CN-122020179-A - Cross-working condition mechanical fault diagnosis method based on depth wavelet self-adaptive graph rolling network

CN122020179ACN 122020179 ACN122020179 ACN 122020179ACN-122020179-A

Abstract

The invention discloses a cross-working condition mechanical fault diagnosis method based on a depth wavelet self-adaptive graph rolling network, and relates to the technical field of mechanical fault diagnosis. The method comprises the steps of integrating depth separable convolution and wavelet convolution through a depth wavelet convolution network, enhancing the extraction capability of low-frequency components in vibration signals, achieving comprehensive multi-scale feature extraction, dynamically selecting optimal neighbor numbers for each node according to node feature cosine similarity by utilizing a self-adaptive graph generating network, constructing exclusive graph topological structures adapting to different working conditions, dynamically adjusting countermeasure intensity according to source domain and target domain differences by adopting a self-adaptive countermeasure network, maintaining stable balance between a feature extractor and a domain discriminator, and improving cross-domain alignment effect. The method can effectively improve the accuracy, robustness and generalization capability of bearing fault diagnosis, and is suitable for real-time fault detection and identification in complex industrial scenes.

Inventors

Fei Xihong
WANG XIAOTIAN
XIA XIN
WANG KANG
FANG TIAN
SHEN HAO

Assignees

安徽工业大学

Dates

Publication Date: 20260512
Application Date: 20260210

Claims (10)

1. A cross-working condition mechanical fault diagnosis method based on a depth wavelet self-adaptive graph rolling network is characterized by comprising the following steps: Step one, acquiring a bearing vibration signal, and preprocessing the bearing vibration signal; Inputting the preprocessed vibration signal into a depth wavelet convolution network to extract multi-scale time-frequency characteristics; Inputting the multi-scale time-frequency characteristics extracted through the depth wavelet convolutional network into the convolutional neural network to extract deep characteristics of data, and generating graph node characteristic representation by means of a full-connection layer; inputting graph node characteristics into a self-adaptive graph generation network, performing nonlinear mapping on the input graph node characteristics by using a multi-layer perceptron to obtain node attribute representation, dynamically selecting the optimal number of neighbors for each node based on node characteristic cosine similarity, and constructing a dedicated graph topological structure aiming at different working conditions; inputting a graph structure constructed through the self-adaptive graph generation network into a multi-receptive field graph convolution network, and extracting multi-scale graph features; step six, inputting the extracted characteristics into a classifier, a structure alignment device and a domain discriminator respectively to execute classification tasks, structure alignment tasks and domain discrimination tasks, and carrying out fault detection and identification on the collected vibration signals.
2. The method for diagnosing the cross-working condition mechanical fault based on the depth wavelet adaptive graph rolling network according to claim 1, wherein the preprocessing in the first step comprises the step of intercepting continuous vibration signals from a signal starting point segment by taking sampling points with fixed window lengths as step sizes, and no overlapping mechanism is adopted in the segmentation process, so that complete independence among segmented signals is ensured.
3. The cross-working condition mechanical fault diagnosis method based on the depth wavelet self-adaptive graph convolution network is characterized in that the depth wavelet convolution network comprises a depth convolution module, a wavelet convolution module and a point-by-point convolution module, the wavelet convolution module extracts multi-scale features of vibration signals through multi-scale decomposition capacity of wavelet transformation, the multi-scale features extracted by the wavelet convolution module are fused with the channel-by-channel convolution features extracted by the depth convolution module, and then the fused multi-scale features are input to the point-by-point convolution module to be fused with cross-channel information, and the multi-channel features are mapped to target channel dimensions to obtain multi-scale time-frequency features.
4. The cross-working condition mechanical fault diagnosis method based on the depth wavelet self-adaptive graph rolling network according to claim 3, wherein the wavelet convolution module builds a wavelet decomposition filter and a reconstruction filter based on a specified wavelet type to carry out two-dimensional wavelet decomposition on an input signal, when the input signal is propagated in the forward direction, the input signal is subjected to multi-scale wavelet decomposition first, each layer of decomposition generates a low-frequency component set and a high-frequency component set of a current scale, then the enhancement features of a convolution layer are combined with scaling to adjust the contribution weights of features of different scales, then the components subjected to multi-scale processing are reconstructed through inverse wavelet transformation, and the components are added with a basic depth convolution output through residual connection to extract multi-scale features.
5. The method for diagnosing a cross-working condition mechanical fault based on a depth wavelet adaptive graph rolling network according to claim 3, wherein the adaptive graph generating network in the fourth step adaptively calculates the association strength between nodes through a similarity measurement mechanism based on node attributes, normalizes the similarity, and generates a dynamic neighbor number for each node according to the statistical characteristics of the similarity between nodes.
6. The cross-working condition mechanical fault diagnosis method based on the depth wavelet adaptive graph rolling network according to claim 5, wherein the adaptive graph generating network selects a plurality of most representative neighbor nodes from all candidate nodes through a top-k ordering mechanism for each node, and generates corresponding edge weights based on similarity among the nodes, so that the generated graph structure can keep consistent with working condition changes.
7. The cross-working condition mechanical fault diagnosis method based on the depth wavelet adaptive graph rolling network according to claim 3, wherein the multi-receptive field graph rolling network in the fifth step extracts local multi-scale features on each node of the graph structure through convolution operation of setting different receptive field ranges, performs information fusion of the data structure, and performs multi-scale graph feature extraction through fusion of different scale features.
8. A cross-working condition mechanical fault diagnosis method based on depth wavelet adaptive graph rolling network according to claim 3, wherein the domain discriminator in the sixth step adopts an adaptive countermeasure network, and dynamically adjusts the countermeasure intensity according to the domain difference between the source domain and the target domain.
9. The method for diagnosing a cross-operating mode mechanical fault based on a depth wavelet adaptive graph rolling network as claimed in claim 8, wherein said adaptive countermeasure network adjusts countermeasure intensity under different operating modes by an adaptive mechanism, and first calculates initial countermeasure intensity According to the initial countermeasure intensity The adaptive countermeasure network adaptively adjusts the countermeasure intensity according to the accuracy of the domain discriminator, and the specific procedure is as follows: Wherein, the In order to adjust the step size of the step, For the accuracy of the domain discriminator, Is a set threshold value.
10. The method for diagnosing a cross-operating mode mechanical failure based on a depth wavelet adaptive graph rolling network as claimed in claim 9, wherein the initial countermeasure strength is The calculation formula of (2) is as follows: Wherein, the In order to counter the strength of the steel, In order to counter the upper limit of the intensity, In order to counter the lower limit of the intensity, In order to be the steepness of the curve, For the current number of iterations, Is the maximum number of iterations.

Description

Cross-working condition mechanical fault diagnosis method based on depth wavelet self-adaptive graph rolling network Technical Field The invention relates to the technical field of mechanical fault diagnosis, in particular to a cross-working condition mechanical fault diagnosis method based on a depth wavelet self-adaptive graph rolling network. Background Bearings are the key components in rotating machinery and are the core content in the field of equipment operation and maintenance. Bearing fault diagnosis research is helpful for improving equipment reliability and safety, reducing downtime and maintenance cost, and prolonging equipment life. The conventional fault diagnosis method mostly depends on a signal processing technology and a machine learning algorithm, such as a Support Vector Machine (SVM), a Random Forest (Random Forest), a clustering algorithm, an extreme learning machine and the like. While machine learning has made significant progress in the field of fault diagnosis, it relies on expert knowledge driven feature engineering resulting in high costs. Meanwhile, the shallow model has the problems of insufficient performance and weak generalization capability under complex working conditions. In recent years, due to the upgrade of computing hardware and acquisition equipment, deep learning techniques have been widely used in the field of fault diagnosis. The deep learning model realizes end-to-end automatic feature extraction through a multi-layer structure, wherein the most representative is Convolutional Neural Network (CNN), and the deep learning model is widely applied due to the strong feature extraction capability. Although the deep learning method typified by CNN has achieved excellent results in the field of mechanical failure diagnosis, there are some limitations. The main reason is that CNN-based methods can only handle data of regular grid structure, but cannot model complex, non-euclidean associations between nodes effectively. While there is little data of a regular grid structure in a practical complex industrial scenario, this limits the application of intelligent fault diagnosis in a practical industrial scenario. The Graph Neural Network (GNN) is capable of modeling the association of nodes with edges, more naturally representing the physical relationships between the device components. And capturing complex spatial relationships and structural modes contained in vibration signals from a plurality of global and local layers by means of strong graph structural feature extraction capability. Particularly in the modeling of the graph structure of the bearing vibration signal. Typical graph-rolling networks such as graph attention network (Graph Attention Networks, GAT), graph sampling and aggregation network (GRAPHSAGE), graph isomorphic network (Graph Isomorphism Network, GIN), time-series graph-rolling network (Topology ADAPTIVE GRAPH Convolutional Networks, TAGCN), etc., have been primarily used in the field of bearing fault diagnosis. Although the GNN-based method exhibits good performance in single-condition bearing fault diagnosis, it still has the problem of difficult graph construction and easy distortion in variable-condition fault diagnosis, so that the cross-condition generalization capability is limited. In an actual industrial scenario, bearings often operate under conditions of continuously changing rotational speed, load, etc. Due to the distribution difference of vibration signals collected under different working conditions, the generalization performance of a fault diagnosis model trained under a single working condition is obviously reduced in an actual industrial scene. Therefore, the cross-working condition fault diagnosis research is developed, and the method has important significance for realizing accurate diagnosis of equipment states, avoiding catastrophic accidents and improving the intelligent level of operation and maintenance. Disclosure of Invention 1. Technical problem to be solved by the invention In order to solve the problems that a traditional graph neural network model in an actual industrial scene is difficult to fully capture key information contained in a sample, the model is insufficient in generalization and robustness due to complex and changeable working conditions, and the like, and meanwhile, the physical consistency, the self-adaption capability and the diagnosis precision of the model are considered, the invention provides a cross-working condition fault diagnosis method based on a depth wavelet self-adaption Domain countermeasure graph convolution network (DEEP WAVELET ADAPTIVE-based Domain ADVERSARIAL GRAPH Convolutional Network, DWA-DAGCN). The invention proposes a deep wavelet convolutional network (DEEP WAVELET-based Adaptive Domain ADVERSARIAL GRAPH Convolutional Network, DTConv). The depth separable convolution and the wavelet convolution are combined to enhance the extraction capability of low-frequency components in the vi