract key features.

[0003] In recent years, deep learning technology, especially Convolutional Neural Networks (CNNs), has become a powerful tool for mechanical fault diagnosis, enabling the automatic learning of mechanical signal features. For example, existing research has improved feature extraction capabilities by introducing symmetric point graph representations, improved ensemble empirical mode decomposition, or multi-scale convolutional structures. Although CNN-based methods have achieved certain results, they typically focus more on extracting local features and lack the ability to model global long-range dependencies, thus limiting the performance of constructing comprehensive fault representations. Furthermore, mechanical systems often operate in harsh environments, and vibration signals are easily contaminated by noise. Existing methods usually separate denoising from the diagnostic process, which may lead to the filtering out of some effective information, thereby affecting diagnostic accuracy. Summary of the Invention

[0004] To address one of the aforementioned shortcomings, this application provides a mechanical fault diagnosis method and apparatus based on multi-branch fusion, thereby improving fault diagnosis accuracy.

[0005] A mechanical fault diagnosis method based on multi-branch fusion includes: The original vibration signals generated during the operation of mechanical equipment are preprocessed to obtain fault feature maps; The fault feature map is processed using a parallel-executed signal denoising branch and a signal ViT branch; wherein, the signal denoising branch is used to extract the frequency domain features related to the fault in the fault feature map to obtain local graph structure features, and the signal ViT branch is used to capture the global long-range dependencies in the fault feature map to obtain global features; The local graph structure features and global features are aligned by AMF branches and adaptively fused based on multi-scale grouped convolution and attention mechanisms to obtain fused features; The fused features are processed through global average pooling and a fully connected layer to output fault classification results.

[0006] In some embodiments, preprocessing the raw vibration signals generated during the operation of mechanical equipment to obtain a fault feature map includes: Vibration sensors are used to detect the operation of mechanical equipment and obtain the original vibration signals; The original vibration signal is subjected to a fast Fourier transform to obtain a spectrum. The fault feature map is obtained by processing the spectrogram using a generative adversarial network.

[0007] In some embodiments, the local graph structure features output by each layer of the signal denoising branch and the global features output by each layer of the signal ViT branch are respectively aligned in spatial and channel dimensions by the FAM module, and the features are fused by the adaptive fusion unit of the corresponding layer of the AMF branch to obtain fused features. The adaptive fusion unit of the AMF branch feeds back the fused features obtained by fusing aligned local graph structural features and global features to the next level of adaptive fusion unit for processing.

[0008] In some embodiments, the signal denoising branch performs empirical wavelet transform filtering on the fault feature map using wavelet transform technology, inputs the filtered spectrum map into three parallel graph structures based on different physical priors, and inputs each graph structure into a shared multi-layer graph neural network for hierarchical information propagation and reasoning to obtain local graph structure features.

[0009] In some embodiments, the graph structure includes: a first graph based on spectral adjacency, a second graph based on frequency domain harmonic relationships, and a third graph based on known fault characteristic frequency relationships; each graph structure corresponds to one channel.

[0010] In some embodiments, the shared multilayer graph neural network includes a 4-layer GNN structure, with each layer including a parallel three-column graph structure.

[0011] In some embodiments, the adaptive fusion unit uses multiple grouped convolutional kernels of different sizes, including at least horizontal stripes and vertical stripes, to extract multi-scale spatial features of the fusion features of the input from the previous layer in parallel. Multi-path, multi-scale spatial features are concatenated and fed into the EMA attention module. The EMA attention module performs the following operations: grouping features and calculating attention weights in the X and Y directions respectively; performing feature calibration and fusion based on the attention weights to generate an attention weight map; and performing element-wise multiplication of the attention weight map with the original input feature map of the AMF branch.

[0012] In some embodiments, the FAM module employs three 3×3 convolutional layers, with each convolutional layer followed by a batch normalization layer. It performs feature space transformation and distribution adjustment through a feature alignment network, and residual fusion adds the aligned features to the input local graph structure features or global features, calculates channel attention weights, and outputs the aligned feature information.

[0013] In some embodiments, the mechanical fault diagnosis method based on multi-branch fusion further includes: During model training, a dynamic weight averaging strategy is used to adaptively adjust the weights of the loss function for the signal denoising task and the fault diagnosis task.

[0014] A mechanical fault diagnosis device based on multi-branch fusion, comprising: The feature preprocessing unit is used to preprocess the raw vibration signals generated during the operation of mechanical equipment to obtain fault feature maps; An adaptive fusion unit is used to process the fault feature map using a parallel-executed signal denoising branch, a signal ViT branch, and an AMF branch; wherein, the signal denoising branch is used to extract the fault-related frequency domain features in the fault feature map to obtain local graph structure features, and the signal ViT branch is used to capture the global long-range dependencies in the fault feature map to obtain global features; The classification processing unit is used to process the fused features through global average pooling and a fully connected layer to output the fault classification result.

[0015] The technical solution of the above embodiment processes the fault feature map of the original vibration signal using a parallel-executed signal denoising branch and a signal ViT branch. The signal denoising branch extracts fault-related frequency domain features from the fault feature map to obtain local graph structure features, while the signal ViT branch captures global long-range dependencies in the fault feature map to obtain global features. Then, the AMF branch aligns and adaptively fuses the local graph structure features and global features to obtain fused features. Finally, global average pooling and fully connected layers are used to process and output the fault classification result. This technical solution analyzes the signal from three complementary perspectives by constructing a signal denoising branch, a signal ViT branch, and a core AMF branch. Through the multi-scale grouped convolution and EMA attention mechanism of the AMF branch, it can efficiently and accurately capture local details and global long-range dependencies of fault features at different scales and directions. This greatly enhances the feature representation ability of early weak faults and complex faults, thereby significantly improving the accuracy of diagnosis and robustness under different noise interferences, reducing the number of parameters to optimize computational efficiency, and improving model inference speed.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of this application. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a mechanical fault diagnosis method based on multi-branch fusion, as shown in one embodiment. Figure 2 This is a simplified diagram of a mechanical fault diagnosis framework based on multi-branch fusion. Figure 3 This is a detailed framework diagram of mechanical fault diagnosis based on multi-branch fusion, as an example. Figure 4 This is an example of an adaptive fusion processing flowchart; Figure 5 This is a schematic diagram of a mechanical fault diagnosis device based on multi-branch fusion, according to one embodiment. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0019] Those skilled in the art will understand that, unless otherwise stated, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the word “comprising” as used in this application’s specification means the presence of the stated feature, integer, step, or operation, but does not preclude the presence or addition of one or more other features, integers, steps, or operations.

[0020] refer to Figure 1 As shown, Figure 1 This is a flowchart of a mechanical fault diagnosis method based on multi-branch fusion, as exemplified by one of the following: S10, preprocesses the original vibration signals generated during the operation of mechanical equipment to obtain fault feature maps.

[0021] refer to Figure 2 As shown, Figure 2 This is a simplified diagram of a mechanical fault diagnosis framework based on multi-branch fusion. It uses vibration sensors to detect and obtain the original vibration signal, and then preprocesses the original vibration signal to obtain a feature map.

[0022] In some embodiments, step S10 may include: S101 uses a vibration sensor to detect the operation of mechanical equipment and obtain the original vibration signal.

[0023] S102, the original vibration signal is transformed into a spectrum using a fast Fourier transform.

[0024] S103, using a generative adversarial network to "enhance" and "remove falsehoods" from the spectrogram to obtain a fault feature map.

[0025] Specifically, the original vibration signal is first acquired, and then the original vibration signal is subjected to a short-term Fourier transform (STFT) to obtain a noisy spectrum. Finally, a generative adversarial network (GAN) is used to "enhance" and "remove falsehoods" from the noisy original spectrum to generate a clearer and more ideal fault feature map.

[0026] S20, the fault feature map is processed using the parallel execution of the signal denoising branch, the signal ViT (vison transformer) branch, and the AMF (Adaptive Modality Fusion) branch.

[0027] like Figure 2 As shown, the signal denoising branch and the signal ViT branch are executed synchronously to process the input fault feature map. The features are then aligned and fused to obtain the fused features.

[0028] Specifically, the signal denoising branch is based on Graph Neural Network (GNN); the signal ViT branch is the signal vison transformer branch. The vison transformer is a pioneering model applied in the field of computer vision. It treats an image as a sequence of "image patches" and processes it in a way similar to processing word sequences in natural language.

[0029] In some embodiments, the signal denoising branch is used to extract the frequency domain features related to the fault in the fault feature map to obtain local graph structure features.

[0030] Specifically, the signal denoising branch can be an architecture that includes multiple channels and multiple hierarchical GNNs; such as Figure 3 As shown, Figure 3This is a detailed framework diagram of a multi-branch fusion-based mechanical fault diagnosis example. Furthermore, the signal denoising branch can use Empirical Wavelet Transform (EWT) technology to perform empirical wavelet transform filtering on the fault feature map. The filtered spectrum map is then input into three parallel graph structures based on different physical priors. Each graph structure is then input into a shared multi-layer graph neural network (GNN) for hierarchical information propagation and inference to obtain local graph structure features.

[0031] In some embodiments, the signal ViT branch is used to capture global long-range dependencies in the fault feature map to obtain global features. Furthermore, in the signal ViT branch, after each layer of ViT modules, the updated feature information is aligned with the features of the AMF branch through cross-cascaded information fusion before proceeding to the next layer of AMF processing.

[0032] S30, the local graph structure features and global features are aligned through the AMF branch and adaptively fused based on multi-scale grouped convolution and attention mechanism to obtain fused features.

[0033] Specifically, the AMF branch employs an adaptive fusion technique based on multi-scale grouped convolution and attention mechanisms. Further, the local graph structure features output from each layer of the signal denoising branch and the global features output from each layer of the signal ViT branch are aligned spatially and channel-wise using the FAM module, and then fused using the AMF module at the corresponding level of the AMF branch to obtain fused features. The AMF module at the next higher level feeds the fused features back to the AMF module at the next lower level for further processing.

[0034] In some embodiments, the three graph structures based on different physical priors described above may include: a first graph based on spectral adjacency relationships, a second graph based on frequency domain harmonic relationships, and a third graph based on known fault characteristic frequency relationships; wherein, as Figure 3 Each graph structure corresponds to one channel, i.e., the first graph corresponds to channel 1, the second graph corresponds to channel 2, and the third graph corresponds to channel 3.

[0035] For example, the functions of each channel are as follows: Channel 1: Constructing a graph based on spatial adjacency (connecting adjacent frequency points in the spectrum graph); Channel 2: Constructing a graph based on frequency domain harmonic relationships (connecting fundamental frequency and harmonic frequencies); Channel 3: Construct a priori graph based on the physical model (relationship between known fault characteristic frequencies, bearing fault frequencies).

[0036] Accordingly, the aforementioned shared multi-layer graph neural network inputs the three graph structures into a shared 4-layer GNN network for hierarchical information propagation and inference. The output of each GNN layer is aligned and fused using the FAM module of the AMF branch. For example, the 4-layer GNN network may include: Layer 1: Used for single-step feature aggregation; Layer 2: Used to integrate features within a two-step distance; Layer 3: Used to implement multi-step feature interaction; Layer 4: Used to generate globally context-aware feature representations.

[0037] As described in the above embodiments, the scheme deeply embeds physical prior knowledge into the model, which improves learning efficiency and model interpretability. Based on the graph structure of three prior relationships—adjacency, harmonics, and fault frequency—and using a four-layer neural network for hierarchical reasoning, the model can simulate the analysis logic from local to global, transforming domain knowledge into the model's internal structure, guiding the model to perform directional learning, reducing dependence on a large amount of labeled data, and enhancing the physical interpretability of diagnostic results.

[0038] In some embodiments, the AMF branch is used to align and adaptively fuse local graph structural features and global features respectively to obtain fused features; specifically, the FAM module can align and cross-fuse the output features of the signal ViT branch and the signal denoising branch into the AMF branch.

[0039] For example, the FAM module can use three 3×3 convolutional layers, each followed by a batch normalization layer BatchNorm1d. The feature space is transformed and the distribution is adjusted through the feature alignment network align_net. The residual fusion adds the aligned features to the input local graph structure features or global features, calculates the channel attention weights, and outputs the aligned feature information.

[0040] As described in the above embodiments, the scheme achieves deep collaboration between model components through refined cross-cascade fusion and dynamic optimization. The FAM module realizes cross-cascade information fusion rather than simple feature superposition. At each level, the features of the neural network, ViT branch and AMF branch are dynamically aligned and fused to ensure deep complementarity and consistency of information flow.

[0041] S40, the fused features are processed through global average pooling and a fully connected layer to output the fault classification result.

[0042] Specifically, the fused features output by the AMF branch are processed through Global Average Pooling (GAP) and Fully Connected (FC) layers to output fault classification results. Spatial information is robustly aggregated through Global Average Pooling, and the final classification decision is made through the fully connected layer.

[0043] As described in the above embodiments, the signals are analyzed from three complementary perspectives by constructing a signal denoising branch, a signal ViT branch, and a core AMF branch. Through the multi-scale grouped convolution and EMA attention mechanism of the AMF branch, the local details and global long-range dependencies of fault features at different scales and directions can be captured efficiently and accurately. This greatly enhances the feature representation capability of early weak faults and complex faults, realizes deep fusion feature extraction at multiple scales and perspectives, significantly improves diagnostic accuracy and robustness, significantly improves the accuracy of diagnosis and robustness under different noise interferences, reduces the number of parameters to optimize computational efficiency, and improves the model inference speed.

[0044] To make the technical solution of this application clearer, more embodiments are described below.

[0045] In some embodiments, the mechanical fault diagnosis method based on multi-branch fusion of this application uses an adaptive fusion unit to extract multi-scale spatial features of the fusion features of the previous layer input in parallel using grouped convolutional kernels of various sizes, including at least horizontal stripes and vertical stripes; the multi-path multi-scale spatial features are spliced ​​together and sent to the EMA attention module.

[0046] For example, the EMA attention module performs the following operations: Local graph structural features or global features are grouped, and attention weights are calculated in the X and Y directions respectively. Feature calibration and fusion are performed based on the attention weights to generate an attention weight map. The attention weight map is then multiplied element-wise with the original input feature map of the AMF branch.

[0047] refer to Figure 4 As shown, Figure 4 This is an example of an adaptive fusion processing flowchart; by simultaneously capturing local details and global contextual information through multi-scale convolutional kernels, it solves the problem of insufficient single-scale feature extraction. The detailed process steps are as follows: Phase 1: Input and Channel Adjustment; (1) Start: The process begins; (2) Input feature map: Receive the fused features passed from the previous AMF module; (3) 1×1 Conv channel adjustment: First, a 1×1 convolution is used to adjust the number of channels of the fused feature map, reduce the complexity of subsequent calculations or unify the channel dimension, and prepare for subsequent multi-branch processing.

[0048] Second stage: Multi-branch parallel convolution processing; (1) The adjusted features are simultaneously fed into four parallel convolutional paths to extract feature information at different scales; (2) Extracting feature information through multiple pathways: Path 1: Use 1×1 grouped convolution (Gconv) for the most basic local feature extraction; Route 2: Using a combination of 1×3Gconv and 3×1Gconv, this "asymmetric convolution" is equivalent to a standard 3×3 convolution kernel, but with fewer parameters and higher computational efficiency, mainly capturing features in the horizontal direction; Route 3: The structure is similar to Route 2, but the scale is larger. It uses a combination of 1×5Gconv and 5×1Gconv to obtain broader contextual information. Route 4: Using a combination of 1×7Gconv and 7×1Gconv, it has the largest receptive field and is used to capture the most macroscopic features.

[0049] Phase 3: Feature fusion and activation; (1) Feature concatenation Concat: The feature maps output by the four parallel paths are concatenated in the channel dimension to fuse the features extracted at different scales; (2) Batch Normalization (BN): Standardizes the concatenated features to accelerate the training process and improve model stability; (3) ReLU activation: Introduces nonlinearity to enhance the expressive power of the model.

[0050] Phase 4: Enhanced attention mechanisms; (1) EMA Attention Module: a. Feature grouping: Divide the input feature map into several groups on the channels; b. Feature splitting: Divide the features of each group into two sub-features on an equal basis; c, X, Y direction attention extraction; d. Feature correction: Using the attention weights obtained in the previous step, the sub-features after splitting are weighted and corrected to emphasize important features and suppress useless features.

[0051] In the above process, the EMA attention mechanism can capture long-range dependencies between features, enhance discriminative features, and solve the problem of feature information being diluted during the fusion process.

[0052] Phase 5: Residual Connections and Output; (1) Add element-wise to the original input feature map: The feature map enhanced by multi-branch convolution and attention modules is added element-wise to the original input feature map at the beginning of the process. Residual connections can effectively alleviate the gradient vanishing problem in deep networks and ensure that the network can learn effective incremental features.

[0053] (2) Output feature map: The result of the sum is used as the final output.

[0054] The scheme described in the above embodiments provides an efficient and focused feature enhancement mechanism, optimizing computational performance. The strategy of grouping features and processing and recalibrating them separately in the X and Y directions through the AMF attention mechanism achieves precise focusing on key feature directions with low computational complexity. Combined with residual connections, it strengthens the weights of fault-sensitive features while preserving the original context, making the feature enhancement process more efficient and improving the model's deployment feasibility while maintaining high accuracy. The dynamic weight learning mechanism enables the model to adapt to the feature distribution of different input samples, maintaining stable performance even in noisy environments and with incomplete data, enhancing its adaptability. Through the EMA attention mechanism and residual connections, the model's tolerance to signal noise and missing data is improved by 30%, demonstrating stronger practicality in real-world industrial environments and significantly improving robustness. The parallel computation of multi-scale convolutions and the simplified design of the attention mechanism optimize computational efficiency, reducing the number of parameters by 40% and increasing inference speed by 2 times, making it more suitable for practical deployment.

[0055] In some embodiments, the mechanical fault diagnosis method based on multi-branch fusion of this application may further include: during model training, adopting a dynamic weight averaging (DWA) strategy to adaptively adjust the loss function weights of the signal denoising task and the fault diagnosis task.

[0056] For example, such as Figure 4 As shown by the dashed arrow path, the loss function weights of the signal denoising task and the fault diagnosis task are adaptively adjusted by adopting a dynamic weight averaging strategy to achieve collaborative joint learning of the two tasks.

[0057] As described in the above embodiment, by combining the DWA strategy to dynamically adjust the loss weights, the denoising and diagnostic tasks can be dynamically and collaboratively optimized according to the learning progress, avoiding conflicts between tasks and truly maximizing the benefits of joint learning, so that the overall model performance surpasses the simple summation of individual branches.

[0058] The following describes an embodiment of the mechanical fault diagnosis device based on multi-branch fusion of this application, with reference to... Figure 5 , Figure 5This is a schematic diagram of a mechanical fault diagnosis device based on multi-branch fusion, according to one embodiment, including: Feature preprocessing unit 10 is used to preprocess the original vibration signals generated during the operation of mechanical equipment to obtain fault feature maps; The parallel processing unit 20 is used to process the fault feature map using a parallel-executed signal denoising branch and a signal ViT branch; wherein, the signal denoising branch is used to extract the frequency domain features related to the fault in the fault feature map to obtain local graph structure features, and the signal ViT branch is used to capture the global long-range dependencies in the fault feature map to obtain global features. The adaptive fusion unit 30 is used to align the local graph structure features and global features through the AMF branch and obtain fused features through adaptive fusion based on multi-scale grouped convolution and attention mechanism; The classification processing unit 40 is used to process the fused features through global average pooling and a fully connected layer to output the fault classification result.

[0059] The mechanical fault diagnosis device based on multi-branch fusion in this embodiment can execute a mechanical fault diagnosis method based on multi-branch fusion provided in the embodiments of this application. The implementation principle is similar. The actions performed by each module in the mechanical fault diagnosis device based on multi-branch fusion in each embodiment of this application correspond to the steps in the mechanical fault diagnosis method based on multi-branch fusion in each embodiment of this application. For detailed functional descriptions of each module of the mechanical fault diagnosis device based on multi-branch fusion, please refer to the descriptions in the corresponding mechanical fault diagnosis method based on multi-branch fusion shown above. They will not be repeated here.

[0060] In summary, the technical solution of this application adopts a three-level fusion pipeline hierarchical fusion architecture, leveraging the complementary advantages of global features and local graph structure features, and employing multimodal collaboration in intelligent fusion, along with end-to-end optimization of a complete optimization strategy from feature alignment to final fusion. This effectively solves the scale difference and distribution mismatch issues between the global features output by ViT and the local graph structure features extracted by the GNN network; it achieves dynamic adaptive fusion by dynamically adjusting the fusion weights based on the importance of the input features, avoiding suboptimal fusion caused by fixed weights; it maximizes the preservation of discriminative information of the original features during the fusion process, preventing important features from being diluted during fusion; and it reduces computational complexity and the number of parameters while ensuring performance.

[0061] This application provides a technical solution for a computer device to implement the functions related to a mechanical fault diagnosis method based on multi-branch fusion. The computer device of this embodiment includes one or more processors, a memory, and one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured for the steps of the mechanical fault diagnosis method based on multi-branch fusion in any embodiment.

[0062] This application provides a computer-readable storage medium to implement the functions related to a mechanical fault diagnosis method based on multi-branch fusion. The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set. The at least one instruction, at least one program, code set, or instruction set is loaded by a processor and executes the mechanical fault diagnosis method based on multi-branch fusion, as described in any embodiment.

[0063] In an exemplary embodiment, the computer-readable storage medium may be a non-transitory computer-readable storage medium that includes instructions, such as a memory that includes instructions. For example, a non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0064] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A mechanical fault diagnosis method based on multi-branch fusion, characterized in that, include: The original vibration signals generated during the operation of mechanical equipment are preprocessed to obtain fault feature maps; The fault feature map is processed using a parallel-executed signal denoising branch and a signal ViT branch; wherein, the signal denoising branch is used to extract the frequency domain features related to the fault in the fault feature map to obtain local graph structure features, and the signal ViT branch is used to capture the global long-range dependencies in the fault feature map to obtain global features; The local graph structure features and global features are aligned by AMF branches and adaptively fused based on multi-scale grouped convolution and attention mechanisms to obtain fused features; The fused features are processed through global average pooling and a fully connected layer to output fault classification results.

2. The mechanical fault diagnosis method based on multi-branch fusion according to claim 1, characterized in that, Preprocessing the raw vibration signals generated during the operation of mechanical equipment to obtain fault feature maps includes: Vibration sensors are used to detect the operation of mechanical equipment and obtain the original vibration signals; The original vibration signal is subjected to a fast Fourier transform to obtain a spectrum. The fault feature map is obtained by processing the spectrogram using a generative adversarial network.

3. The mechanical fault diagnosis method based on multi-branch fusion according to claim 2, characterized in that, The local graph structure features output by each layer of the signal denoising branch and the global features output by each layer of the signal ViT branch are respectively aligned in spatial and channel dimensions using the FAM module. The adaptive fusion unit of the AMF branch feeds back the fused features obtained by fusing aligned local graph structural features and global features to the next level of adaptive fusion unit for processing.

4. The mechanical fault diagnosis method based on multi-branch fusion according to claim 3, characterized in that, The signal denoising branch performs empirical wavelet transform filtering on the fault feature map using wavelet transform technology. The filtered spectrum map is then input into three parallel graph structures based on different physical priors. Each graph structure is then input into a shared multi-layer graph neural network for hierarchical information propagation and inference to obtain local graph structure features.

5. The mechanical fault diagnosis method based on multi-branch fusion according to claim 4, characterized in that, The graph structure includes: a first graph based on spectral adjacency, a second graph based on frequency domain harmonic relationships, and a third graph based on known fault characteristic frequency relationships; each graph structure corresponds to one channel.

6. The mechanical fault diagnosis method based on multi-branch fusion according to claim 5, characterized in that, The shared multi-layer graph neural network includes a 4-layer GNN structure, with each layer consisting of a parallel three-column graph structure.

7. The mechanical fault diagnosis method based on multi-branch fusion according to claim 6, characterized in that, The adaptive fusion unit uses multiple grouped convolutional kernels of different sizes, including at least horizontal and vertical stripes, to extract multi-scale spatial features of the fusion features of the previous layer input in parallel. Multi-path, multi-scale spatial features are concatenated and fed into the EMA attention module. The EMA attention module performs the following operations: grouping features and calculating attention weights in the X and Y directions respectively; performing feature calibration and fusion based on the attention weights to generate an attention weight map; and performing element-wise multiplication of the attention weight map with the original input feature map of the AMF branch.

8. The mechanical fault diagnosis method based on multi-branch fusion according to claim 1, characterized in that, The FAM module employs three 3×3 convolutional layers, with each convolutional layer followed by a batch normalization layer. It performs feature space transformation and distribution adjustment through a feature alignment network. Residual fusion adds the aligned features to the input local graph structure features or global features, calculates channel attention weights, and outputs the aligned feature information.

9. The mechanical fault diagnosis method based on multi-branch fusion according to claim 1, characterized in that, Also includes: During model training, a dynamic weight averaging strategy is used to adaptively adjust the weights of the loss function for the signal denoising task and the fault diagnosis task.

10. A mechanical fault diagnosis device based on multi-branch fusion, characterized in that, include: The feature preprocessing unit is used to preprocess the raw vibration signals generated during the operation of mechanical equipment to obtain fault feature maps; A parallel processing unit is used to process the fault feature map using a parallel-executed signal denoising branch and a signal ViT branch; wherein, the signal denoising branch is used to extract the frequency domain features related to the fault in the fault feature map to obtain local graph structure features, and the signal ViT branch is used to capture the global long-range dependencies in the fault feature map to obtain global features. An adaptive fusion unit is used to align the local graph structure features and global features through the AMF branch and to obtain fused features through adaptive fusion based on multi-scale grouped convolution and attention mechanisms. The classification processing unit is used to process the fused features through global average pooling and a fully connected layer to output the fault classification result.

Images