Rotary machinery noise robust fault diagnosis method based on mixed domain graph neural network

Abstract

The invention requests to protect a rotating machine noise robust fault diagnosis method based on a mixed domain graph neural network. Comprising the following steps: collecting vibration data, and dividing an original vibration signal into a plurality of sections of subsamples with the length of 1024 by using non-overlapping sampling; converting the constructed subsample signal vibration signal into a frequency domain signal by using fast Fourier transform (FFT); taking the subsample frequency domain signals as nodes, the frequency spectrums as node features and the fault types as node labels, and then constructing the nodes into a graph structure; dividing the generated graph structure data set into a training set and a test set; and performing fault diagnosis on a test set by using the trained model, realizing self-adaptive weight adjustment by using a gating mechanism to fuse mechanical features, outputting a diagnosis result, calculating a performance index and performing result visualization analysis. The result shows that compared with a single-domain graph neural network, the method has higher accuracy and noise immunity, and end-to-end rotating machine fault diagnosis under the strong noise condition can be effectively achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery, and in particular to a rotating machinery fault diagnosis method based on hybrid domain adaptive weighted fusion. Background Art

[0002] As mechanical components that bear loads and transmit motion, rotating machinery is widely used in various fields such as machinery, electric power, aerospace, etc. However, due to its importance and the characteristics of being prone to failure, rotating machinery faults have become one of the main causes of equipment operation accidents in China, accounting for about 30% of the total number of accidents. The health state of rotating machinery is closely related to whether the machinery can operate normally. Therefore, it is very important to detect and diagnose rotating machinery faults and the fault locations.

[0003] The fault diagnosis of rotating machinery has been studied by many scholars at home and abroad, and many methods have been proposed. In traditional bearing fault diagnosis, signal processing technologies such as wavelet transform and empirical mode decomposition are used for the collected vibration signals to obtain fault signal characteristics. However, these methods only process the signals and do not obtain exact fault diagnosis results. Although deep learning has achieved great success in the field of fault diagnosis, among which convolutional neural networks are used more. Convolutional neural networks (CNNs) have translational invariance and locality and can effectively process Euclidean data such as image signals. However, it is very difficult to use CNNs to process non-Euclidean data such as graph-structured data. Therefore, a new field - graph neural networks (GNNs) has emerged. Graph data is an irregular data based on non-Euclidean space. The graph data composed of nodes and edges connecting the nodes has a more comprehensive information expression ability. In the field of mechanical fault diagnosis, the internal relationships contained in the data have not been given due attention in the diagnosis model, and there are still some deficiencies in the diagnosis of rotating machinery. For deep learning, the deeper the network structure, the more features can be extracted, but at the same time, the more training parameters are brought, and there will be problems of gradient explosion and disappearance during training. At the same time, it is impossible to accurately diagnose signals with noise, while the graph convolutional neural network structure can extract features well without too many network layer structures.

[0004] After retrieval, the application publication number is CN113505817A, a method for adaptively weighted training of bearing fault diagnosis model samples under unbalanced data, including the following steps: S1. Obtain training samples; S2. Construct a bearing fault diagnosis model; S3. Input the training samples into the bearing fault diagnosis model to extract features of each training sample; S4. Predict the labels of each training sample according to the features of each training sample; S5. Calculate the error between the predicted label and the true label, and the weights of each training sample; S6. Perform weighted backpropagation on the error according to the weights of each training sample to update the model; S7. Repeat steps S3 - S6 until the bearing fault diagnosis model converges or reaches the specified number of iteration steps. This solution realizes the adaptive weighted training of samples in the process of training the bearing intelligent fault diagnosis model under unbalanced data, enables the training process to focus on the unconverged samples, weakens the dominance of the majority class in the unbalanced data during training, and improves the performance of the intelligent diagnosis model under unbalanced data.

[0005] This patent mainly focuses on the processing of unbalanced data, but does not elaborate on how the model generalizes to different types of bearings and fault modes; the convolution neural network used in this patent, which consists of multiple convolution layers and pooling layers, to extract data features, results in an increase in the training time cost; there is no discussion in this patent about the robustness of the model in a strong noise environment, and in the actual industrial environment, sensor data often contains a large amount of noise. There is a lack of comparative experiments with other existing methods in this patent.

[0006] The present invention verifies the generalization of the model for different data types through the CWRU bearing dataset and the Jiangnan University bearing dataset; in the model construction of the present invention, both the spectrum - domain GCN and the spatial - domain GAT only use two - layer networks for feature extraction, reducing the time cost; the present invention designs a hybrid feature extraction method based on parallel spectrum - domain graph neural network and spatial - domain graph neural network. By extracting features in the spectrum and spatial neighborhood information and taking advantage of their respective advantages to complement each other. The spectrum - domain GCN effectively suppresses high - frequency noise through frequency - domain filtering, and the spatial - domain GAT sensitively captures local changes through the interaction between nodes and their neighborhoods. This hybrid feature extraction mechanism shows stronger robustness in a strong noise environment. The hybrid - domain model in the present invention realizes comparative experiments with other single - domain models, and the results show that the hybrid - domain model is superior to the single - domain models. Summary of the Invention

[0007] In order to solve the problems existing in the existing fault diagnosis of rotating machinery, such as low accuracy, high training difficulty, and inability to maintain good anti-noise performance and generalization performance under different strong noise conditions, the present invention provides a fault diagnosis method based on hybrid-domain adaptive weighted fusion. The designed HD-GNN model of this method takes graph-structured data as input, and has achieved the highest fault diagnosis accuracy rate in the rotating machinery dataset. Moreover, the performance advantage of the model increases continuously with the increase of the noise intensity, and still can maintain a high accuracy rate. A noise-robust fault diagnosis method for rotating machinery based on hybrid-domain graph neural network is proposed. The technical solution of the present invention is as follows:

[0008] A noise-robust fault diagnosis method for rotating machinery based on hybrid-domain graph neural network, which comprises the following steps:

[0009] Step 1: Collect vibration data, and use non-overlapping sampling to divide the original vibration signal into multiple sub-samples with a length of 1024;

[0010] Step 2: Use the fast Fourier transform FFT to convert the sub-sample signal vibration signal constructed in Step 1 into a frequency-domain signal;

[0011] Step 3: Take the sub-sample frequency-domain signal as a node, the spectrum as a node feature, and the fault type as a node label, and then construct these nodes into a graph structure;

[0012] Step 4: Divide the graph structure dataset generated in Step 3 into a training set and a test set;

[0013] Step 5; Use the trained model to perform fault diagnosis on the test set, use a gating mechanism to achieve adaptive adjustment of weights for mechanical feature fusion, output the diagnosis result, calculate the performance index and perform result visualization analysis.

[0014] Further, in Step 2, using the fast Fourier transform FFT to convert the sub-sample signal vibration signal constructed in Step 1 into a frequency-domain signal specifically includes:

[0015] The mathematical model description of the fast Fourier transform is that in graph spectrum theory, the Laplacian matrix of a graph is defined as:

[0016] L = D - A

[0017] where A is the adjacency matrix of the graph, and A ∈ R n*n , D is the diagonal matrix of the graph and where the i in D ii represents the element on the diagonal in the degree matrix. j represents the number of columns in the adjacency matrix A, and n represents the dimension of the matrix R. A graph can be divided into an undirected graph and a directed graph. For an undirected graph, A ij represents that an edge connects node v i and vj , while in a directed graph, A ij represents v i pointing to v j ; for the construction of the adjacency matrix, a 0-1 weighting method is used, that is, it is 1 when there is an edge connection between nodes and 0 when there is no edge connection;

[0018] Finally, the normalized Laplacian matrix L is defined as:

[0019] L = I n - D -1 / 2 AD -1 / 2

[0020] where, I n is the identity matrix. Performing eigenvalue decomposition on the Laplacian matrix L, we have:

[0021] Lu i = λ i u i

[0022] where: λ i is the eigenvalue of L, and u i is the eigenvector corresponding to λ i of L, and u T u = I, ‖u‖ = 1. Diagonalizing L, we have:

[0023] L = UΛU T

[0024] where: U = [u1, u2, u3,...., u n , Λ = diag([λ1,..., λ n ); U represents the eigenvector matrix of the Laplacian matrix L corresponding to the eigenvalue λ i , and u n is the eigenvector of L corresponding to λ i , and λ n corresponds to the eigenvalue of the Laplacian matrix L. Define the graph Fourier transform as: Its inverse transform is:

[0025] where, x is the vertex domain signal, is the spectral domain signal, that is, the essence of the graph Fourier transform is to use the eigenvectors of L as basis functions to transform the graph signal from the vertex domain to the frequency domain.

[0026] Furthermore, the mathematical models of the graph convolutional neural network and the graph attention network operations in step 2 are described as:

[0027] H (l+1) = σ(D -1 / 2 AD -1 / 2 H(l) W (l) )

[0028] Among them, H represents the node feature matrix, σ represents the non-linear activation function, and W represents the weight parameter matrix.

[0029]

[0030] Among them represents the parameters of the single-layer forward neural network, u ∈ N i represents the neighbor nodes of node i, W is the projection matrix, and α ij represents the attention coefficient between nodes, || represents the concatenation operation, and LeakReLU represents the non-linear activation function.

[0031] Furthermore, the mathematical model description of the gating mechanism operation in step 3 is as follows:

[0032] First, concatenate the node features in the spectral domain and the node features in the spatial domain. The formula is as follows:

[0033] f concat =[f spatial, f spectral

[0034] Among them, f spatial, f spectral represent the node features based on the spatial domain and the node features based on the spectral domain respectively, and f concat represents the feature after their concatenation;

[0035] The gating signal calculation is used to dynamically adjust the contribution degrees of different feature sources in the gating mechanism, and then the model can adaptively assign a weight value to each node. The formula is as follows;

[0036] g = σ(Wf concat + b)

[0037] Among them, W and b represent the weight matrix and bias matrix of each layer respectively, and g represents the finally calculated gating signal; this function is used to convert the input features into an appropriate representation to calculate an effective gating signal;

[0038] Finally, adaptively weighted fusion of features from different sources is performed using the adaptively assigned weight values. The formula is as follows;

[0039] f spatial_gated = g ⊙ f spatial

[0040] f spectral_gated = (1 - g) ⊙ f spectral

[0041] f fused ​= f spatial_gated + f spectral_gated

[0042] where f spatial_gated and f spectral_gated represent the node features after adaptive weighting, and f fused represents the new node features after weighted fusion.

[0043] Furthermore, the graph structure includes a k-nearest neighbor graph, a fully connected graph, a path graph, and an ER random graph, where k = 5 in the k-nearest neighbor graph and ER = 0.5 in the ER random graph.

[0044] Furthermore, the gating mechanism in step 5 optimizes the feature fusion based on the spectral domain and the spatial domain by dynamically adjusting the weights, so as to improve the robustness and diagnostic accuracy of the model in a strong noise environment.

[0045] Furthermore, the graph neural network (GCN) based on the spectral domain and the graph neural network (GAT) based on the spatial domain respectively use the Laplacian matrix and the attention coefficient mechanism for node feature extraction.

[0046] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the fault diagnosis method based on hybrid domain adaptive weighted fusion as described in any one of the above.

[0047] A non-transitory computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the fault diagnosis method based on hybrid domain adaptive weighted fusion as described in any one of the above.

[0048] A computer program product includes a computer program, and when the computer program is executed by a processor, it implements the fault diagnosis method based on hybrid domain adaptive weighted fusion as described in any one of the above.

[0049] The advantages and beneficial effects of the present invention are as follows:

[0050] The rotating machinery fault diagnosis method of the present invention introduces the adaptive weighted fusion of node features based on the spectral domain and the spatial domain. Under strong noise conditions, the graph neural networks (GNNs) in the spectral domain and the spatial domain have their own advantages. The spectral domain GNN uses frequency domain filtering to suppress high-frequency noise and capture global structure information, but the calculation is complex and it is less sensitive to local changes. The spatial domain GNN is sensitive to local information through node and neighborhood convolution operations, has high computational efficiency, and shows stronger robustness in a strong noise environment. Combining the two, the spectral domain GNN provides global structure support, and the spatial domain GNN enhances the response ability to local changes. The combination of the two can improve the generalization ability and robustness of the model in a complex noise environment.

[0051] The fault diagnosis experiment shows that the HD-GNN model of the present invention not only controls the training time within an acceptable range, but also has higher accuracy and stability. Moreover, the model has more obvious advantages in processing noisy data, and the algorithm performance continuously expands with the increase of noise intensity.

[0052] The innovation of the present invention lies in applying the Hybrid Domain Graph Neural Network (HD-GNN) to the fault diagnosis of rotating machinery. Specifically, the innovation is reflected in the following aspects:

[0053] The innovation of the present invention mainly includes the following steps:

[0054] 1. Hybrid domain feature extraction: Combine the graph convolutional network (GCN) based on the spectral domain and the graph attention network (GAT) based on the spatial domain to extract features from the spectral and spatial neighborhood information respectively.

[0055] 2. Adaptive weighted fusion mechanism: By introducing a gating mechanism, dynamically adjust the weights of the spectral domain and spatial domain features to achieve adaptive fusion.

[0056] 3. Flexible use of graph structures: Not only use a single graph structure, but also adopt a variety of graph structures including k-nearest neighbor graph, fully connected graph, path graph, and ER random graph.

[0057] End-to-end fault diagnosis: The entire method from vibration signal acquisition, preprocessing, feature extraction to fault diagnosis can be automatically completed without manual intervention.

[0058] The corresponding beneficial effects are as follows:

[0059] 1. In the aspect of rotating machinery fault diagnosis, compared with traditional fault diagnosis methods and single-domain graph neural networks, the HD-GNN model proposed by the present invention not only improves the diagnostic accuracy, but also performs particularly outstanding under strong noise conditions, ensuring the robustness and stability under complex working conditions.

[0060] 2. Since the model can automatically adjust the feature weights, it reduces the sensitivity to noise, makes the diagnostic results more reliable, helps to monitor and maintain the equipment in real time, reduces the downtime caused by faults, and lowers the maintenance cost.

[0061] 3. Through the innovative design of the hybrid domain graph neural network, the present invention provides a new solution for the fault diagnosis of rotating machinery, promoting the application of graph neural networks in the field of non-stationary data processing.

[0062] The ingenuity or difference lies in:

[0063] 1. Hybrid Domain Feature Extraction: Traditionally, signal processing and feature extraction methods in the spectral domain and spatial domain are separate. This invention combines the two for the first time, using spectral domain GCN to suppress high-frequency noise and spatial domain GAT to capture local changes. This combination shows stronger robustness in a strong noise environment.

[0064] 2. Adaptive Weighted Fusion Mechanism: By introducing a gating mechanism, dynamic adjustment of spectral domain and spatial domain features is achieved, enabling the model to automatically select the most favorable feature combination according to specific situations, improving the diagnostic accuracy and generalization ability of the model.

[0065] 3. Flexible Application of Graph Structure: In the field of rotating machinery fault diagnosis, this invention not only uses a single graph structure but also adopts multiple graph structures, providing the model with diverse information input methods and enhancing the adaptability and performance of the model.

[0066] 4. End-to-End Fault Diagnosis Process: The entire method requires no manual intervention and automatically completes the whole process from vibration signal acquisition to fault diagnosis, simplifying the fault diagnosis process and improving efficiency.

[0067] The following are two specific embodiments of this invention: Brief Description of the Drawings

[0068] Figure 1 is the model structure diagram of the preferred embodiment provided by this invention;

[0069] Figure 2 is the model flow chart;

[0070] Figure 3 is the comparison diagram of T-SNE and confusion matrix of the model on the CWRU dataset;

[0071] Figure 4 is the comparison diagram of classification accuracy of the model under strong noise conditions of -20dB, -15dB, -10dB, -5dB for the CWRU data.

[0072] Figure 5 is the comparison diagram of T-SNE and confusion matrix of the model on the JNU dataset;

[0073] Figure 6 The data is JNU, and it is the comparison diagram of classification accuracy of the model under strong noise conditions of -20dB, -15dB, -10dB, -5dB. Detailed Description of the Embodiment

[0074] Next, the technical solutions in the embodiments of this invention will be clearly and detailedly described in conjunction with the drawings in the embodiments of this invention. The described embodiments are only a part of the embodiments of this invention.

[0075] The technical solution of the present invention to solve the above technical problems is as follows:

[0076] As Figure 1 shown, a fault diagnosis method based on hybrid-domain adaptive weighted fusion in a specific embodiment includes:

[0077] Step 1: Collect vibration data, and use non-overlapping sampling to divide the original vibration signal into multiple sub-samples with a length of 1024.

[0078] Step 2: Use the fast Fourier transform (FFT) to convert the vibration signal of the sub-sample signal constructed in Step 1 into a frequency-domain signal.

[0079] Step 3: Take the sub-sample frequency-domain signal as a node, the spectrum as a node feature, and the fault type as a node label, and then construct these nodes into a graph structure. Here, four graph structures are selected as the input of the model, namely the k-nearest neighbor graph, the fully connected graph, the path graph, and the ER random graph. In the k-nearest neighbor graph, set k = 5, and in the ER random graph, ER = 0.5. There is an undirected edge connecting any two nodes in the four graph structures, and the weight of the edge is weighted by 0-1, that is, when there is an edge connection between two nodes, the weight is 1, otherwise it is 0.

[0080] Step 4: Divide the graph structure data set generated in Step 3 into a training set and a test set according to a ratio of 6:4.

[0081] Step 5: Use the trained model to perform fault diagnosis on the test set, use the gating mechanism to achieve adaptive adjustment of weights for mechanical feature fusion, and then use Softmax for node classification to output the diagnosis result. And visualize the result.

[0082] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0083] In an embodiment of the present invention, the fault diagnosis method based on hybrid-domain adaptive weighted fusion is verified. The following will give a possible embodiment to non-restrictively elaborate on its specific implementation scheme.

[0084] The data set involved in this embodiment uses the open bearing database of Case Western Reserve University and the bearing data set of Jiangnan University to verify the performance of HD-GNN.

[0085] The CWRU dataset acquisition system is driven by a three-phase induction motor, and the vibration signals of the faulty bearings are recorded through two accelerometers. The motor loads of the system are 0HP, 1HP, 2HP, and 3HP respectively, the rotational speed is 1772 r / min, and data is collected at a sampling frequency of 12 KHz. Faults are formed by introducing single-point defects on the inner ring, outer ring, and rolling elements of the bearing respectively through electrical discharge machining technology. The damage sizes at each location are 7 inch, 14 inch, and 21 inch respectively. Together with the normal state, 10 different state data samples are generated, as shown in Table 1. The constructed graph structure data is input into the network model proposed in this paper.

[0086] Table 1 CWRU Fault Data Description

[0087]

[0088]

[0089] To verify the superiority of the proposed model, the influence of different graph structures on the diagnostic results under different models was compared and analyzed in a noise-free environment and a strong-noise environment. The comparison of the results in the noise-free environment is shown in Table 2. Under the condition of no noise, the accuracy of the models is close to 100%, proving that the model in this paper also has good robustness in a noise-free environment. The signal-to-noise ratios set for each model are -20 dB, -15 dB, -10 dB, and -5 dB respectively, where dB represents the unit of SNR, P signal represents the signal power, and P noise represents the noise power. The formula is as follows.

[0090] SNR(dB) = 10log 10 (P signal / P noise )

[0091] Table 2 CWRU No Noise

[0092]

[0093] Under the same conditions where the graph structure is a fully connected graph and the noise environment is -15 dB, the T-SNE visualization and confusion matrix of the final classification results of each model are respectively as Figure 3 shown. Under the same circumstances, the model proposed in this paper has a higher classification accuracy compared to other single-domain models, and its noise resistance and robustness are better than those of single-domain models.

[0094] The bearing fault dataset provided by Jiangnan University is used to evaluate and verify the generalization ability and performance advantages of the proposed fault diagnosis method. This dataset collects vibration signals from bearings with a high sampling frequency of 50 kHz, ensuring that the obtained signals have sufficient details and accuracy. An experimental dataset is collected, which contains vibration signals from three different motor speeds, simulating various operating conditions of bearings in actual applications. This dataset covers four states of bearings, namely normal state, inner race fault, outer race fault, and rolling element fault. See Table 3 for details.

[0095] Table 3 JNU Fault Data Description

[0096]

[0097] The result comparison under the noise-free environment is shown in Table 4. Under the noise-free condition, the accuracy of the model is close to 100%, proving that the model in this paper also has good robustness in the noise-free environment.

[0098] Table 4 JNU Noise-free

[0099]

[0100] To reduce the influence of randomness and ensure the fairness of the comparative experiment, the experiments of all models are conducted 10 times, and the final evaluation index is the average value. To verify the noise resistance and robustness of the model under strong noise conditions, this paper conducts comparative experiments in the noise environments with signal-to-noise ratios of -5dB, -10dB, -15dB, and -20dB. It can be seen from Figure 4 that the model designed in this paper is superior to the other four models under strong noise conditions and in the four graph structures, showing better noise resistance and robustness.

[0101] The following are two specific embodiments of the present invention:

[0102] Embodiment 1: Verification of the HD-GNN Model Based on the CWRU Dataset

[0103] Implementation Environment and Data Preparation

[0104] - Dataset Source: This embodiment uses the open bearing database of Case Western Reserve University (CWRU). This dataset records the bearing vibration signals under different fault states, including inner race defects, outer race defects, rolling element defects, and normal states. Each state has three different damage sizes.

[0105] - Hardware and Software Environment: The experiment was run on a server equipped with an NVIDIA GeForce RTX 2080Ti GPU, using the Python programming language, with the deep learning framework being PyTorch and the library related to graph neural networks being DGL (DeepGraph Library).

[0106] Implementation Steps

[0107] 1. Data Collection and Preprocessing: Data was collected from the CWRU dataset at a sampling rate of 12KHz, and then the original vibration signals were divided into sub-samples of length 1024 through non-overlapping sampling. Next, the vibration signals were converted into frequency-domain signals using the Fast Fourier Transform (FFT).

[0108] 2. Graph Structure Construction: The frequency-domain signals were converted into graph structures, where the frequency-domain signals served as nodes, the spectra served as node features, and the fault types served as node labels. Four graph structures were constructed: k-nearest neighbor graph (k = 5), fully connected graph, path graph, and ER random graph (ER = 0.5). All edges in the graph structures were undirected edges, and the weights were weighted with 0 - 1.

[0109] 3. Model Training and Testing:

[0110] - Model Design: An HD-GNN model was constructed, which included a spectral-domain graph convolutional neural network (GCN) and a spatial-domain graph attention network (GAT), and a gating mechanism was set between them for feature fusion.

[0111] - Data Partitioning: The constructed graph structure dataset was partitioned into a training set and a test set in a ratio of 6:4.

[0112] - Model Training: The HD-GNN model was trained using the training set. After feature extraction, adaptive weighted fusion was achieved through the gating mechanism, and finally, a Softmax layer was used for node classification.

[0113] 4. Performance Evaluation: The model was tested in an environment without noise and with different noise intensities (-5dB, -10dB, -15dB, -20dB). The performance and robustness of the model were evaluated by calculating the classification accuracy, confusion matrix, and T-SNE visualization analysis.

[0114] Experimental Results

[0115] - In an environment without noise, the accuracy of the model was close to 100%, proving that HD-GNN has excellent recognition ability under ideal conditions.

[0116] - In a -15dB noise environment, compared with other single-domain models, the classification accuracy of HD-GNN was significantly higher, verifying its noise resistance and robustness.

[0117] - Under different noise intensities, the average classification accuracy of HD-GNN is better than that of other models. Especially under the strong noise condition of -20dB, the performance advantage is more significant.

[0118] Example 2: Application of HD-GNN model on the bearing database of Jiangnan University

[0119] Implementation environment and data preparation

[0120] - Data set source: In this example, the bearing fault database of Jiangnan University is used. Different from the CWRU data set, this database may contain more complex equipment configurations and working conditions, and the data noise characteristics are also different. - Hardware and software environment: The same as in Example 1, the experiment runs on the same server, using Python and PyTorch frameworks, and using the DGL library to construct the graph neural network.

[0121] Implementation steps

[0122] 1. Data collection and preprocessing: Collect vibration data, use the same preprocessing steps as CWRU, convert the data into frequency domain signals, and divide them into sub-samples with a length of 1024.

[0123] 2. Graph structure construction: Adopt the same graph structure construction method as in Example 1 to ensure the consistency of the model input.

[0124] 3. Model application and performance evaluation:

[0125] - Model application: Directly apply the HD-GNN model trained on the CWRU data set, omit the model training step, and directly perform fault diagnosis on the test set of the bearing database of Jiangnan University.

[0126] - Performance evaluation: Evaluate the generalization ability of the model on the Jiangnan University data set, and verify the robustness and generalization performance of the model on different data sets by calculating the classification accuracy and comparing with the CWRU data set.

[0127] Experimental results

[0128] - On the Jiangnan University data set, although the working conditions and noise characteristics are different, the HD-GNN model can still maintain a high classification accuracy, proving its generalization ability on different data sets.

[0129] - In a noise-free environment, the accuracy of the model is close to 100%, proving that HD-GNN has excellent recognition ability under ideal conditions.

[0130] - In a -15 dB noise environment, compared with other single-domain models, the classification accuracy of HD-GNN is significantly higher, and its noise resistance and robustness are verified.

[0131] - Under different noise intensities, the average classification accuracy of HD-GNN is better than that of other models. Especially under the strong noise condition of -20 dB, the performance advantage is more significant.

[0132] - Compared with the CWRU dataset, the performance difference of the model on the Jiangnan University dataset is not significant, indicating that the hybrid-domain graph neural network has stable robustness in processing non-stationary vibration signals.

[0133] The systems, devices, modules or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions.

[0134] Computer-readable media includes both permanent and non-permanent, removable and non-removable media and can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0135] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or also includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising the element.

[0136] The above embodiments should be understood as only illustrative of the present invention and not restrictive of the scope of protection of the present invention. After reading the content described in the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for robust fault diagnosis of rotating machinery noise based on hybrid domain graph neural network, characterized in that: The following steps are involved: Step 1: Collect vibration data and divide the original vibration signal into multiple sub-samples with a length of 1024 using non-overlapping sampling; Step 2: Use fast Fourier transform (FFT) to convert the sub-sample signal vibration signal constructed in step 1 into a frequency domain signal; Step 3: Take the sub-sample frequency domain signal as the node, the spectrum as the node feature, and the fault type as the node label, and then construct these nodes into a graph structure; Step 4: Divide the graph structure dataset generated in step 3 into a training set and a test set; Step 5: Use the trained model to perform fault diagnosis on the test set, use the gating mechanism to adaptively adjust the weights to mechanically fuse features, output the diagnosis results, calculate the performance indicators and perform visual analysis of the results.

2. The method for robust fault diagnosis of rotating machinery noise based on hybrid domain graph neural network according to claim 1 is characterized in that: The step 2 uses fast Fourier transform (FFT) to convert the sub-sample signal vibration signal constructed in step 1 into a frequency domain signal, specifically comprising: The mathematical model of the fast Fourier transform is described as follows: In graph spectrum theory, the Laplacian matrix of a graph is defined as: L=DA Where A is the adjacency matrix of the graph, and A∈R n*n , D is the diagonal matrix of the graph and Where D ii The i in represents the diagonal elements in the degree matrix. j represents the number of columns in the adjacency matrix A, and n represents the dimension of the matrix R. Graphs can be divided into undirected graphs and directed graphs. For undirected graphs, A ij Represents an edge connecting nodes v i and v j , while in a directed graph, A ij Indicates v i Point to v j ; For the construction of the adjacency matrix, a 0-1 weighting method is used, that is, when there is an edge connection between nodes, it is 1, and when there is no edge connection, it is 0; Finally, the normalized Laplacian matrix L is defined as: L=I n -D -1 / 2 AD -1 / 2 Among them, I n is the unit matrix, and the Laplace matrix L is decomposed into the following features: Is i λ i the i Where: i is the eigenvalue of L, u i is the λ corresponding to L i The eigenvector of T u=I,‖u‖=1, diagonalize L, and we have: L=UΛU T Among them: U=[u1,u2,u3,....,u n ],Λ=diag([λ1,...,λ n ]); U represents the Laplace matrix L corresponding to the eigenvalue λ i The eigenvector matrix, u n is L corresponding to λ i The eigenvector of n Corresponding to the eigenvalue of the Laplace matrix L. The Fourier transform of the graph is defined as: Its inverse transform is: Among them, x is the vertex domain signal, It is a spectral domain signal, that is, the essence of the graph Fourier transform is to use the eigenvector of L as the basis function to transform the graph signal from the vertex domain to the frequency domain.

3. The method for robust fault diagnosis of rotating machinery noise based on hybrid domain graph neural network according to claim 2 is characterized in that: The mathematical model of the graph convolutional neural network and graph attention network operation in step 2 is described as: H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) ) Where H represents the node feature matrix, σ represents the nonlinear activation function, and W represents the weight parameter matrix. in Represents the parameters of a single-layer feedforward neural network, u∈N i represents the neighbor nodes of node i, W is the projection matrix, α ij represents the attention coefficient between nodes, || represents the cascade operation, and LeakReLU represents the nonlinear activation function.

4. The method for robust fault diagnosis of rotating machinery noise based on hybrid domain graph neural network according to claim 1 is characterized in that: The mathematical model of the gate control mechanism operation in step 3 is described as: (1) First, the node features in the spectral domain and the node features in the spatial domain are concatenated. The formula is as follows: f concat =[f spatial, f spectral ] where f spatial, f spectral They represent the node features based on the spatial domain and the node features based on the spectral domain respectively, and f concat Indicates the characteristics after splicing; (2) The gate signal calculation is used in the gating mechanism to dynamically adjust the contribution of different feature sources. Then the model can adaptively assign a weight value to each node. The formula is as follows; g=σ(Wf concat +b) Where W and b represent the weight matrix and bias matrix of each layer respectively, and g represents the final calculated gating signal; This function is used to convert the input features into appropriate representations to calculate effective gating signals; (3) Finally, the features of different sources are adaptively weighted and fused through adaptively assigned weight values. The formula is as follows: f spatial_gated =g⊙f sspatial f spectral_gated =(1-g)⊙f spectral f fused =f spatial_gated +f spectral_gated where f spatial_gated and f spectral_gated represents the node features after adaptive weighting, f fused Represents the new node features after weighted fusion.

5. The method for robust fault diagnosis of rotating machinery noise based on hybrid domain graph neural network according to claim 1 is characterized in that: The graph structure includes a k-nearest neighbor graph, a fully connected graph, a path graph and an ER random graph, wherein k=5 in the k-nearest neighbor graph and ER=0.5 in the ER random graph.

6. The method for robust fault diagnosis of rotating machinery noise based on hybrid domain graph neural network according to claim 1 is characterized in that: The gating mechanism of step 5 optimizes the fusion of spectral-domain and spatial-domain features by dynamically adjusting weights to improve the robustness and diagnostic accuracy of the model in a strong noise environment.

7. The method for robust fault diagnosis of rotating machinery noise based on hybrid domain graph neural network according to claim 1 is characterized in that: The spectral domain-based graph neural network (GCN) and the spatial domain-based graph neural network (GAT) respectively use the Laplacian matrix and attention coefficient mechanism to extract node features.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the rotating machinery noise robust fault diagnosis method based on the hybrid domain graph neural network as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for robust fault diagnosis of rotating machinery noise based on a hybrid domain graph neural network as described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for robust fault diagnosis of rotating machinery noise based on a hybrid domain graph neural network as described in any one of claims 1 to 7 is implemented.

Images