CN-121997274-A - Mechanical fault diagnosis method and device based on multi-branch fusion

Abstract

The application relates to a mechanical fault diagnosis method and device based on multi-branch fusion, wherein the method comprises the steps of preprocessing an original vibration signal to obtain a fault characteristic diagram, processing fault characteristics through an AMF (advanced mechanical fiber) branch by utilizing a signal denoising branch and a signal ViT branch which are executed in parallel, wherein the signal denoising branch is used for extracting frequency domain characteristics related to faults in the fault characteristic diagram to obtain local diagram structural characteristics, the signal ViT branch is used for capturing global long-range dependency relations in the fault characteristic diagram to obtain global characteristics, the AMF branch is used for aligning the local diagram structural characteristics and the global characteristics and obtaining fusion characteristics based on multi-scale grouping convolution and self-adaptive fusion of an attention mechanism, and the fusion characteristics are processed through a global average pooling layer and a full-connection layer to output fault classification results.

Inventors

• PENG WEICHAO
• HUANG JIANFENG
• WAN KAI
• CHEN SHAOBING

Assignees

• 惠州学院

Dates

Publication Date

20260508

Application Date

20260318

Claims (10)

1. 1. A mechanical fault diagnosis method based on multi-branch fusion, comprising: preprocessing an original vibration signal generated in the operation of mechanical equipment to obtain a fault characteristic diagram; processing the fault feature map 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 frequency domain features related to faults in the fault feature map to obtain local map structural features, and the signal ViT branch is used for capturing global long-range dependency relations in the fault feature map to obtain global features; the local graph structural features and the global features are aligned through AMF branches, and fusion features are obtained based on multi-scale grouping convolution and self-adaptive fusion of an attention mechanism; And processing the fusion characteristics through a global average pooling layer and a full connection layer to output a fault classification result.
2. 2. The multi-branch fusion-based mechanical fault diagnosis method according to claim 1, wherein preprocessing an original vibration signal generated during operation of a mechanical device to obtain a fault signature comprises: Detecting the operation of mechanical equipment by adopting a vibration sensor to obtain an original vibration signal; Obtaining a spectrogram from the original vibration signal through fast Fourier transform; And processing the spectrogram by using a generated countermeasure network to obtain a fault characteristic diagram.
3. 3. The multi-branch fusion-based mechanical fault diagnosis method according to claim 2, wherein the local graph structural feature of each layer output of the signal denoising branch and the global feature of each layer output of the signal ViT branch are respectively aligned in space and channel dimension by a FAM module; And the adaptive fusion unit of the AMF branch feeds back fusion characteristics obtained by fusing the aligned local graph structural characteristics and global characteristics to the adaptive fusion unit of the next level for processing.
4. 4. The mechanical fault diagnosis method based on multi-branch fusion according to claim 3, wherein the signal denoising branch performs empirical wavelet transform filtering processing on the fault feature map through wavelet transform technology, the filtered spectrogram is input into three parallel map structures based on different physical priors, and each map structure is respectively input into a shared multi-layer map neural network for hierarchical information propagation and reasoning to obtain the local map structural feature.
5. 5. The method for diagnosing a mechanical failure based on multi-branch fusion according to claim 4, wherein said graph structures comprise a first graph based on a frequency spectrum adjacency relationship, a second graph based on a frequency domain harmonic relationship, and a third graph based on a known failure feature frequency relationship, each graph structure corresponding to a channel.
6. 6. The multi-branch fusion-based mechanical fault diagnosis method according to claim 5, wherein the shared multi-layer graph neural network comprises a 4-layer GNN structure, each layer comprising a parallel three-column graph structure.
7. 7. The mechanical fault diagnosis method based on multi-branch fusion according to claim 6, wherein the adaptive fusion unit extracts multi-scale spatial features of fusion features of the previous layer input in parallel using a plurality of different-sized grouping convolution kernels including at least horizontal stripes and vertical stripes; The method comprises the steps of splicing multiple paths of multi-scale space features and sending the spliced multi-path multi-scale space features to an EMA attention module, wherein the EMA attention module groups the features, calculates attention weights in the X direction and the Y direction respectively, performs feature calibration and fusion according to the attention weights to generate an attention weight map, and performs element-by-element multiplication operation on the attention weight map and an original input feature map of an AMF branch.
8. 8. The mechanical fault diagnosis method based on multi-branch fusion according to claim 1, wherein the FAM module adopts three 3×3 convolution layers, each convolution layer is followed by a batch normalization layer, feature space transformation and distribution adjustment are performed through a feature alignment network, residual fusion adds the aligned features to the input local graph structural features or global features, calculates channel attention weights, and outputs the aligned feature information.
9. 9. The multi-branch fusion-based mechanical fault diagnosis method according to claim 1, further comprising: In the model training process, a dynamic weight average strategy is adopted to adaptively adjust the loss function weights of the signal denoising task and the fault diagnosis task.
10. 10. A mechanical fault diagnosis device based on multi-branch fusion, comprising: The feature preprocessing unit is used for preprocessing an original vibration signal generated in the operation of the mechanical equipment to obtain a fault feature map; The parallel processing unit is used for processing the fault characteristic map by utilizing a signal denoising branch and a signal ViT branch which are executed in parallel, wherein the signal denoising branch is used for extracting frequency domain characteristics related to faults in the fault characteristic map to obtain local map structural characteristics, and the signal ViT branch is used for capturing global long-range dependency relations in the fault characteristic map to obtain global characteristics; The self-adaptive fusion unit is used for aligning the local graph structural features and the global features through AMF branches and obtaining fusion features based on multi-scale grouping convolution and self-adaptive fusion of an attention mechanism; and the classification processing unit is used for processing the fusion characteristics through a global average pooling layer and a full connection layer to output a fault classification result.

Description

Mechanical fault diagnosis method and device based on multi-branch fusion Technical Field The application relates to the technical field of machine equipment fault diagnosis, in particular to a mechanical fault diagnosis method and device based on multi-branch fusion. Background The reliable running of the rolling bearing serving as a core component of the rotary machine is crucial to industrial application, the traditional state monitoring method mainly relies on extracting distinguishing features reflecting the health state of the bearing from vibration signals, however, under complex working conditions, fault information of the rolling bearing is often covered by noise and other interferences, so that the traditional method is difficult to effectively extract the key features. In recent years, deep learning technology, particularly convolutional neural networks (Convolutional Neural Network, CNN), has become a powerful tool for mechanical fault diagnosis and can realize automatic learning of mechanical signal characteristics, for example, the capability of feature extraction has been improved by introducing symmetrical point diagram representation, improved integrated empirical mode decomposition or a multi-scale convolution structure and other modes. Although convolution neural network-based approaches achieve some success, they are often more focused on local feature extraction, and lack modeling capabilities for global long-range dependencies, thereby limiting the performance of building a comprehensive fault characterization. In addition, mechanical systems are often in harsh operating environments, vibration signals are susceptible to noise pollution, and existing methods often separate denoising from diagnostic processes, which may result in part of the effective information being filtered out, thereby affecting diagnostic accuracy. Disclosure of Invention In order to solve one of the defects, the application provides a mechanical fault diagnosis method and device based on multi-branch fusion, which improves the fault diagnosis precision. A mechanical fault diagnosis method based on multi-branch fusion, comprising: preprocessing an original vibration signal generated in the operation of mechanical equipment to obtain a fault characteristic diagram; processing the fault feature map 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 frequency domain features related to faults in the fault feature map to obtain local map structural features, and the signal ViT branch is used for capturing global long-range dependency relations in the fault feature map to obtain global features; the local graph structural features and the global features are aligned through AMF branches, and fusion features are obtained based on multi-scale grouping convolution and self-adaptive fusion of an attention mechanism; And processing the fusion characteristics through a global average pooling layer and a full connection layer to output a fault classification result. In some embodiments, preprocessing a raw vibration signal generated during operation of a mechanical device to obtain a fault signature includes: Detecting the operation of mechanical equipment by adopting a vibration sensor to obtain an original vibration signal; Obtaining a spectrogram from the original vibration signal through fast Fourier transform; And processing the spectrogram by using a generated countermeasure network to obtain a fault characteristic diagram. In some embodiments, the local graph structural feature output by each layer of the signal denoising branch and the global feature output by each layer of the signal ViT branch are respectively aligned with the features of the space and the channel dimension through the FAM module, and the feature fusion is performed through the adaptive fusion units of the corresponding layers of the AMF branch to obtain the fusion feature; And the adaptive fusion unit of the AMF branch feeds back fusion characteristics obtained by fusing the aligned local graph structural characteristics and global characteristics to the adaptive fusion unit of the next level for processing. In some embodiments, the signal denoising branch performs empirical wavelet transform filtering processing on the fault feature map through wavelet transform technology, inputs the filtered spectrogram into three parallel map structures based on different physical priors, and inputs each map structure into a shared multi-layer map neural network to perform hierarchical information propagation and reasoning to obtain local map structural features. In some embodiments, the graph structures include a first graph based on a spectral adjacency, a second graph based on a frequency domain harmonic relationship, and a third graph based on a known fault signature frequency relationship, each graph structure corresponding to a channel. In some embodiments,