Intelligent diagnosis method for composite faults of rolling bearings based on deep graph convolutional network
By converting the vibration signal into graph structure data and using the graph convolution neural network model, the feature extraction problem in composite fault signals is solved, and high-precision rolling bearing fault diagnosis is achieved, which is suitable for a variety of working conditions and sample imbalance.
Patent Information
- Application Number
- CN202310374137.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-10
- Publication Date
- 2025-04-18
- Estimated Expiration
- 2043-04-10
AI Technical Summary
The prior art is difficult to effectively extract various fault characteristics from the composite fault signal, resulting in low accuracy in rolling bearing fault diagnosis, especially when processing non-stationary and nonlinear signals.
The method based on the depth graph convolution network is adopted to convert the original vibration signal into graph structure data, and the graph convolution neural network model is used to fully mine the characteristics of nodes and edges, thereby improving the accuracy of composite fault diagnosis.
The end-to-end diagnosis model is realized, which eliminates the complex feature engineering in traditional diagnostic methods, and can extract useful fault characteristics under the same working conditions, different working conditions and sample imbalance, effectively avoiding the wrong classification between a single fault and a compound fault.
Smart Images
Figure CN116662873B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an intelligent diagnosis method for composite faults of rolling bearings based on a deep graph convolutional network. Background Art
[0002] Rolling bearings are an important component of rotating machinery and are also the most easily damaged components. According to statistics, in rotating machinery using rolling bearings, about 30% of mechanical failures are related to bearings, and their operating status directly affects the overall performance of the equipment. In actual engineering, multiple bearing failures coexist to form composite failures. The vibration signal of the composite failure state is not a simple superposition of a single fault signal, but is coupled with the vibration signals of other components through a complex transmission path. The failures of different components interact and influence each other, making the characteristics of the composite fault signal non-stationary and nonlinear, which brings great difficulties to the determination and location of bearing faults. Therefore, how to extract the characteristics of each fault from the composite fault signal is the focus and difficulty of current research.
[0003] The most important step in the traditional rolling bearing fault diagnosis method is feature extraction, which aims to extract useful fault information from the signal, thereby improving the accuracy of fault diagnosis. Common feature extraction methods include principal component analysis (PCA), wavelet transform (WT), short-time Fourier transform (STFT), etc. At present, most fault diagnosis methods require manual feature selection when processing non-stationary and nonlinear signals. Different feature selection determines the effect of fault diagnosis; on the other hand, the algorithm performance of existing feature extraction methods will gradually decrease with the increase of data volume. In order to overcome the above shortcomings, deep learning theory has been introduced into the field of fault diagnosis. Compared with traditional shallow models, deep learning obtains health information from collected signals by constructing an "end-to-end" model that does not rely on prior knowledge, and can meet the high requirements for fault diagnosis accuracy in the "big data" era. However, traditional deep learning methods can only learn features from the vertices of the input data, while ignoring the information contained in the edges formed between vertices. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a rolling bearing composite fault intelligent diagnosis method based on a deep graph convolutional network, which converts the original vibration signal into graph structure data, and in the graph domain, utilizes the feature mining capability of the graph convolutional neural network model to fully mine the features of nodes and edges to improve the accuracy of composite fault diagnosis.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: an intelligent diagnosis method for composite faults of rolling bearings based on a deep graph convolutional network, comprising the following steps:
[0006] (1) Building graphs based on time series
[0007] The original time series X with a signal length of L is normalized, and the normalized data is expressed as:
[0008] X nol =normalization(X) (1)
[0009] Where X nol is the normalized time series, normalization(·) represents the normalization method;
[0010] Secondly, the original time series of the signal length L is divided into sub-samples of length d, with no overlap between each sample, and a corresponding label is assigned to each sub-sample. The obtained sub-sample set can be expressed as:
[0011]
[0012]
[0013] Where ∏ represents the constructed super sample set, represents subsamples, y1,…,y n represents the label, n represents the number of sub-sampling;
[0014] Perform fast Fourier transform FFT on each sub-sample and use the obtained spectrum as a new sample. The process can be expressed as:
[0015] Where FFT(·) means converting the subsamples into the frequency domain and taking the first half of the result;
[0016] The final labeled dataset is represented as:
[0017]
[0018] (2) Constructing graph structure data
[0019] Construct a weighted graph with the labeled dataset obtained above, and construct a graph with n nodes by treating each subsample as a node in the graph; estimate the distance between samples and define a threshold ε; if the cosine similarity is greater than the threshold, there will be an edge between the two nodes; the node x can be obtained by i Neighbor nodes:
[0020]
[0021] in yes The neighbor nodes of , ε is the selected radius, and the threshold ε is 0. Compute Node and The cosine similarity of
[0022] The weight between each two nodes is calculated through a threshold Gaussian kernel weight function:
[0023]
[0024] Where β is the bandwidth variance of the Gaussian function;
[0025] (3) Fault diagnosis structure
[0026] The processed data is put into the deep graph convolution model for processing, and the processing results are output; the deep graph convolution model uses momentum stochastic gradient descent SGD as the optimizer for hyperparameter optimization; the momentum of momentum stochastic gradient descent SGD is 0.9, the number of iterations epochs is 100, the batch size is 64, the learning rate decay strategy is used to adjust the learning rate, the weight decay value is initialized to 0.0005, the learning rate learn rate is 0.001, the number of network nodes is 640, and the number of edges is 5760; the deep graph convolution model includes an input layer, a first convolution layer, a first normalization layer, a second convolution layer, a second normalization layer, a fully connected layer, and an output layer;
[0027] In the first graph convolution layer, the input node feature map size is 512, the output node feature map size is 1024, and the number of graph convolution kernels is 512×1024;
[0028] The first normalization layer normalizes the data after the convolution operation. The number of channels of the output feature map is 1024, the step size is 1, and the ReLU activation function is used for nonlinear mapping;
[0029] In the second graph convolution layer, the input node feature map size is 1024, the output node feature map size is 1024, and the number of graph convolution kernels is 1024×1024;
[0030] The second normalization layer normalizes the data after the convolution operation. The number of channels of the output feature map is 1024, the step size is 1, and the ReLU activation function is used for nonlinear mapping.
[0031] In the fully connected layer, the input feature map size is 1024, the output feature map size is 512, and the activation function ReLU is used;
[0032] In the output layer, the input feature map size is 512, and the activation function Softmax is used.
[0033] As a preferred solution, in step 3, the processed data is placed in a deep graph convolution model and processed repeatedly for more than 5 times, and the average value of all output processing results is taken as the final result.
[0034] As a preferred solution, in step 2, cosine similarity is used to estimate the distance between samples.
[0035] As a preferred solution, in step 1, the original time series X with a Max-min normalized signal length of L is used.
[0036] As another preferred solution, in step 2, the distance between two nodes is calculated using the Euclidean algorithm:
[0037] in, and i1 ,y i2 Represents nodes. The closer the value of d is to 1, the higher the similarity.
[0038] As a preferred solution, in step 2, the distance between two nodes is calculated using the Manhattan distance algorithm:
[0039] in, and i1 ,y i2 Represents nodes. The closer the value of d is to 1, the higher the similarity.
[0040] As another preferred solution, in step 2, the distance between two nodes is calculated using the Chebyshev algorithm:
[0041] in, and i1 ,y i2 Represents nodes. The closer the value of d is to 1, the higher the similarity.
[0042] The beneficial effects of the present invention are:
[0043] (1) The DGCN model can realize an end-to-end diagnosis mode, eliminating the complex feature engineering in traditional diagnosis methods. The graph constructed using vertices and edges can provide more information for the training of the diagnosis model. The experimental results show that this method can extract useful fault features under the same working conditions, different working conditions and sample imbalance, and various fault types can be clearly distinguished. This method has great advantages in the treatment of rolling bearing composite faults under the same working conditions, different working conditions and sample imbalance.
[0044] (2) Aiming at the problem that the high correlation between compound fault and single fault samples can easily lead to classification errors, this method can effectively avoid the misclassification between single fault and compound fault. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is the composite fault diagnosis process of the present invention;
[0046] Figure 2 It is the visualization of the input layer, one convolution layer, and two convolution layers of the present invention;
[0047] Figure 3 It is a schematic diagram of the final diagnosis result of the embodiment of the present invention. DETAILED DESCRIPTION
[0048] The specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.
[0049] like Figure 1-3 As shown, a rolling bearing composite fault intelligent diagnosis method based on deep graph convolutional network includes the following steps:
[0050] 1 Data Preprocessing
[0051] The vibration signals collected from the XJTU-SY test bench, the test data set used is shown in Table 1, and 6 bearings under operating conditions are selected. First, the original time series X with the original signal length L (L is 1024000) is normalized to the maximum and minimum; the normalized data is expressed as:
[0052] X nol =normalization(X) (1)
[0053] Secondly, a sliding window is used to intercept non-overlapping vibration signals. Each sub-sample d contains 1024 data points, and a corresponding label is assigned to each sub-sample. The obtained sub-sample set can be expressed as:
[0054]
[0055] The number of subsamples n for each fault type is 1000, and the total number of samples for the eight fault types is 8000;
[0056] Perform fast Fourier transform FFT on each sub-sample and use the obtained spectrum as a new sample. The process can be expressed as:
[0057] Where FFT(·) means converting the subsamples into the frequency domain and taking the first half of the result;
[0058] The final labeled dataset is represented as:
[0059]
[0060] The division of training set and test set is shown in Table 2. The fault types of data set A_1 include normal (N_1), outer ring fault (OF_1), cage fault (CF) and inner and outer ring combined fault (OIF). The fault types of data set A_2 include normal (N_2), outer ring fault (OF_2), inner ring, rolling element and cage combined fault (ICRF), inner ring fault (IF). The faults under two working conditions are diagnosed respectively. 80% samples of each bearing health state are used for training, and the remaining samples are used for testing.
[0061] In order to compare the same working condition with different working conditions, in data set B, the data of different working conditions are regarded as a whole, and only one normal bearing is considered. Therefore, there are 7 health states in total. 80% of the samples of each bearing health state are used for training, and the remaining samples are used for testing. In actual situations, compound fault samples are more difficult to collect than single fault samples. Therefore, in data set C, the percentage of normal bearing training samples is 50%, the percentage of single fault training samples is 30%, and the percentage of compound fault training samples is 20% or 10%.
[0062] Table 1 Datasets used for testing
[0063]
[0064] Table 2 Dataset description
[0065]
[0066] 2 Structure diagram
[0067] Take 8 fault samples under 2 working conditions respectively, each signal length L is 1024000, the subsample d length is 1024, and the number of subsamples n is 1000. Every 10 subsamples are used to form a ε-radius of 0, 2β 2 The graph with 1024 nodes and 10 nodes can be expressed as:
[0068]
[0069] Each sample will first calculate the cosine similarity with its 10 nearest neighbors, and obtain the edge weight through formula 10, which can be expressed as:
[0070]
[0071] Then move to the next 10 neighbors until all samples are traversed.
[0072] The method used can effectively extract graph features. Node information is embedded in each layer, and the fault features in the original data are extracted by the graph convolution layer. This study is for node classification, so the input layer and the structure of the network after one convolution and two convolution layers do not change. Each layer of the network is visualized by Gephi, such as Figure 2 shown.
[0073] 3 Fault diagnosis structure
[0074] The processed data is put into the deep graph convolution model for processing, and the processing results are output:
[0075] The deep graph convolution model uses momentum stochastic gradient descent (SGD) as the optimizer for hyperparameter optimization. The momentum of SGD is 0.9, the number of iterations (epochs) is 100, the batch size is 64, and the learning rate decay strategy is also used to adjust the learning rate. The weight decay value is initialized to 0.0005, the learning rate is 0.001, the number of network nodes is 640, and the number of edges is 5760. The network connection mode of the model is input layer-convolution layer 1→normalization layer 1→convolution layer 2→normalization layer 2→fully connected layer→output layer. The detailed network structure of the model is shown in Table 3.
[0076] Table 3 Model structure
[0077]
[0078]
[0079] DGCN includes input layer, first convolution layer, first normalization layer, second convolution layer, second normalization layer, fully connected layer, and output layer;
[0080] In the first graph convolution layer, the input node feature map size is 512, the output node feature map size is 1024, and the number of graph convolution kernels is 512×1024.
[0081] The first normalization layer normalizes the data after the convolution operation. The number of channels of the output feature map is 1024, the step size is 1, and the ReLU activation function is used for nonlinear mapping;
[0082] In the second graph convolution layer, the input node feature map size is 1024, the output node feature map size is 1024, and the number of graph convolution kernels is 1024×1024.
[0083] The second normalization layer normalizes the data after the convolution operation. The number of channels of the output feature map is 1024, the step size is 1, and the ReLU activation function is used for nonlinear mapping.
[0084] In the fully connected layer, the input feature map size is 1024, the output feature map size is 512, and the activation function ReLU is used;
[0085] In the output layer, the input feature map size is 512, and the activation function Softmax is used.
[0086] In order to reduce the randomness of the results, the processed data is put into the deep graph convolution DGCN model again and repeated 9 times, and the average of the 10 test results is taken as the final result.
[0087] Results Analysis
[0088] like Figure 3 As shown, (a) and (b) are the confusion matrices of compound fault diagnosis under the same working condition, (c) is the confusion matrix of compound fault diagnosis under different working conditions, and (d) is the confusion matrix of compound fault diagnosis under sample imbalance. The results show that this method is effective for compound fault diagnosis under the same working condition, different working conditions and sample imbalance.
[0089] The above-mentioned embodiments are only illustrative of the principles and effects of the invention, as well as some embodiments of its application, and are not intended to limit the invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the invention, and all of these belong to the protection scope of the invention.
Claims
1. An intelligent diagnosis method for composite faults of rolling bearings based on deep graph convolutional networks, comprising the following steps: (1) Building graphs based on time series The original time series X with a signal length of L is normalized, and the normalized data is expressed as: X nol =normalization(X) (1) Where X nol is the normalized time series, normalization(·) represents the normalization method; Secondly, the original time series of the signal length L is divided into sub-samples of length d, with no overlap between each sample, and a corresponding label is assigned to each sub-sample. The obtained sub-sample set can be expressed as: Where ∏ represents the constructed super sample set, represents the subsample, y n represents the label, n represents the number of sub-sampling; Perform a fast Fourier transform (FFT) on each sub-sample and use the obtained spectrum as a new sample. The process can be expressed as: Where FFT(·) means converting the subsamples into the frequency domain and taking the first half of the result; The final labeled dataset is represented as: (2) Constructing graph structure data Construct a weighted graph with the labeled dataset obtained above, and construct a graph with n nodes by treating each subsample as a node in the graph; estimate the distance between samples and define a threshold ε; if the cosine similarity is greater than the threshold, there will be an edge between the two nodes; the node x can be obtained by i Neighbor nodes: in yes The neighbor nodes of , ε is the selected radius, and the threshold ε is 0. Compute Node and The cosine similarity of The weight between each two nodes is calculated through a threshold Gaussian kernel weight function: Where β is the bandwidth variance of the Gaussian function; (3) Fault diagnosis structure The processed data is put into the deep graph convolution model for processing, and the processing results are output; the deep graph convolution model uses momentum stochastic gradient descent SGD as the optimizer for hyperparameter optimization; the momentum of momentum stochastic gradient descent SGD is 0.9, the number of iterations epochs is 100, the batch size is 64, the learning rate decay strategy is used to adjust the learning rate, the weight decay value is initialized to 0.0005, the learning rate learn rate is 0.001, the number of network nodes is 640, and the number of edges is 5760; the deep graph convolution model includes an input layer, a first convolution layer, a first normalization layer, a second convolution layer, a second normalization layer, a fully connected layer, and an output layer; In the first graph convolution layer, the input node feature map size is 512, the output node feature map size is 1024, and the number of graph convolution kernels is 512×1024; The first normalization layer normalizes the data after the convolution operation. The number of channels of the output feature map is 1024, the step size is 1, and the ReLU activation function is used for nonlinear mapping; In the second graph convolution layer, the input node feature map size is 1024, the output node feature map size is 1024, and the number of graph convolution kernels is 1024×1024; The second normalization layer normalizes the data after the convolution operation. The number of channels of the output feature map is 1024, the step size is 1, and the ReLU activation function is used for nonlinear mapping. In the fully connected layer, the input feature map size is 1024, the output feature map size is 512, and the activation function ReLU is used; In the output layer, the input feature map size is 512, and the activation function Softmax is used.
2. The method for intelligent diagnosis of composite rolling bearing faults based on deep graph convolutional network according to claim 1, characterized in that: In step 3, the processed data is placed in the deep graph convolution model and processed repeatedly for more than 5 times, and the average value of all output processing results is taken as the final result.
3. The rolling bearing composite fault intelligent diagnosis method based on deep graph convolutional network according to claim 1, characterized in that: In step 2, cosine similarity is used to estimate the distance between samples.
4. The rolling bearing composite fault intelligent diagnosis method based on deep graph convolutional network according to claim 1, characterized in that: In step 1, the original time series X with a signal length of L is normalized using Max-min.
5. The method for intelligent diagnosis of composite rolling bearing faults based on deep graph convolutional network according to claim 1, characterized in that: In step 2, the distance between two nodes is calculated using the Euclidean algorithm: in, and i1 ,y i2 Represents nodes. The closer the value of d is to 1, the higher the similarity.
6. The rolling bearing composite fault intelligent diagnosis method based on deep graph convolutional network according to claim 1, characterized in that: In step 2, the distance between two nodes is calculated using the Manhattan distance algorithm: in, and i1 ,y i2 Represents nodes. The closer the value of d is to 1, the higher the similarity.
7. The method for intelligent diagnosis of composite rolling bearing faults based on deep graph convolutional network according to claim 1, characterized in that: In step 2, the distance between two nodes is calculated using the Chebyshev algorithm: in, and i1 ,y i2 Represents nodes. The closer the value of d is to 1, the higher the similarity.