A bearing fault diagnosis method, device and equipment based on graph convolutional network

Abstract

The present invention discloses a bearing fault diagnosis method, device and equipment based on a graph convolutional network, and relates to the technical field of rotating machinery fault detection. The present invention utilizes an association graph construction method based on a neighboring algorithm KNN and a node embedding method to construct target domain data and source domain data into an association graph with classified connection of feature nodes, so that an intrinsic association relationship between the target domain data and the source domain data is established, thereby reducing the characteristic distribution deviation of the target domain data and the source domain data, and then introduces a pseudo-label consistency learning strategy to adjust the classifier decision boundary so that it can more accurately classify the target domain data, and simultaneously uses an adaptive global threshold and an adaptive local threshold to obtain the average confidence of the accurately classified target domain data to calculate the global threshold, and uses the global threshold to judge the diagnosis result of the rolling bearing; deeper and more complex fault feature information in the target domain data can be captured, thereby greatly improving the accuracy of rolling bearing fault identification.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery fault detection, and in particular to a bearing fault diagnosis method, device and equipment based on a graph convolutional network. Background Art

[0002] Rolling bearings, as key components in rotating machinery, build bridges between the stationary and rotating parts of equipment, but they are also the main source of failures. According to statistics, about 30%-40% of the failures in rotating machinery come from bearing components. The main reason for this is that bearings operate under high temperature, high pressure and high load conditions for a long time, which can easily cause structural damage. Therefore, monitoring the operating conditions of bearings can enable staff to detect and solve faults in a timely manner, reduce the accident rate, and ensure the safe operation of mechanical equipment. However, since bearings are usually located inside the equipment, the initial symptoms of failures are not obvious, manual troubleshooting is difficult, and the later hazards are great. With the continuous development of industrial Internet and sensor technology, bearing fault diagnosis technology is becoming increasingly intelligent. Intelligent fault diagnosis collects a large amount of operating data through sensors, and then uses data mining technology to find hidden fault modes. Compared with traditional methods that rely on manual experience, it is more efficient and the diagnosis is more accurate. Among them, the deep learning-based method automatically learns advanced features, providing a powerful tool for intelligent fault diagnosis.

[0003] Bearing fault diagnosis methods based on deep learning can automatically mine fault features and achieve high-precision and high-efficiency diagnosis, and have become a current research hotspot; but most of them perform feature learning based on the assumption that fault samples are independent and identically distributed, ignoring the intrinsic connection between fault samples; for bearing samples of the same fault type, they will share some similar fault features, and by mining the correlation between them, the typical feature information of the fault type can be better learned; for samples of different fault types, there may be some implicit connections between them, which may not be directly reflected in the original features, but can be mined through the interaction and information transfer between samples, thereby helping to extract deep feature representations that distinguish different fault types; graph neural networks transmit and update information through the connection relationship between samples, which can help the model learn more complex and in-depth fault feature representations.

[0004] The bearing fault diagnosis method based on graph neural network aims to train the model with the collected historical bearing data (training set), and then apply it to determine the fault category of the current bearing (test set); this method is based on a premise assumption that the training set and test set of the bearing should obey the same distribution, but this assumption is difficult to meet in industrial scenarios; because the load and speed of the bearing change during the production and operation of mechanical equipment; the load refers to the pressure acting on the bearing, such as when a crane lifts goods of different weights, the load borne by the bearing will change accordingly; the speed refers to the speed at which the bearing rotates, and the corresponding speed will be adjusted according to the process requirements, such as the machine tool will switch to different speeds during rough processing and fine processing; this change in the working condition of the bearing causes the distribution difference between the collected training set and test set data, resulting in a domain shift problem, that is, the model trained in the training set cannot be well generalized in the test set; therefore, for the current fault diagnosis method based on deep learning, because it is based on the assumption that the fault samples are independent and identically distributed for feature learning, it ignores the intrinsic connection between the fault samples, and it is difficult to capture deeper and more complex fault feature representations, resulting in low accuracy in fault identification. Summary of the invention

[0005] The embodiments of the present invention provide a bearing fault diagnosis method, device and equipment based on a graph convolutional network, which can solve the problem of low accuracy of fault identification due to the existing deep learning-based fault diagnosis method in the prior art, which performs feature learning based on the assumption that fault samples are independent and identically distributed, ignoring the intrinsic connection between fault samples, and having difficulty in capturing deeper and more complex fault feature representations.

[0006] An embodiment of the present invention provides a bearing fault diagnosis method based on a graph convolutional network, comprising the following steps:

[0007] Obtain vibration signals of rolling bearings under different working conditions; decompose the vibration signals of rolling bearings under different working conditions into source domain data containing fault labels and target domain data without fault labels;

[0008] The feature extraction module is constructed using graph convolutional network (GCN), and the domain classification layer and fault classification layer are constructed using graph convolutional network (GCN) and maximum mean difference (MMD) respectively. The feature extraction module, domain classification layer and fault classification layer are cascaded to construct a network diagnosis model.

[0009] In the feature extraction module of the network diagnosis model, an association graph construction method and a node embedding method based on the neighboring algorithm KNN are introduced, and the association graph construction method and the node embedding method are integrated into the feature extraction module to obtain an improved feature extraction module; a pseudo-label consistency learning strategy is introduced into the domain classification layer of the network diagnosis model to obtain an improved domain classification layer; an adaptive threshold is introduced into the fault classification layer of the network diagnosis model to obtain an improved fault classification layer; and the improved feature extraction module, the improved domain classification layer and the improved fault classification layer are cascaded to form an improved network diagnosis model;

[0010] The target domain data and the source domain data are input into the improved network diagnosis model; the association graph construction method in the improved feature extraction module is used to construct an association graph with interconnected vertices, and the node embedding method in the improved feature extraction module is used to embed the features in the target domain data and the source domain data into the vertices in the association graph respectively, so as to form an association graph with classified connection of feature nodes; the pseudo-label consistency learning strategy in the improved domain classification layer is used to perform weak and strong enhancement on the target domain data in the association graph, and the pseudo-labels of the weakly enhanced data are used to supervise the strongly enhanced data, and the decision boundary of the target domain data classification is adjusted to the sparse area of ​​the target domain to classify the target domain data; the adaptive thresholds of the adaptive global threshold and the adaptive local threshold contained in the improved fault classification layer are used to set the classified target domain data to different thresholds, and the fault diagnosis results of the rolling bearing are output according to different thresholds.

[0011] Preferably, the construction of the feature extraction module is based on the characteristics of the graph convolutional network GCN for feature extraction and pattern recognition on graph structure data;

[0012] The maximum mean difference (MMD) is the difference between two data with different distributions. and Mapped to the common regenerated Hilbert space RKHS, and measure the distribution distance; the maximum mean difference MMD graph convolutional network GCN is used to construct the domain classification layer and the fault classification layer respectively.

[0013] Preferably, the construction of the association graph includes:

[0014] The association graph is constructed based on the KNN algorithm, that is, the Euclidean distance between each data feature in the target domain and the source domain is calculated to find the nearest vertex of each feature. N neighbors, and associate the feature vertex with N The neighbors are connected by edges, and the edge weights are assigned using the Gaussian kernel function; the graph composition equation based on the neighboring algorithm KNN is:

[0015]

[0016] in: Represents a vertex and The Euclidean distance between x ia express No. Dimensional data, each The dimension is ; x ja express No. Dimensional data, each The dimension is ; To return the vertex In the node set Center front N A set of vertices that are close to each other, and use express; Represents a vertex and vertices The weights of the connected edges; ζ represents the bandwidth of the Gaussian kernel;

[0017] In the association diagram G =( V , E ) is constructed, where Represents a collection of nodes. n is the number of nodes, E Represents a collection of edges.

[0018] Preferably, the association diagram G =( V , E ) After the construction is completed, the feature extraction module captures the association relationship between the target domain samples and the source domain samples in the association graph, including:

[0019] The constructed association graph is input into the graph convolutional network GCN to capture the association relationship between the target domain samples and the source domain samples, learn the spatial structure features, and the target domain and source domain fault data are l The spatial structure features learned by the +1 layer are:

[0020]

[0021] Where: σ(·) represents the nonlinear activation function; represents the adjacency matrix with self-connected edges; express Matrix of Indicates Features learned by the layer; express The weight parameters of the layer.

[0022] Preferably, after the association relationship between the target domain samples and the source domain samples is captured, the original feature distributions in the target domain samples and the source domain samples are aligned, and the multi-kernel maximum mean difference MK-MMD is used to align the target domain samples based on weak enhancement. and source domain Perform feature difference measurement;

[0023] The value of the multi-core maximum mean difference MK-MMD is used as the structural alignment loss, and the feature extraction module aligns the global features, the structural alignment loss L M The equation is:

[0024]

[0025] in: L M represents structural alignment loss; MMD represents maximum mean difference.

[0026] Preferably, the using of the pseudo-labels of the weakly enhanced data to supervise the strongly enhanced data comprises:

[0027] Weakly enhance samples in the target domain The class probability distribution In order to ensure the reliability of pseudo labels, the corresponding category with the maximum category probability greater than the threshold is selected as the hard label; the filtering equation of the pseudo label is:

[0028]

[0029] in: Representation sample Pseudo labels that meet the threshold; Indicates t In the iteration k Thresholds for categories; Representation sample No. k The probability of the class label;

[0030] The pseudo-label of the weakly enhanced sample in the target domain is used as the supervision information of the corresponding enhanced sample in the target domain to obtain the pseudo-label consistency loss. L PLC The equation is obtained as:

[0031]

[0032] in: B Indicates the batch size, μ Indicates the proportion of samples in a batch that meet the threshold. H represents the cross entropy loss of two probability distributions; Represents the number of samples in the target domain; Represents a strongly enhanced sample in the target domain The class probability distribution of ; Represents the class probability of the weakly enhanced samples in the target domain.

[0033] Preferably, the adaptive threshold sets the classified target domain data to different thresholds, including:

[0034] The global threshold is obtained by using the average confidence of the target domain data, and the global threshold is set as the exponential moving average of the confidence of each iteration. τ t , the equation for obtaining it is:

[0035]

[0036] Among them: the initialization value is 1 / K , K is the number of categories, is the momentum decay hyperparameter in EMA; Represents a model output sample probability; Indicates t- Threshold for 1 iteration; Indicates the probability of the sample label with the largest model output; B Indicates the size of the batch;

[0037] Adaptive local thresholding reflects each feature category k The learning status of each class is predicted using the expected Get the adaptive local threshold, expect The equation for obtaining is:

[0038]

[0039] Among them: the initialization value is 1 / K , K is the number of categories, is the momentum decay hyperparameter in EMA; Indicates t- Expected value of 1 iteration; B Indicates the size of the batch; Indicates that the sample is a category k probability;

[0040] Combine the adaptive global threshold and the adaptive local threshold to obtain the final adaptive threshold. The equation for obtaining is:

[0041]

[0042] in: represents the maximum normalization function.

[0043] The embodiment of the present invention also provides a bearing fault diagnosis device based on a graph convolutional network, comprising:

[0044] The detection module is used to obtain the vibration signal of the rolling bearing under different working conditions; decompose the vibration signal of the rolling bearing under different working conditions into source domain data containing fault labels and target domain data without fault labels;

[0045] Model module, using graph convolutional network GCN to build feature extraction module, using graph convolutional network GCN and maximum mean difference MMD to build domain classification layer and fault classification layer respectively; and cascading the feature extraction module, domain classification layer and fault classification layer to build a network diagnosis model;

[0046] In the feature extraction module of the network diagnosis model, an association graph construction method and a node embedding method based on the neighboring algorithm KNN are introduced, and the association graph construction method and the node embedding method are integrated into the feature extraction module to obtain an improved feature extraction module; a pseudo-label consistency learning strategy is introduced into the domain classification layer of the network diagnosis model to obtain an improved domain classification layer; an adaptive threshold is introduced into the fault classification layer of the network diagnosis model to obtain an improved fault classification layer; and the improved feature extraction module, the improved domain classification layer and the improved fault classification layer are cascaded to form an improved network diagnosis model;

[0047] The diagnosis module is used to input the target domain data and the source domain data into the improved network diagnosis model; use the association graph construction method in the improved feature extraction module to construct an association graph with interconnected vertices, use the node embedding method in the improved feature extraction module to embed the features in the target domain data and the source domain data on the vertices in the association graph respectively, and form an association graph with classified connection of feature nodes; use the pseudo-label consistency learning strategy in the improved domain classification layer to perform weak and strong enhancement on the target domain data in the association graph, use the pseudo-label of the weakly enhanced data to supervise the strongly enhanced data, and adjust the decision boundary of the target domain data classification to the sparse area of ​​the target domain to classify the target domain data; use the adaptive thresholds of the adaptive global threshold and the adaptive local threshold contained in the improved fault classification layer to set the classified target domain data to different thresholds, and output the fault diagnosis results of the rolling bearing according to different thresholds.

[0048] An embodiment of the present invention further provides an electronic device, including a memory and a processor;

[0049] The memory is used to store computer programs;

[0050] The processor is used to implement the steps of the bearing fault diagnosis method based on graph convolutional network as described above when executing the computer program stored in the memory.

[0051] The embodiment of the present invention provides a bearing fault diagnosis method, device and equipment based on graph convolutional network. Compared with the prior art, the beneficial effects are as follows:

[0052] The present invention utilizes an association graph construction method based on a KNN neighboring algorithm and a node embedding method to construct the target domain data and the source domain data into an association graph with classified connections of feature nodes, so as to establish an intrinsic association relationship between the target domain data and the source domain data, thereby reducing the characteristic distribution deviation of the target domain data and the source domain data, and then introduces a pseudo-label consistency learning strategy to adjust the classifier decision boundary so that it can more accurately classify the target domain data, and simultaneously uses an adaptive global threshold and an adaptive local threshold to obtain the average confidence of the accurately classified target domain data to calculate the global threshold, and uses the global threshold to judge the diagnosis result of the rolling bearing; it can capture deeper and more complex fault feature information in the target domain data, thereby greatly improving the accuracy of rolling bearing fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of the overall process of a bearing fault diagnosis method based on a graph convolutional network provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of the overall framework of a network in a bearing fault diagnosis method based on a graph convolutional network provided in an embodiment of the present invention;

[0055] Figure 3 A schematic diagram of an association constructed based on a KNN method for a bearing fault diagnosis method based on a graph convolutional network provided in an embodiment of the present invention;

[0056] Figure 4 A schematic diagram of a classifier of a bearing fault diagnosis method based on a graph convolutional network provided in an embodiment of the present invention misclassifying target domain data. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0058] See also Figure 1, an embodiment of the present invention provides a bearing fault diagnosis method based on a graph convolutional network, comprising the following steps:

[0059] Step 1: According to the vibration signals of rolling bearings under different working conditions, source domain training data containing fault labels and target domain training data without fault labels are obtained.

[0060] Step 2: Input the source domain training data and the target domain training data into the diagnosis model.

[0061] Step 3: The diagnostic model is constructed in the order of feature extraction module and feature transfer module. The feature transfer module includes domain classification layer and fault classification layer.

[0062] Step 4: Use the joint loss function to constrain the diagnostic model for training, and use the target domain verification data without fault labels to verify the trained diagnostic model, so as to use the verified diagnostic model to diagnose rolling bearing faults.

[0063] Step 5: The joint loss function includes the classification loss function of the source domain data, the domain classification loss function, and the distance loss function between the source domain data and the target domain data features.

[0064] Graph Convolutional Networks:

[0065] Graph Convolutional Network (GCN) can perform effective feature extraction and pattern recognition on graph structured data. G =( V , E ), Represents a collection of nodes. E ={e ij =( v i , v j )| v i , v j ∈ V} represents the set of edges, The adjacency matrix of the graph is represented as follows;

[0066]

[0067] Laplacian matrix of graph L = D - A ,in: L is a symmetric matrix, D Represents the degree matrix of the graph, and the calculation formula is as follows:

[0068]

[0069] The feedforward propagation formula of the spectral graph convolutional network is:

[0070]

[0071] in: X represents the input feature matrix, σ(·) represents the nonlinear activation function, , I is the identity matrix, according to get.

[0072] How big is the mean difference between multiple cores?

[0073] Maximum Mean Discrepancy (MMD) is the sum of two data with different distributions. and Mapped to the public Reproducing Kernel Hilbert Space (RKHS), and then measure the distribution distance; the empirical estimate of MMD is calculated as:

[0074]

[0075] Where: H k Indicates the use of kernel function in RKHS space k Define the distribution distance, Indicates that the mapping function maps sample data to RKHS, Represents learning through feature extractor Features, express Learned features.

[0076] Since the choice of kernel function will greatly affect the final result of MMD, Multiple Kernel Maximum Mean Discrepancy (MK-MMD) came into being. It uses a linear combination of multiple kernel functions to replace the original kernel function. The formula is as follows:

[0077]

[0078] in: m represents the number of kernel functions, Represents the weight parameters of different kernel functions.

[0079] This paper proposes a new framework combining graph convolutional network with unsupervised domain adaptation, a pseudo-label consistency fault diagnosis method based on graph convolutional network (Graph Convolutional Network Pseudo-Label Consistency, GCN_PLC), the overall framework is as follows Figure 2 As shown in the figure, it includes two modules: a graph convolutional feature extractor and a classifier shared by the two domains. The training of the model includes three parts of losses. For labeled source domain data, the cross entropy classification loss is calculated to measure the effectiveness of the classifier; the multiple kernel maximum mean discrepancy (MK-MMD) structural loss between the two domains is calculated to reduce the distribution difference; for unlabeled target domain data, the pseudo-label consistency loss is calculated to train the model to obtain the discrimination information of the target domain, thereby adjusting the decision boundary of the classifier.

[0080] Specifically:

[0081] Graph construction and node embedding:

[0082] In order to fully explore the intrinsic correlation of fault signal data, the present invention uses a graph convolutional network to learn spatial feature representation; the fault diagnosis method based on the graph convolutional network is to construct the fault samples into a graph structure for node feature learning, and then convert the fault diagnosis task into a node classification problem.

[0083] The present invention adopts a KNN-based approach to construct an association graph, such as Figure 3 As shown, the nearest vertex is found by calculating the Euclidean distance K neighbors, connect them with edges, and consider the different influences of different neighbor nodes on them, and use Gaussian kernel function to assign edge weights; the KNN-based graph construction process is expressed as:

[0084]

[0085] in: Represents a vertex and The Euclidean distance between x ia express No. Dimensional data, each The dimension is ; x ja express No. Dimensional data, each y The dimension is ; To return the vertex In the node set Center front K A set of vertices that are close to each other, and use express; Represents a vertex and vertices The weight of the connected edge; ζ represents the bandwidth of the Gaussian kernel.

[0086] After the above series of operations, the association diagram G =( V , E ) is constructed, where Represents a collection of nodes. n is the number of nodes, E Represents a set of edges; then inputs the feature extractor multi-layer GCN to capture the association between samples, learn spatial structure features, and fault data in l + 1 The features learned by the layer are represented as follows:

[0087]

[0088] Where: σ(·) represents the nonlinear activation function; represents the adjacency matrix with self-connected edges; express Matrix of Indicates Features learned by the layer; express The weight parameters of the layer.

[0089] The source domain and target domain data use the above graph construction method and shared feature extractor, which is helpful to extract more universal feature representation, thereby reducing the difficulty of migration; the present invention uses a pseudo-label consistency strategy to learn the complex discrimination information of unlabeled target domain data; to implement this strategy, two different enhancement operations are applied to the same target domain sample before graph construction, one weak enhancement operation is to add random changes to the signal and amplify its amplitude; the other strong enhancement is to divide the signal into a certain number of fragments and disrupt the order, and then add a random jitter to it; in summary, the feature extractor The feature representations of source domain samples, target domain weakly enhanced samples, and target domain strongly enhanced samples will be output respectively.

[0090] Feature difference measurement and classifier training:

[0091] In order to ensure that the original features of the source domain and the target domain are better distributed and aligned, MK-MMD is used to perform weak enhancement on the target domain. and source domain The feature difference is measured; the smaller the MK-MMD value is, the smaller the distribution deviation between the two domains is. Therefore, this result can be used as a loss to optimize the feature extractor to achieve global feature alignment. The structural alignment loss is expressed as follows:

[0092]

[0093] in: L M represents structural alignment loss; MMD represents maximum mean difference.

[0094] The classifier outputs through the fully connected layer and passes through the Softmax function to obtain the predicted probability of each category. Assume , and Represent the source domain samples , weakly enhanced samples in the target domain And the target domain strong enhancement sample The classification loss is based on the true label and predicted label of the source domain data, and the feature extractor and classifier parameters are updated through back propagation. For multi-classification tasks, the cross entropy loss function is used to calculate the classification loss, and its formula is as follows:

[0095]

[0096] in: For sample One-hot label.

[0097] Pseudo-label consistency learning:

[0098] Through the above feature alignment and classifier training, the source domain data can be accurately classified and the overall distribution difference between the two domains can be reduced; however, since only the source domain data has labels, the classifier only maximizes the inter-class difference of the source domain during the training process. In this process, some unique discriminative information in the target domain may be ignored. Therefore, the classifier may misclassify the target domain data, such as Figure 4 As shown; To solve this problem, the present invention uses a pseudo-label consistency learning strategy to enhance the model's learning of target domain discrimination information with the help of pseudo-labels to adjust the decision boundary to fall in the sparse area of ​​the target domain.

[0099] Pseudo-label learning is a classic method to solve semi-supervised problems. Pseudo-labels of unlabeled data are obtained through a pre-trained model, and then trained simultaneously with a small amount of labeled data until the model converges. The pseudo-labels enable the model to fully learn the information of the unlabeled data. Inspired by pseudo-label learning, in unsupervised domain adaptation methods, the target domain pseudo-labels are used to learn its discriminative features. Pseudo-label learning uses pseudo-labels to supervise the original samples for self-training, while the pseudo-label consistency learning used in the present invention performs weak and strong enhancements on the same target domain samples, and then uses the pseudo-labels of the weakly enhanced samples to supervise the strongly enhanced samples, and the training model learns the discriminative information of the target domain. This data enhancement method can better improve the accuracy and robustness of the classifier decision boundary for the classification of target domain data.

[0100] Specifically, for the weakly enhanced samples in the target domain The class probability distribution , in order to ensure the reliability of pseudo labels, the corresponding category with the maximum category probability greater than the threshold is selected as the hard label; the filtering formula of pseudo labels is:

[0101]

[0102] in: Representation sample Pseudo labels that meet the threshold; Indicates t In the iteration k Thresholds for categories; Representation sample No. k The probability of the class label.

[0103] The pseudo-label of the weakly enhanced sample in the target domain is used as the supervision information of the corresponding enhanced sample in the target domain, so the pseudo-label consistency loss is obtained. L PLC The calculation formula is as follows:

[0104]

[0105] in: B Indicates the batch size, μ Indicates the proportion of samples in a batch that meet the threshold. H represents the cross entropy loss of two probability distributions; Represents the number of samples in the target domain; Represents a strongly enhanced sample in the target domain The class probability distribution of ; Represents the class probability of the weakly enhanced samples in the target domain.

[0106] Finally, the entire framework combines the structural alignment loss, source domain classification loss, and target domain consistency loss to train and update the model parameters. The total loss is:

[0107]

[0108] in: α and β Represents the weight coefficient.

[0109] Adaptive Threshold:

[0110] Traditional pseudo-label learning selects high-confidence samples by setting a fixed threshold. Although the method is simple and efficient, it has three shortcomings:

[0111] ① In the early stage of model training, only a small number of samples have confidence levels exceeding the fixed threshold, and the data utilization rate is low, so a large amount of information in unlabeled samples is discarded.

[0112] ②The fixed threshold does not take into account the learning difficulty of different categories, which may lead to category imbalance and poor model performance.

[0113] ③The fixed threshold method cannot be adjusted according to the learning situation of the model during the training process, and there is a certain degree of blindness.

[0114] Therefore, the present invention adaptively adjusts the confidence threshold according to the learning situation of the model.

[0115] The adaptive threshold is divided into two parts: the adaptive global threshold and the adaptive local threshold. The adaptive global threshold is used to reflect the learning state of the overall data of the unlabeled target domain, and the value increases steadily during training to ensure that the noise pseudo-labels are discarded in the later stage. The present invention uses the average confidence of the model on the target domain data to calculate the global threshold. Indicates that t Indicates t iterations; however, it is time-consuming to calculate the confidence of all data in each iteration, so the global threshold is estimated as the exponential moving average (EMA) of the confidence of each iteration, calculated as follows:

[0116]

[0117] Among them: the initialization value is 1 / K , K is the number of categories, is the momentum decay hyperparameter in EMA, Represents a model output sample probability; Indicates t-1The threshold of the number of iterations; Indicates the probability of the sample label with the largest model output; B Indicates the batch size.

[0118] Adaptive local thresholding is used to reflect each specific category k The learning status of , considering the intra-class diversity (the difference between different samples in the same fault category) and the inter-class proximity relationship (the similarity between different fault categories), is obtained using the expectation of the model prediction for each category, and Indicates that the calculation formula is:

[0119]

[0120] Among them: the initialization value is 1 / K , K is the number of categories, is the momentum decay hyperparameter in EMA; Indicates t-1 The expected value of the iterations; B Indicates the size of the batch; Indicates that the sample is a category k probability.

[0121] Combine the global threshold and the local threshold to get the final adaptive threshold, and use The specific calculation is:

[0122]

[0123] in: represents the maximum normalization function.

[0124] Different from the existing models for solving cross-domain fault diagnosis problems, the present invention proposes a framework combining graph convolutional networks and domain adaptation, which mainly includes a feature extraction module, a pseudo-label consistency module and an adaptive threshold module; the graph convolutional network generalizes the convolutional neural network to graph data, and learns node embedding by aggregating neighbor node features through information transmission. It not only learns the feature representation of connected samples, but also transmits and updates information through the connection relationship between samples at a deeper level to capture the intrinsic relationship of the data, so it can learn more complex feature representations; since the pseudo-label consistency module provides supervision information for the unlabeled target domain to a certain extent, the model can learn the discrimination information of the unlabeled target domain, so it can improve the model's recognition ability for the target domain data; and the data enhancement operation in this module increases the diversity of the target domain data, which helps the model learn more robust feature representations; because the adaptive threshold module updates the threshold according to the training process of the model, and sets different thresholds considering the differences in the training of each fault category, the reliability of the pseudo-label is improved, thereby ensuring the effect of pseudo-label consistency learning.

[0125] The embodiment of the present invention also provides a bearing fault diagnosis device based on a graph convolutional network, comprising:

[0126] The detection module is used to obtain the vibration signals of the rolling bearing under different working conditions; and decompose the vibration signals of the rolling bearing under different working conditions into source domain data containing fault labels and target domain data without fault labels.

[0127] Model module, uses graph convolutional network GCN to build a feature extraction module, uses graph convolutional network GCN and maximum mean difference MMD to build domain classification layer and fault classification layer respectively; and cascades the feature extraction module, domain classification layer and fault classification layer to build a network diagnosis model.

[0128] An association graph construction method and a node embedding method based on the KNN neighboring algorithm are introduced into the feature extraction module of the network diagnosis model, and the association graph construction method and the node embedding method are integrated into the feature extraction module to obtain an improved feature extraction module; a pseudo-label consistency learning strategy is introduced into the domain classification layer of the network diagnosis model to obtain an improved domain classification layer; an adaptive threshold is introduced into the fault classification layer of the network diagnosis model to obtain an improved fault classification layer; and the improved feature extraction module, the improved domain classification layer and the improved fault classification layer are cascaded to form an improved network diagnosis model.

[0129] The diagnosis module is used to input the target domain data and the source domain data into the improved network diagnosis model; use the association graph construction method in the improved feature extraction module to construct an association graph with interconnected vertices, use the node embedding method in the improved feature extraction module to embed the features in the target domain data and the source domain data on the vertices in the association graph respectively, and form an association graph with classified connection of feature nodes; use the pseudo-label consistency learning strategy in the improved domain classification layer to perform weak and strong enhancement on the target domain data in the association graph, use the pseudo-label of the weakly enhanced data to supervise the strongly enhanced data, and adjust the decision boundary of the target domain data classification to the sparse area of ​​the target domain to classify the target domain data; use the adaptive thresholds of the adaptive global threshold and the adaptive local threshold contained in the improved fault classification layer to set the classified target domain data to different thresholds, and output the fault diagnosis results of the rolling bearing according to different thresholds.

[0130] An embodiment of the present invention further provides an electronic device, including a memory and a processor.

[0131] The memory is used to store computer programs.

[0132] When the processor is used to execute the computer program stored in the memory, the steps of the above-mentioned bearing fault diagnosis method based on graph convolutional network are implemented.

[0133] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A bearing fault diagnosis method based on graph convolutional network, characterized in that: The following steps are involved: Obtain vibration signals of rolling bearings under different working conditions; decompose the vibration signals of rolling bearings under different working conditions into source domain data containing fault labels and target domain data without fault labels; The feature extraction module is constructed using graph convolutional network (GCN), and the domain classification layer and fault classification layer are constructed using graph convolutional network (GCN) and maximum mean difference (MMD) respectively. The feature extraction module, domain classification layer and fault classification layer are cascaded to construct a network diagnosis model. In the feature extraction module of the network diagnosis model, the association graph construction method and node embedding method based on the neighboring algorithm KNN are introduced, and the association graph construction method and node embedding method are integrated into the feature extraction module to obtain an improved feature extraction module; A pseudo-label consistency learning strategy is introduced into the domain classification layer of the network diagnosis model to obtain an improved domain classification layer; an adaptive threshold is introduced into the fault classification layer of the network diagnosis model to obtain an improved fault classification layer; and cascading the improved feature extraction module, the improved domain classification layer and the improved fault classification layer to form an improved network diagnosis model; Inputting the target domain data and the source domain data into the improved network diagnosis model; An association graph with interconnected vertices is constructed using the association graph construction method based on the neighboring algorithm KNN in the improved feature extraction module. The features in the target domain data and the source domain data are respectively embedded in the vertices in the association graph using the node embedding method in the improved feature extraction module to form an association graph with classified connection of feature nodes. The target domain data in the association graph are weakly and strongly enhanced using the pseudo-label consistency learning strategy in the improved domain classification layer, and the strong enhanced data is supervised by the pseudo-label of the weakly enhanced data, and the decision boundary of the target domain data classification is adjusted to the sparse area of ​​the target domain to classify the target domain data. The classified target domain data is set to different thresholds using the adaptive thresholds of the adaptive global threshold and the adaptive local threshold contained in the improved fault classification layer, and the fault diagnosis results of the rolling bearing are output according to different thresholds.

2. A bearing fault diagnosis method based on graph convolutional network according to claim 1, characterized in that: The construction of the feature extraction module is based on the characteristics of the graph convolutional network GCN for feature extraction and pattern recognition on graph structure data; The maximum mean difference (MMD) is the difference between two data with different distributions. and Mapped to the common regenerated Hilbert space RKHS, and measure the distribution distance; the maximum mean difference MMD graph convolutional network GCN is used to construct the domain classification layer and the fault classification layer respectively.

3. A bearing fault diagnosis method based on graph convolutional network according to claim 1, characterized in that: The construction of the association graph includes: The association graph is constructed based on the neighboring algorithm KNN, that is, the Euclidean distance between each data feature in the target domain and the source domain is calculated to find the N nearest neighbors of each feature vertex, the feature vertex is connected to the N neighbors with edges, and the edge weight is assigned using the Gaussian kernel function; the graph construction equation based on the neighboring algorithm KNN is: Where: d ij Represents vertex x i and x j The Euclidean distance between ia Represents x i The a-th dimension of data, each x has a dimension of m; x ja Represents x j The a-th dimension data of , each dimension of x is m; KNN(·) returns the vertex x i In the node set The first N vertices with the closest distance are collected and Ne(x i ) means; e ij Represents vertex x i and vertex x j The weight of the connected edge; ζ represents the bandwidth of the Gaussian kernel; When the association graph G = (V, E) is constructed, where V ∈ R n represents a set of nodes, n is the number of nodes, and E represents a set of edges.

4. A bearing fault diagnosis method based on graph convolutional network according to claim 3, characterized in that: After the association graph G=(V, E) is constructed, the feature extraction module captures the association relationship between the target domain samples and the source domain samples in the association graph, including: The constructed association graph is input into the graph convolutional network GCN to capture the association relationship between the target domain samples and the source domain samples, and learn the spatial structure features. The spatial structure features learned by the target domain and source domain fault data in the l+1 layer are: Where: σ(·) represents the nonlinear activation function; represents the adjacency matrix with self-connected edges; express The matrix of l Represents the features learned by the lth layer; θ l+1 Represents the weight parameter of the l+1 layer.

5. A bearing fault diagnosis method based on graph convolutional network according to claim 4, characterized in that: After capturing the association relationship between the target domain samples and the source domain samples, the original feature distributions in the target domain samples and the source domain samples are aligned, and the multi-core maximum mean difference MK-MMD is used to align the target domain X based on weak enhancement. w and source domain X s Perform feature difference measurement; The value of the multi-core maximum mean difference MK-MMD is used as the structural alignment loss, and the feature extraction module aligns the global features. The structural alignment loss L M The equation is: L M =MMD 2 (X s ,X w ) Where: L M represents structural alignment loss; MMD represents maximum mean difference.

6. A bearing fault diagnosis method based on graph convolutional network according to claim 1, characterized in that: The method of using the pseudo-labels of the weakly enhanced data to supervise the strongly enhanced data includes: Weakly enhance samples in the target domain The class probability distribution In order to ensure the reliability of pseudo labels, the corresponding category with the maximum category probability greater than the threshold is selected as the hard label; the filtering equation of the pseudo label is: in: Representation sample Pseudo labels that meet the threshold; τ t (k) represents the threshold of the k-th category in the t-th iteration; Representation sample The probability of the k-th class label; The pseudo-label of the weakly enhanced sample in the target domain is used as the supervision information of the corresponding enhanced sample in the target domain to obtain the pseudo-label consistency loss, whose loss L PLC The equation is obtained as: Where: B represents the size of the batch, μ represents the proportion of samples in a batch that meet the threshold, H represents the cross entropy loss of two probability distributions; n t Represents the number of samples in the target domain; Represents a strongly enhanced sample in the target domain The class probability distribution of ; Represents the class probability of the weakly enhanced samples in the target domain.

7. A bearing fault diagnosis method based on graph convolutional network according to claim 1, characterized in that: The adaptive threshold sets the classified target domain data to different thresholds, including: The global threshold is obtained by using the average confidence of the target domain data, and the global threshold is set to the exponential moving average τ of the confidence of each iteration t , the equation for obtaining it is: Where: the initialization value is 1 / K, K is the number of categories, and λ∈(0,1) is the momentum decay hyperparameter in EMA; Represents a model output sample The probability of t-1 represents the threshold of the t-1th iteration; max(q i ) indicates the maximum sample label probability output by the model; B indicates the batch size; The adaptive local threshold reflects the learning status of each feature category k, using the expected E predicted by the model for each category. t (k) Obtain the adaptive local threshold, the expected E t The equation for obtaining (k) is: Where: the initialization value is 1 / K, K is the number of categories, λ∈(0,1) is the momentum decay hyperparameter in EMA; E t-1 represents the expected value of the t-1th iteration; B represents the batch size; q i (k) represents the probability that the sample belongs to category k; Combine the adaptive global threshold and the adaptive local threshold to obtain the final adaptive threshold, the adaptive threshold τ t The equation for obtaining (k) is: Where: MaxNorm(·) represents the maximum normalization function.

8. A bearing fault diagnosis device based on graph convolutional network, characterized in that: include: The detection module is used to obtain the vibration signal of the rolling bearing under different working conditions; decompose the vibration signal of the rolling bearing under different working conditions into source domain data containing fault labels and target domain data without fault labels; Model module, using graph convolutional network GCN to build feature extraction module, using graph convolutional network GCN and maximum mean difference MMD to build domain classification layer and fault classification layer respectively; and cascading the feature extraction module, domain classification layer and fault classification layer to build a network diagnosis model; In the feature extraction module of the network diagnosis model, the association graph construction method and node embedding method based on the neighboring algorithm KNN are introduced, and the association graph construction method and node embedding method are integrated into the feature extraction module to obtain an improved feature extraction module; A pseudo-label consistency learning strategy is introduced into the domain classification layer of the network diagnosis model to obtain an improved domain classification layer; an adaptive threshold is introduced into the fault classification layer of the network diagnosis model to obtain an improved fault classification layer; and cascading the improved feature extraction module, the improved domain classification layer and the improved fault classification layer to form an improved network diagnosis model; The diagnosis module is used to input the target domain data and the source domain data into the improved network diagnosis model; use the association graph construction method based on the neighboring algorithm KNN in the improved feature extraction module to construct an association graph with interconnected vertices, and use the node embedding method in the improved feature extraction module to embed the features in the target domain data and the source domain data into the vertices in the association graph respectively, so as to form an association graph with classified connection of feature nodes; use the pseudo-label consistency learning strategy in the improved domain classification layer to perform weak and strong enhancement on the target domain data in the association graph, and use the pseudo-label of the weakly enhanced data to supervise the strongly enhanced data, and adjust the decision boundary of the target domain data classification to the sparse area of ​​the target domain to classify the target domain data; use the adaptive thresholds of the adaptive global threshold and the adaptive local threshold contained in the improved fault classification layer to set the classified target domain data to different thresholds, and output the fault diagnosis results of the rolling bearing according to different thresholds.

9. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the steps of a bearing fault diagnosis method based on a graph convolutional network as described in any one of claims 1 to 7 when executing the computer program stored in the memory.

Images