Cross-working-condition mechanical fault diagnosis method based on deep wavelet adaptive graph convolutional network

Abstract

The invention discloses a cross-working-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network, and relates to the technical field of mechanical fault diagnosis. According to the method, deep separable convolution and wavelet convolution are fused through a deep wavelet convolution network, so that the capability of extracting low-frequency components in vibration signals is enhanced, and comprehensive multi-scale feature extraction is realized; dynamically selecting the optimal number of neighbors for each node according to the node feature cosine similarity by using an adaptive graph generation network, and constructing an exclusive graph topological structure adapted to different working conditions; the self-adaptive adversarial network is adopted to dynamically adjust the adversarial strength according to the difference between the source domain and the target domain, the stable balance between the feature extractor and the domain discriminator is maintained, and the cross-domain alignment effect is improved. The method can effectively improve the accuracy, robustness and generalization ability of bearing fault diagnosis, and is suitable for real-time fault detection and recognition in a complex industrial scene.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, and more specifically, to a cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network. Background Technology

[0002] Bearings, as critical components in rotating machinery, are a core element of equipment operation and maintenance. Research on bearing fault diagnosis helps improve equipment reliability and safety, reduce downtime and maintenance costs, and extend equipment lifespan. Traditional fault diagnosis methods largely rely on signal processing techniques and machine learning algorithms, such as Support Vector Machines (SVM), Random Forests, clustering algorithms, and Extreme Learning Machines. While machine learning has made significant progress in fault diagnosis, its reliance on expert-driven feature engineering leads to high costs. Furthermore, shallow models generally suffer from insufficient performance and weak generalization ability under complex operating conditions.

[0003] In recent years, upgrades in computing hardware and data acquisition equipment have led to the widespread use of deep learning technology in the field of fault diagnosis. Deep learning models achieve end-to-end automatic feature extraction through multi-layered structures, with Convolutional Neural Networks (CNNs) being the most representative and widely used due to their powerful feature extraction capabilities. Although deep learning methods, represented by CNNs, have achieved excellent results in mechanical fault diagnosis, some limitations remain. The main reason is that CNN-based methods can only process data with regular grid structures and cannot effectively model complex, non-Euclidean relationships between nodes. However, regular grid structures are almost nonexistent in real-world complex industrial scenarios, which limits the application of intelligent fault diagnosis in practical industrial settings.

[0004] Graph Neural Networks (GNNs) can model the relationships between nodes and edges, representing the physical relationships between equipment components more naturally. Leveraging their powerful graph structure feature extraction capabilities, they can capture the complex spatial relationships and structural patterns inherent in vibration signals at multiple global and local levels. They have shown particular potential in graph structure modeling of bearing vibration signals. Typical graph convolutional networks, such as Graph Attention Networks (GAT), GraphSAGE, Graph Isomorphism Networks (GIN), and TopologyAdaptive Graph Convolutional Networks (TAGCN), have been initially applied in bearing fault diagnosis. Although GNN-based methods have shown good performance in bearing fault diagnosis under single operating conditions, they still face challenges in graph construction and distortion under varying operating conditions, limiting their generalization ability across operating conditions.

[0005] In real-world industrial scenarios, bearings often operate under constantly changing conditions, such as rotational speed and load. Because the vibration signals collected under different operating conditions exhibit varying distributions, fault diagnosis models trained only under a single operating condition show a significant decrease in generalization performance in real-world industrial environments. Therefore, conducting research on cross-operating-condition fault diagnosis is of great significance for achieving accurate equipment condition diagnosis, preventing catastrophic accidents, and improving the level of intelligent operation and maintenance. Summary of the Invention

[0006] 1. The technical problem that the invention aims to solve To address the challenges of traditional graph neural network models in real-world industrial scenarios, such as their inability to fully capture key information from samples and insufficient generalization and robustness due to complex and variable working conditions, while simultaneously ensuring physical consistency, adaptability, and diagnostic accuracy, this invention proposes a cross-working-condition fault diagnosis method based on a Deep Wavelet-based Adaptive-based Domain Adversarial Graph Convolutional Network (DWA-DAGCN). This invention proposes a Deep Wavelet-based Adaptive Domain Adversarial Graph Convolutional Network (DTConv). By combining depthwise separable convolution with wavelet convolution, it enhances the extraction of low-frequency components from vibration signals and improves the comprehensiveness of node feature representation, thereby improving the model's fault diagnosis performance. Simultaneously, an Adaptive Graph Generation Network (AGGNet) is designed. The optimal number of neighbors for each node is dynamically selected based on the cosine similarity of node features, and a dedicated graph topology is constructed for different working conditions, thereby enhancing the model's generalization ability under different conditions. Furthermore, this invention proposes a dynamic adversarial algorithm (Adaptive Adversarial Network, AANet). Based on the domain differences between the source and target domains, the adversarial strength is adaptively adjusted to maintain a stable adversarial balance between the feature extractor and the domain discriminator under complex working conditions, improving cross-domain alignment and enhancing the model's robustness.

[0007] 2. Technical Solution To achieve the above objectives, the technical solution provided by this invention is as follows: The present invention provides a cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network, comprising the following steps: Step 1: Acquire bearing vibration signals and preprocess them; Step 2: Input the preprocessed vibration signal into a deep wavelet convolutional network for multi-scale time-frequency feature extraction; Step 3: Input the multi-scale time-frequency features extracted by the deep wavelet convolutional network into the convolutional neural network to extract the deep features of the data, and generate graph node feature representations with the help of fully connected layers; Step 4: Input the graph node features into the adaptive graph generation network, use a multilayer perceptron to perform nonlinear mapping on the input graph node features to obtain node attribute representations, dynamically select the optimal number of neighbors for each node based on the cosine similarity of node features, and construct a dedicated graph topology structure for different working conditions. Step 5: Input the graph structure constructed by the adaptive graph generation network into the multi-receptive field graph convolutional network to perform multi-scale graph feature extraction; Step 6: Input the extracted features into the classifier, structure aligner, and domain discriminator respectively to perform classification, structure alignment, and domain discrimination tasks; perform fault detection and identification on the collected vibration signals.

[0008] Furthermore, the preprocessing in step one includes segmenting the continuous vibration signal from the signal start point, with sampling points of fixed window length as the step size, without using any overlapping mechanism during the segmentation process to ensure that the segmented signals are completely independent.

[0009] Furthermore, the deep wavelet convolutional network includes a deep convolutional module, a wavelet convolutional module, and a pointwise convolutional module. The wavelet convolutional module extracts multi-scale features from the vibration signal through the multi-scale decomposition capability of wavelet transform. The multi-scale features extracted by the wavelet convolutional module are fused with the channelwise convolutional features extracted by the deep convolutional module and then input into the pointwise convolutional module for cross-channel information fusion, mapping the multi-channel features to the target channel dimension to obtain multi-scale time-frequency features.

[0010] Furthermore, the wavelet convolution module constructs wavelet decomposition filters and reconstruction filters based on a specified wavelet type to perform two-dimensional wavelet decomposition on the input signal. During forward propagation, the input signal is first decomposed into multi-scale wavelet components, with each layer generating a set of low-frequency and high-frequency components at the current scale. Then, features are enhanced through convolutional layers, and the contribution weights of features at different scales are adjusted by scaling. Subsequently, the multi-scale processed components are reconstructed through inverse wavelet transform, and added to the output of the basic depth convolution via residual connection to extract multi-scale features.

[0011] Furthermore, the adaptive graph generation network in step four adaptively calculates the association strength between nodes through a similarity measurement mechanism based on node attributes, normalizes the similarity, and generates a dynamic number of neighbors for each node based on the statistical characteristics of the similarity between nodes.

[0012] Furthermore, the adaptive graph generation network selects several of the most representative neighbor nodes for each node from all candidate nodes through a top-k sorting mechanism, and generates corresponding edge weights based on the similarity between nodes, so that the generated graph structure can remain consistent with changes in working conditions.

[0013] Furthermore, in step five, the multi-receptive-field graph convolutional network extracts local multi-scale features at each node of the graph structure by setting convolution operations with different receptive field ranges, and performs information fusion of the data structure. By fusing features at different scales, multi-scale graph feature extraction is performed.

[0014] Furthermore, the domain discriminator in step six employs an adaptive adversarial network to dynamically adjust the adversarial strength based on the domain differences between the source and target domains.

[0015] Furthermore, the adaptive adversarial network adjusts the adversarial strength under different operating conditions through an adaptive mechanism, first calculating the initial adversarial strength. Based on the initial confrontation strength The adaptive adversarial network adaptively adjusts the adversarial strength based on the accuracy of the domain discriminator. The specific process is as follows:

[0016] in, To adjust the step size, For the accuracy of the domain discriminator, The set threshold value.

[0017] Furthermore, the initial intensity of confrontation The calculation formula is as follows:

[0018] in, To counter the intensity, To counter the strength limit, To counteract the lower limit of intensity, For the steepness of the curve, This represents the current iteration number. This represents the maximum number of iterations.

[0019] 3. Beneficial effects Compared with the prior art, the technical solution provided by this invention has the following advantages: (1) The present invention provides a cross-condition mechanical fault diagnosis method based on deep wavelet adaptive graph convolutional network. By fusing deep separable convolution and wavelet convolution, it enhances the ability to extract low-frequency components in vibration signals, enabling the model to achieve comprehensive feature extraction and high extraction accuracy even in complex industrial environments. This effectively avoids the problem of the model being unable to fully decompose and extract multi-scale features under complex conditions, thus providing higher accuracy and reliability for bearing fault diagnosis.

[0020] (2) The present invention provides a cross-condition mechanical fault diagnosis method based on deep wavelet adaptive graph convolutional network. It designs an adaptive graph generation network (AGGNet), dynamically selects the optimal number of neighbors for each node according to the cosine similarity of node features, and constructs a dedicated graph topology for different working conditions. This provides dynamic adaptation capability for feature association modeling under complex working conditions, thereby alleviating the problem that traditional graph construction methods cannot adapt to changes in data distribution under complex working conditions.

[0021] (3) The present invention provides a cross-condition mechanical fault diagnosis method based on deep wavelet adaptive graph convolutional network. It designs a novel adaptive adversarial network (AANet) that can adaptively adjust the adversarial strength according to the real-time domain difference, thereby enhancing the generalization ability and training stability of the model. Attached Figure Description

[0022] Figure 1 The flowchart shows the cross-condition fault diagnosis method based on deep wavelet adaptive domain adversarial graph convolutional network of the present invention. Figure 2 This is a structural diagram of the Depth Wavelet Adaptive Domain Adversarial Graph Convolutional Network (DWA-DAGCN Network) of the present invention; Figure 3 This is a structural diagram of the depth wavelet convolutional network (DTConv module) of the present invention; Figure 4 This is a diagram illustrating the wavelet convolution process of the present invention. Figure 5 This is a diagram illustrating the core mechanism of the adaptive graph generation network (AGGNet module) of the present invention. Figure 6 This is a CWRU confusion matrix diagram for the diagnostic assessment of this invention; Figure 7 This is a HUST confusion matrix diagram for the diagnostic assessment of this invention; Figure 8 This is a visualization of the diagnostic assessment features of CWRU t-SNE according to the present invention.

[0023] Figure 9 This is a visualization of the HUST t-SNE features for diagnostic evaluation according to the present invention. Detailed Implementation

[0024] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments.

[0025] Example 1 This embodiment utilizes a publicly available bearing fault dataset from Case Western Reserve University (CWRU). This dataset was acquired using a dedicated platform, including a motor, torque sensor, power meter, and the rolling bearing under test. Vibration signals were acquired at frequencies of 12 kHz or 48 kHz. The dataset covers different fault types (such as inner race faults, outer race faults, and rolling element faults) and damage of varying sizes, simulating bearing operation under different health conditions. The dataset also includes motor operating data at multiple speeds, suitable for validating domain adaptive and transfer learning algorithms under varying operating conditions.

[0026] This embodiment also utilizes the publicly available HUST dataset from Hanoi University of Science and Technology (HUST). The dataset was acquired using a dedicated platform comprising a dynamometer, powder brake, torque sensor, bearing under test, accelerometer, 1-horsepower induction motor, measurement module, and frequency converter, at a sampling frequency of 51200Hz. The dataset covers six fault types (inner ring fault, outer ring fault, rolling element fault, and their pairwise combinations) for five types of bearings under three operating conditions. It effectively verifies the model's generalization and robustness under complex operating conditions and unknown equipment conditions, serving as an ideal benchmark for cross-condition fault diagnosis.

[0027] Combination Figure 1 The present embodiment of a cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network includes the following steps: Step 1: Construct a bearing vibration signal dataset containing various working conditions. After preprocessing the original vibration signals, divide them into training, validation, and test sets. The preprocessing includes: introducing Gaussian noise into the original vibration signals to simulate industrial site interference; normalizing and standardizing the signals; segmenting continuous vibration signals in the dataset using a non-overlapping segmentation strategy with a fixed window length of 1024; and assigning a health status label corresponding to the source dataset to each segmented sample. The specific process is as follows: First, Gaussian noise is introduced into the original vibration signal to simulate signal interference in an industrial environment, thereby improving the model's robustness to weak fault characteristics and environmental noise. Then, the signal is normalized to ensure the input data is of a uniform magnitude; further standardization is performed to eliminate numerical differences between different channels and enhance the stability of the model training process. Next, starting from the signal's starting point, segments are extracted in increments of 1024 sampling points until the signal ends. No overlapping mechanisms are used during segmentation to ensure complete independence between samples. Each segment sample is assigned a health status label (0-9, 10 classes) corresponding to its source dataset. Finally, the training and validation sets are divided in an 8:2 ratio for subsequent model training.

[0028] Step 2: Input the vibration signal samples from the training set into the constructed deep wavelet convolutional network (hereinafter referred to as DTConv network) to achieve high-quality multi-scale feature extraction while reducing computational costs.

[0029] The DTConv network integrates depthwise separable convolution with wavelet convolution, effectively capturing high-frequency and low-frequency components in the signal and preserving discriminative time-frequency local features. While maintaining the high efficiency of depthwise separable convolution, it significantly enhances the multi-scale feature extraction capability. By weightedly fusing multi-channel wavelet features through pointwise convolutional layers, it achieves cross-channel information integration, thereby solving the problem of incomplete feature extraction in bearing fault diagnosis.

[0030] Combination Figure 3 , Figure 4 The DTConv network consists of a deep convolution module, a wavelet convolution module, and a pointwise convolution module. Through the synergy of multi-scale decomposition and reconstruction of wavelet convolution and deep convolution operations, it retains the local feature capture capability of traditional deep convolution and uses the "scaling-translation" characteristics of wavelet convolution to separate multi-frequency components of the signal, providing comprehensive feature support for fault diagnosis under complex working conditions.

[0031] The wavelet convolution branch is the core feature extraction unit, achieving multi-band feature separation and enhancement through iterative fusion of multi-scale wavelet transform and convolution operations. The wavelet convolution module first constructs wavelet decomposition and reconstruction filters based on a specified wavelet type (e.g., db1), performing two-dimensional wavelet decomposition on the input signal, splitting it into low-frequency approximation (LL) components and three high-frequency detail (LH, HL, HH) components. During forward propagation, the input signal undergoes multi-scale wavelet decomposition, with each layer generating a set of low-frequency and high-frequency components at the current scale. Then, features are enhanced through convolutional layers, and a scaling module adjusts the contribution weights of features at different scales to strengthen fault-related frequency bands. Subsequently, the multi-scale processed components are reconstructed through inverse wavelet transform and added to the output of the base depth convolution via residual concatenation, preserving the original feature information while incorporating the enhanced features from the multi-scale wavelet processing. This process enables the wavelet convolution module to simultaneously capture both the low-frequency global trend and the high-frequency local impact of the signal, overcoming the limitation of traditional convolution features having a single perspective.

[0032] The wavelet convolution module independently extracts multi-scale features for each input channel, leveraging the wavelet's multi-band decomposition capability to assign differentiated time-frequency features to different channels. The multi-scale features extracted by the wavelet convolution module are fused with the channel-wise convolution features extracted by the depthwise convolution module and then input into the pointwise convolution module for cross-channel information fusion, mapping the multi-channel features to the target channel dimension and achieving efficient feature integration.

[0033] The DTConv network significantly reduces the number of parameters while maintaining accuracy, making it suitable for lightweight industrial deployments. In the field of bearing fault diagnosis, when processing bearing vibration signals, the DTConv network first performs multi-scale decomposition and enhancement of each channel signal to capture fault characteristics of different channels. Then, it fuses multi-channel, multi-scale features through point-by-point convolution, providing subsequent modules with high-quality features that combine multi-frequency band discriminative power and global correlation.

[0034] In this embodiment, the DTConv network overcomes the limitations of traditional convolution in extracting non-stationary signals by leveraging wavelet transform's multi-scale decomposition capabilities, adaptively separating multi-scale components of the signal. Its depthwise separable structure and residual connection design ensure efficient and stable feature extraction while preserving fault details. This enables it to handle high-noise and multi-condition environments in tasks such as bearing fault diagnosis, accurately capturing early, minute fault features, providing highly discriminative feature inputs for subsequent modules, and ultimately achieving high-precision diagnosis under complex conditions.

[0035] Figure 4 The process of wavelet convolution can be summarized as follows:

[0036] Where X is the input, The low-frequency components of X correspond to the global features of X. These are the horizontal, vertical, and diagonal high-frequency components of X, respectively.

[0037] , , , Four wavelet filters, For convolution operations, For the transpose convolution operation, i is the current level of the cascaded wavelet decomposition and , For wavelet transform, Y is the output of the convolution. It is the i-th level depthwise convolution The weight tensor has four times the number of input channels as X. For inverse wavelet transform, This is the final output of the i-th level wavelet convolution.

[0038] Step 3: Input the features extracted by the deep wavelet convolutional network DTConv into the convolutional neural network CNN to extract deep features of the data, and use fully connected layers to generate graph node feature representations; Specifically, the multi-scale, time-frequency features extracted by the Deep Wavelet Convolutional Network (DTConv) are used as input and passed to a Convolutional Neural Network (CNN) to extract deeper features from the data. The convolutional layers of the CNN network focus on capturing patterns in local regions, gradually extracting more complex feature representations. The fully connected layers further process these features to generate more comprehensive graph node feature representations, providing more accurate input for subsequent classification tasks or deep learning models such as graph convolutional networks.

[0039] Step 4: Input the graph node features into the Adaptive Graph Generation Network (AGGNet) to dynamically construct a graph topology that adapts to the current working conditions. The AGGNet network improves upon the original graph generation network by introducing a multilayer perceptron, an adaptive adjacency construction mechanism, and a similarity-based edge weight generation strategy. This ensures that the generated graph structure remains consistent with changes in working conditions, thereby improving the feature propagation quality and modeling capabilities of the subsequent graph neural network.

[0040] The AGGNet network dynamically selects the optimal number of neighbors for each node based on the cosine similarity of node features, and constructs a dedicated graph topology for different operating conditions, combined with... Figure 5 The details are as follows: The AGGNet network first uses a multilayer perceptron to perform a nonlinear mapping on the features of the input graph nodes to obtain a more expressive representation of node attributes. This attribute matrix can more fully characterize the potential relationships between nodes, providing a foundation for subsequent dynamic graph construction. Subsequently, through a similarity measurement mechanism based on node attributes, the association strength between nodes is adaptively calculated, and the similarity is normalized to enhance the stability and comparability of the graph structure under different operating conditions.

[0041] After similarity evaluation, the AGGNet network introduces an adaptive selection strategy with a variable number of neighbors. Based on the statistical characteristics of similarity between nodes, it generates a dynamic number of neighbors for each node to avoid redundant connections and information dilution caused by a fixed neighborhood size. Through a top-k ranking mechanism, it selects several of the most representative neighbor nodes for each node from all candidate nodes, ensuring that the generated adjacency relationships closely reflect the changing trends of node feature distribution under different operating conditions, thereby guaranteeing the targeted nature of the graph structure construction process.

[0042] After determining adjacent nodes, the AGGNet network further generates corresponding edge weights based on the similarity between nodes. The adaptive assignment of edge weights effectively enhances the importance of key connections while suppressing the influence of weakly related edges on feature propagation, ensuring that the constructed graph maintains structural effectiveness and consistent feature representation even under high noise or multi-condition perturbations. Ultimately, through this dynamic topology construction strategy, the AGGNet network can generate targeted graph structures in different operating environments, enabling the model to maintain good generalization and robustness under cross-condition and variable operating conditions.

[0043] Thanks to this task-adaptive graph construction method, the information interaction relationships between nodes are effectively reconstructed, providing high-quality topological input for deep feature extraction in subsequent graph convolutional networks. This method effectively alleviates the problem of insufficient adaptability of traditional static graph structures under complex conditions, thereby improving the robustness and accuracy of the entire fault diagnosis model under challenging conditions.

[0044] Figure 4 The core mechanism of AGGNet can be summarized as follows:

[0045] in, This is a matrix of node attributes. This is a multilayer perceptron operation. The node cosine similarity matrix The normalized cosine similarity matrix of the nodes. The adaptive number of neighbors for each node. B For batch sizes, , , The number of neighbors corresponding to each node.

[0046] Step 5: Input the topology map adapted to the current working conditions, constructed by the Adaptive Graph Generation Network (AGGNet), into the Multi-Receptive Field Graph Convolutional Network (MRF-GCN) for multi-scale graph feature extraction. The MRF-GCN network performs multi-scale feature extraction by setting different receptive field ranges and performs information fusion of data structures to achieve excellent feature representation. Specifically, the key advantage of the MRF-GCN network lies in its ability to aggregate features at different scales within the graph structure, thereby capturing detailed information that is difficult to capture in traditional graph convolutional networks. Compared to standard graph convolutional networks, MRF-GCN, by setting different receptive field ranges, can better adapt to various features and information levels when handling complex situations, avoiding the information loss caused by a fixed receptive field. The specific process is as follows: In the MRF-GCN network, convolutional operations with different receptive fields are first used to extract local, multi-scale features at each node of the graph structure. These features at different scales can capture local details and the relationships between the global structure, thus ensuring the diversity and richness of information. Through this multi-receptive field feature extraction method, the MRF-GCN network not only enhances the model's ability to capture local features but also effectively improves the transmission and fusion of global information, ensuring more comprehensive and detailed information exchange between nodes.

[0047] Furthermore, the MRF-GCN network employs a data structure fusion strategy during graph convolution, achieving more accurate and robust feature representations by fusing features at different scales. This feature fusion strategy effectively mitigates conflicts between features of different scales during transmission, optimizes information flow during graph convolution, and generates higher-quality feature representations. These optimized feature representations further drive the training of the graph neural network, enhancing the model's robustness and generalization ability under complex conditions.

[0048] In this process, the MRF-GCN network fully leverages the receptive field's ability to aggregate information at different scales. Through multi-level information extraction and fusion, it maximizes the preservation of potentially useful information within the graph structure while suppressing interference from redundant information. This multi-receptive-field-based graph convolutional network provides powerful support for the Adaptive Graph Generation Network (AGGNet), enabling the fault diagnosis model to maintain excellent performance even under various complex operating conditions, greatly enhancing the model's practical application value.

[0049] Step Six: Input the extracted features into the classifier, structure aligner, and domain discriminator (Adaptive Adversarial Network AANet) to perform classification, structure alignment, and domain discrimination tasks, respectively. In particular, the AANet network plays a key role in cross-domain fault diagnosis tasks. It can adaptively adjust the adversarial strength according to the domain differences between the source and target domains, maintain a stable adversarial balance between the feature extractor and the domain discriminator under complex conditions, and thus effectively improve the cross-domain alignment effect.

[0050] Traditional adversarial networks have achieved some success in handling cross-condition tasks, but they typically rely on a fixed rate of change as the adversarial strength coefficient. This fixed mechanism struggles to cope with the constantly changing domain differences in real-world industrial environments. Specifically, when the difference between the source and target domains is small, excessively strong adversarial strength weakens the discriminative power of features, leading to a decline in classification performance; conversely, when the domain differences are large, insufficient adversarial strength results in inadequate domain alignment, causing domain shift problems. Therefore, the fixed rate of change in traditional adversarial training limits the balance of the adversarial process between the feature extractor and the domain discriminator, affecting the model's stability and generalization ability under complex conditions.

[0051] To address this issue, this embodiment proposes AANet (Adaptive Adversarial Network), which dynamically adjusts the adversarial strength based on real-time calculated domain differences, thereby optimizing the model training process. AANet adjusts the adversarial strength under different conditions through an adaptive mechanism, ensuring a dynamic balance between the feature extractor and the domain discriminator, and improving the model's generalization ability and stability.

[0052] First, the AANet network calculates the initial adversarial strength. The calculation formula is as follows:

[0053] in, To counter the intensity, To counter the strength limit, To counteract the lower limit of intensity, For the steepness of the curve, This represents the current iteration number. This represents the maximum number of iterations.

[0054] Based on the initial adversarial strength, the AANet network adaptively adjusts the adversarial strength according to the accuracy of the domain discriminator to ensure a stable adversarial balance between the feature extractor and the domain discriminator. Specifically, when the accuracy of the domain discriminator is high (i.e., it can easily distinguish between the source and target domains), indicating a large domain difference, AANet reduces the adversarial strength to avoid excessive feature mixing; conversely, when the accuracy of the domain discriminator is low (i.e., the domain difference is small), AANet increases the adversarial strength to promote more adequate domain alignment. The specific process is as follows:

[0055] in, To adjust the step size, For the accuracy of the domain discriminator, This is the set threshold value. This mechanism allows the AANet network to dynamically adjust the adversarial strength based on actual domain differences, thereby more effectively learning domain-invariant features and improving the model's adversarial training performance.

[0056] Throughout the training process, the parameters of the AANet network are updated using the backpropagation (BP) algorithm, while the parameters of the feature extractor, label classifier, and domain discriminator are optimized using gradient descent. The parameter update formula for the AANet network is as follows:

[0057]

[0058] in, These represent the parameters of the feature extractor and the domain discriminator in the AANet network, respectively. For the total loss function, The learning rate is used. During training, the AANet network can adjust the adversarial strength in real time to maintain the balance between the feature extractor and the domain discriminator, thus making the model more stable and accurate in cross-condition diagnostic tasks.

[0059] By introducing the AANet network, the adversarial training between the feature extractor and the domain discriminator becomes more flexible and efficient, enabling it to adapt to domain differences under dynamically changing working conditions. This improves the model's generalization ability and cross-domain alignment, providing strong support for bearing fault diagnosis tasks under complex working conditions.

[0060] Step 7: Deploy the trained model to the production equipment to perform fault detection and identification on the real-time collected vibration signals, and dynamically optimize the identification performance in cross-domain situations by combining graph neural network mechanism.

[0061] To verify the fault diagnosis performance of the method proposed in this embodiment, and to comprehensively evaluate the classification capability and feature distribution of the diagnostic system, such as... Figure 6 , Figure 7 , Figure 8 and Figure 9 As shown, the most difficult task among the 12 cross-domain tasks in CWRU is introduced. Q The most difficult of the 30 cross-domain missions and HUST's 6 missions. Q The model prediction results are analyzed using a confusion matrix of 20 and t-SNE feature visualization techniques. Figure 6 In the CWRU dataset confusion matrix shown in the image, DWA-DAGCN ( Figure 6 (Left) Compared to DAGCN ( Figure 6 (Right) Overall accuracy has been comprehensively improved, with DWA-DAGCN achieving an overall diagnostic accuracy of 94.22%, while DAGCN's overall accuracy was only 90.00%. Figure 7 In the HUST dataset confusion matrix shown in the image, DWA-DAGCN ( Figure 7 Left) for DAGCN ( Figure 7 The proposed method (right) has undergone comprehensive optimization, achieving an overall diagnostic accuracy of 88.16% for DWA-DAGCN and only 84.60% for DAGCN. This demonstrates that the proposed method has high diagnostic accuracy and good generalization ability. Figure 8 and Figure 9 The t-SNE feature distributions of DWA-DAGCN and DAGCN on the CWRU and HUST datasets are presented respectively. The domain alignment t-SNE plots show that the scatter points of the source and target domains of DWA-DAGCN have significantly higher overlap compared to DAGCN, indicating that DWA-DAGCN has superior domain alignment capabilities. The classification t-SNE plots show that the categories of DWA-DAGCN have clearer classification boundaries than DAGCN, demonstrating that the proposed method has high diagnostic accuracy and good generalization ability. Comparative experiments were conducted on the CWRU and HUST datasets. The comparison methods included Baseline (DAGCN), CORAL, ManbaLG, DeepGD, FANDA, and BDC. The experimental results are shown in Tables 1 and 2, demonstrating that the proposed method outperforms the aforementioned existing methods in both diagnostic accuracy and generalization ability.

[0062] Table 1. Comparative experimental results based on the CWRU dataset

[0063] Table 2. Comparative experimental results based on the HUST dataset

Claims

1. A cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network, characterized in that, Includes the following steps: Step 1: Acquire bearing vibration signals and preprocess them; Step 2: Input the preprocessed vibration signal into a deep wavelet convolutional network for multi-scale time-frequency feature extraction; Step 3: Input the multi-scale time-frequency features extracted by the deep wavelet convolutional network into the convolutional neural network to extract the deep features of the data, and generate graph node feature representations with the help of fully connected layers; Step 4: Input the graph node features into the adaptive graph generation network, use a multilayer perceptron to perform nonlinear mapping on the input graph node features to obtain node attribute representations, dynamically select the optimal number of neighbors for each node based on the cosine similarity of node features, and construct a dedicated graph topology structure for different working conditions. Step 5: Input the graph structure constructed by the adaptive graph generation network into the multi-receptive field graph convolutional network to perform multi-scale graph feature extraction; Step 6: Input the extracted features into the classifier, structure aligner, and domain discriminator respectively to perform the classification task, structure alignment task, and domain discrimination task; Fault detection and identification are performed on the collected vibration signals.

2. The cross-condition mechanical fault diagnosis method based on deep wavelet adaptive graph convolutional network according to claim 1, characterized in that, The preprocessing in step one includes segmenting the continuous vibration signal from the signal start point, with sampling points of fixed window length as the step size. No overlapping mechanism is used in the segmentation process to ensure that the segmented signals are completely independent.

3. A cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 1 or 2, characterized in that, The deep wavelet convolutional network includes a deep convolution module, a wavelet convolution module, and a pointwise convolution module. The wavelet convolution module extracts multi-scale features from the vibration signal through the multi-scale decomposition capability of wavelet transform. The multi-scale features extracted by the wavelet convolution module are fused with the channelwise convolution features extracted by the deep convolution module and then input into the pointwise convolution module for cross-channel information fusion, mapping the multi-channel features to the target channel dimension to obtain multi-scale time-frequency features.

4. The cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 3, characterized in that, The wavelet convolution module constructs wavelet decomposition filters and reconstruction filters based on a specified wavelet type to perform two-dimensional wavelet decomposition on the input signal. During forward propagation, the input signal is first decomposed into multi-scale wavelet components, with each layer generating a set of low-frequency and high-frequency components at the current scale. Then, the features are enhanced through convolutional layers, and the contribution weights of features at different scales are adjusted by scaling. Subsequently, the multi-scale processed components are reconstructed through inverse wavelet transform, and then added to the output of the basic depth convolution via residual connection to extract multi-scale features.

5. The cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 3, characterized in that, The adaptive graph generation network in step four adaptively calculates the association strength between nodes through a similarity measurement mechanism based on node attributes, normalizes the similarity, and generates a dynamic number of neighbors for each node based on the statistical characteristics of the similarity between nodes.

6. The cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 5, characterized in that, The adaptive graph generation network selects several of the most representative neighbor nodes for each node from all candidate nodes through a top-k sorting mechanism, and generates corresponding edge weights based on the similarity between nodes, so that the generated graph structure can remain consistent with changes in working conditions.

7. The cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 3, characterized in that, The multi-receptive-field graph convolutional network in step five extracts local multi-scale features at each node of the graph structure by setting convolution operations with different receptive field ranges, and performs information fusion of the data structure. By fusing features at different scales, multi-scale graph feature extraction is performed.

8. The cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 3, characterized in that, The domain discriminator in step six employs an adaptive adversarial network, which dynamically adjusts the adversarial strength based on the domain differences between the source and target domains.

9. A cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 8, characterized in that, The adaptive adversarial network adjusts the adversarial strength under different operating conditions through an adaptive mechanism. First, the initial adversarial strength is calculated. Based on the initial confrontation strength The adaptive adversarial network adaptively adjusts the adversarial strength based on the accuracy of the domain discriminator. The specific process is as follows: in, To adjust the step size, For the accuracy of the domain discriminator, The set threshold value.

10. A cross-condition mechanical fault diagnosis method based on a deep wavelet adaptive graph convolutional network according to claim 9, characterized in that, Initial combat strength The calculation formula is as follows: in, In order to counter the intensity, To counter the upper limit of strength, To counteract the lower limit of intensity, For the steepness of the curve, This represents the current iteration number. This represents the maximum number of iterations.

Images