A graph neural network fault diagnosis method, device and storage medium

Abstract

The application discloses a kind of graph neural network fault diagnosis method, device and storage medium, wherein method includes: collecting the vibration acceleration signal of rotating machinery under different experimental conditions, obtains sample according to vibration acceleration signal, and sample is labeled, obtains sample dataset;Construct the graph neural network of multi-scale space-time information fusion drive for identifying different working condition faults;According to sample dataset, the graph neural network is trained and optimized;In test phase, the test sample of different working conditions is input into the graph neural network after training, and the probability distribution of sample fault is output, the fault of maximum probability is obtained as the prediction sample label.The application starts from multi-dimensional time sequence information feature extraction, constructs the convolution kernel of different scale size under single branch respectively to carry out different scale feature extraction to original vibration signal, effectively overcome the problem of insufficient single scale structure feature extraction and large amount of calculation of multi-branch structure, and can be widely applied in the field of mechanical equipment fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical equipment fault diagnosis, and in particular to a graph neural network fault diagnosis method and device and storage medium. BACKGROUND

[0002] In recent years, with the continuous development of modern industrial technology and the modernization level of mechanical equipment, rotating machinery tends to be more complex, large-scale, high-performance, high-efficiency and highly automated. However, rotating machinery equipment works in harsh environments such as high load and complex and variable working conditions for a long time, and various faults such as wear and tear and fracture of core components such as gears and bearings inevitably occur during operation, which causes unstable performance or failure of the rotating machinery equipment. If it cannot be found and repaired in time, it will cause great safety hazards and economic losses. Studies have shown that if the fault diagnosis technology is used to intervene in maintenance in the early stage of failure, the failure rate and maintenance cost can be reduced. Therefore, how to accurately and intelligently diagnose the running status of rotating machinery is a problem that needs to be solved. Therefore, it is of great significance to carry out accurate intelligent diagnosis algorithm research to ensure the safe operation of equipment and improve production efficiency.

[0003] Deep learning generally refers to a deep neural network constructed by multiple hidden layers. The basic idea is to simulate the hierarchical structure of the human brain, process external input data through multiple nonlinear transformations, so as to learn multi-level abstract features and hidden structure representation from a large amount of data, and realize intelligent identification and prediction of input data. Therefore, the fault diagnosis method based on deep learning does not need to perform complex signal mechanism analysis, but automatically extracts effective features of data through neural network for fault diagnosis and classification, promoting the development and application of intelligent fault diagnosis method.

[0004] With the development of computer technology and data acquisition technology, a large amount of equipment operation data can be collected in industrial field, which promotes the fault diagnosis field to enter the intelligent era. However, the running condition of rotating machinery equipment is complex and variable, and slight condition changes may cause a large change in the collected vibration signal, resulting in different data distribution of the sample to be diagnosed and the training sample, so that the original fault diagnosis model is not completely suitable for the new running condition. If fault diagnosis models are established for different tasks respectively, not only time and manpower are consumed, but also sufficient fault samples are needed. However, in industrial applications, rotating machinery cannot run with faults for a long time, and the cost of collecting and labeling fault data for each task is very high. SUMMARY

[0005] To at least partially solve one of the technical problems in the prior art, the purpose of the present application is to provide a graph neural network fault diagnosis method, device and storage medium.

[0006] The technical scheme adopted by the present application is:

[0007] A graph neural network fault diagnosis method, comprising the following steps:

[0008] Collecting vibration acceleration signals of rotating machinery under different experimental conditions, obtaining samples according to the vibration acceleration signals, and labeling the samples to obtain a sample dataset;

[0009] A multi-scale spatio-temporal information fusion driven graph neural network for identifying different working condition faults is constructed;

[0010] The graph neural network is trained and optimized according to the sample dataset;

[0011] In the test phase, the test samples of different working conditions are input into the trained graph neural network, and the probability distribution of the sample fault is output, and the fault with the maximum probability is obtained as the predicted sample label;

[0012] The graph neural network includes a multi-scale feature extraction module, a channel attention module, a graph data construction module, a graph convolution module, and a label classification module.

[0013] The multi-scale feature extraction module is used to extract information of different scales from the input signal.

[0014] The channel attention module is used to model the multi-scale information of different channels, and a channel attention mechanism is added in the convolution layer to selectively enhance the features with large information quantity, so that the subsequent processing fully utilizes these features and suppresses useless features.

[0015] The graph data construction module is used to introduce the spatial information between samples,

[0016] The graph convolution module is used to model the spatial information and extract high-dimensional features that fuse spatio-temporal information,

[0017] The label classification module is used to output the class probability of the known class fault.

[0018] Further, the sample obtained according to the vibration acceleration signal comprises:

[0019] A data segment containing at least one cycle length of fault feature frequency in all fault categories is intercepted from the vibration acceleration signal to form a sample;

[0020] The sample includes training samples and test samples, and the training samples and test samples are samples under different working conditions, the data feature distributions of the two are different, but the class spaces of the two are the same, and the number of fault categories contained is the same.

[0021] Further, the multi-scale feature extraction module comprises a plurality of feature extraction units; wherein the first feature extraction unit comprises a plurality of convolution layers with different sizes of one-dimensional convolution kernels, a channel attention layer, a batch normalization layer, a maximum pooling layer with one-dimensional pooling kernels, and the subsequent feature extraction units comprise a convolution layer with a single scale one-dimensional convolution kernel, a batch normalization layer, and a maximum pooling layer with a one-dimensional pooling kernel.

[0022] Further, the expression of the multi-scale feature extraction module is:

[0023] Z l =C(σ(f i (Z l-1 )+b l ))

[0024] wherein Z l is the output of the lth layer, Z l-1 is the output of the (l-1)th layer, b l is the bias of the lth layer, f i represents a convolution operation of different scales, σ is a nonlinear activation function, and C represents a merging operation of features output by different scale convolution operations in the channel dimension.

[0025] Further, the channel attention module specifically performs the following steps:

[0026] A1, performing global maximum pooling and global average pooling operations on the feature map X output by the multi-scale feature extraction module; wherein the pooling operation compresses the H*1 feature map of each channel into a 1*1 number, and the number includes the global information of the channel; the vectors Z1 and Z2 after the pooling operation represent the global feature information, and the expression is as follows:

[0027] Z1, Z2=F sq (X)

[0028] In the formula, X represents the feature map output by the multi-scale convolution, F sq represents global maximum pooling and global average pooling, and Z1 and Z2 represent the output of the feature map after global maximum pooling and average pooling.

[0029] A2, mapping the vectors Z1 and Z2 through a fully connected layer to output S1 and S2, then adding S1 and S2 and activating to output a channel attention vector S, the dimension of the channel attention vector S is equal to the channel dimension of the multi-scale feature X, and the expression is:

[0030] S=σ(FC(Z1)+FC(Z2))

[0031] In the formula, FC(·) represents a shared fully connected layer, σ(·) represents a Sigmoid activation function, and S represents a channel attention vector.

[0032] After the sigmoid activation operation, all values in the channel attention vector S are compressed to 0 to 1, the multi-scale feature input X is multiplied by the channel attention vector S, and the feature map Y of the fusion channel attention mechanism is obtained, and the expression is:

[0033] Y=F sc (S, X)=S T X

[0034] In the formula, F sc (·) represents the scale operation on X.

[0035] Further, the graph data construction module includes a full connection layer and a feature matrix and its transpose matrix are multiplied to obtain an adjacency matrix, the adjacency matrix selects top-k neighbors of each node according to a top-k sorting mechanism and sets the values to 1, indicating that there is a connection edge between the nodes, and the values in the rest of the adjacency matrix are set to 0, indicating that there is no connection edge between the nodes.

[0036] Further, the expression of the adjacency matrix is:

[0037]

[0038] In the formula, is the output of the MLP, A is the constructed adjacency matrix, is a sparse matrix with only 0 and 1, normalize(·) represents a normalization function, Top represents, and k(A) represents.

[0039] Further, the graph convolution module includes two feature extraction units, each feature extraction unit including a graph convolution layer and a batch normalization layer.

[0040] The expression of the graph convolution module is:

[0041]

[0042] In the formula, H 0 and H 1 are the feature representations learned by the first layer and the second layer graph convolution respectively, A is the adjacency matrix, W 0 and W 1 are the learnable parameters of the graph convolution layer.

[0043] Another technical solution adopted by the application is:

[0044] A graph neural network fault diagnosis device, comprising:

[0045] At least one processor;

[0046] at least one memory for storing at least one program;

[0047] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0048] Another technical solution adopted by the present application is:

[0049] A computer-readable storage medium, wherein a processor-executable program is stored, the processor-executable program is used to execute the method as described above when executed by a processor.

[0050] The beneficial effects of the present application are: the present application extracts multi-dimensional time sequence information features, constructs single-branch convolution kernels of different scales to extract different scale features of the original vibration signal, and only a small amount of calculation is required to extract multi-scale fault features of time sequence data, effectively overcoming the problems of insufficient single-scale structure feature extraction and large calculation amount of multi-branch structure. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following introduces the drawings of the related technical solutions in the embodiments of the present application or the prior art. It should be understood that the drawings in the following introduction are only for the convenience of clearly describing part of the embodiments in the technical solutions of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the premise of the drawings.

[0052] Figure 1 is a schematic diagram of a multi-scale space-time information fusion driven graph neural network in the embodiments of the present application;

[0053] Figure 2 is a structural schematic diagram of multi-scale convolution in the embodiments of the present application;

[0054] Figure 3 is a structural schematic diagram of channel attention mechanism in the embodiments of the present application;

[0055] Figure 4 is a schematic diagram of graph data construction in the embodiments of the present application;

[0056] Figure 5 is a schematic diagram of ablation experiment comparison results under variable working conditions in the embodiments of the present application;

[0057] Figure 6 is a schematic diagram of comparison results of different diagnostic model experiments in the embodiments of the present application;

[0058] Figure 7 is a step flowchart of a graph neural network fault diagnosis method in the embodiments of the present application. DETAILED DESCRIPTION

[0059] Embodiments of the present application are described below in detail with reference to the accompanying drawings, wherein the same or similar numerals represent the same or similar elements or elements having the same or similar functions throughout the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation on the present application. For the step numbers in the following embodiments, they are only set for the convenience of setting the description, and the order between the steps is not limited in any way, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0060] In the description of the present application, it should be understood that the orientation description, such as the orientation or position relationship indicated by up, down, front, back, left, right, etc. is based on the orientation or position relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and is not intended to indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0061] In the description of the present application, the meaning of several is one or more, and the meaning of multiple is more than two, greater than, less than, more than, etc. are understood as not including the number, and above, below, etc. are understood as including the number. If it is described as first, second, it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or the order of indicated technical features.

[0062] In the description of the present application, unless otherwise explicitly limited, the words such as setting, installing, connecting, etc. should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0063] As Figure 7 As shown in the figure, in order to solve the problem that the fault diagnosis effect of the planetary gearbox is poor in a strong noise environment and under variable working conditions, and the generalization ability is poor, the embodiment provides a graph neural network fault diagnosis method. The method first constructs a multi-scale convolution kernel to extract different scale features from the original time series signal to enhance the fault feature expression ability; then a channel attention mechanism is constructed, and different weights are adaptively assigned to the features extracted by different scale convolution kernels according to the importance of channel features, and the features are strengthened; finally, the multi-scale features output by convolution are constructed into graph data in the spatial domain for graph convolution learning, so as to effectively utilize the time series multi-dimensional information and spatial structure correlation information of the data, realize the deep fusion of the spatial and temporal domain fault information under multi-scale, and improve the accuracy of diagnosis and the generalization performance of the model. The method specifically includes the following steps:

[0064] S1, collect vibration acceleration signals of the rotating machinery under different experimental conditions, obtain samples according to the vibration acceleration signals, and label the samples to obtain a sample data set.

[0065] The data set is collected and calibrated. The vibration acceleration signals of the rotating machinery under different experimental conditions are collected, a certain length of data segment is intercepted from the vibration acceleration signals to constitute a sample, and the sample is labeled to obtain a large number of different working condition sample data sets {x, y}. The data set is composed of sample data x and its corresponding label y.

[0066] The sample includes training samples and test samples, the training samples and the test samples are samples under different working conditions, the data feature distributions of the two are different, but the class spaces of the two are the same, the number of fault classes contained is the same, and the ability of the proposed diagnosis method to diagnose data faults under different working conditions is verified.

[0067] S2, construct a multi-scale spatio-temporal information fusion driven graph neural network for identifying faults under different working conditions.

[0068] Referring to Figure 1 , the graph neural network comprises a multi-scale feature extraction module, a channel attention module, a graph data construction module, a graph convolution module and a label classification module. The multi-scale feature extraction module is used to extract information of different scales from the input signal. The channel attention module is used to model the multi-scale information of different channels. Adding a channel attention mechanism in the convolution layer can selectively enhance features with large information quantity, so that the subsequent processing can fully utilize these features and suppress useless features. The graph data construction introduces spatial information between samples. The graph convolution module is used to model the spatial information and extract high-dimensional features that fuse the spatial and temporal information. The label classifier is used to output the class probability of the known class fault.

[0069] Specifically, the multi-scale feature extractor adopts multiple one-dimensional convolution kernels of different sizes. The channel attention module retains channel information through global pooling, and then maps it through a fully connected layer and outputs the importance of each channel through Sigmoid. The graph data construction module determines whether there is an edge between samples by calculating the cosine similarity between sample features. The graph convolution module extracts spatial information between samples through graph convolution for fault diagnosis. The label classifier output layer adopts a Softmax output function.

[0070] Referring to Figure 2As an optional implementation, the constructed multi-scale feature extraction module includes a plurality of feature extraction units, the first unit includes a convolution layer including a plurality of one-dimensional convolution kernels of different sizes, a channel attention layer, a batch normalization layer, and a maximum pooling layer of a one-dimensional pooling kernel, and subsequent feature extraction units include a convolution layer of a single scale one-dimensional convolution kernel, a batch normalization layer, and a maximum pooling layer of a one-dimensional pooling kernel. The mathematical model of the multi-scale feature extraction module is as follows:

[0071] Z l =C(σ(f i (Z l-1 )+b l )) (1)

[0072] Wherein Z l is the output of the lth layer, Z l-1 is the output of the (l-1)th layer, b l is the bias of the lth layer, f i represents a convolution operation of different scales, sigma is a nonlinear activation function, and C represents a merging operation of features output by different scale convolution operations in the channel dimension. It should be noted that the difference between this step and other methods is that the multi-scale convolution is directly implemented by using one-dimensional convolution kernels of different sizes, which is different from other methods that use different coarse-grained vibration signals. Other methods input the input signal into different convolution neural network branches after artificially dividing the input signal into different coarse-grained signals to extract features, and the quality of multi-scale features depends on the division of multi-scale coarse-grained layers. It is difficult to effectively measure and fuse the multi-scale coarse-grained layers with their feature weights; meanwhile, multi-scale feature fusion is performed in the high-dimensional full connection layer of different branch networks, resulting in high model complexity and large parameter quantity.

[0073] Referring to Figure 3 As an optional implementation, the constructed channel attention module guides the calculation resources to the part with the largest amount of information in the input signal through the attention mechanism, which can improve the performance in the deep network. The channel attention mechanism is specially used to model the relevance of each channel. The output of the convolution layer does not consider the dependency of each channel. Adding a channel attention mechanism in the convolution layer can selectively enhance features with a large amount of information, so that the subsequent processing can fully utilize these features and suppress useless features. The steps include:

[0074] Step 1: Global max pooling and global average pooling operations are performed on the feature map X output by the multi-scale convolution. In order to more efficiently calculate the channel attention features, the spatial dimension of the feature map needs to be compressed. Both average pooling and max pooling can compress the spatial dimension, so both pooling operations are used. The pooling operation compresses the H*1 feature map of each channel into a 1*1 value, which includes the global information of the channel. The Z1 and Z2 vectors after pooling can represent the global feature information, so the pooling operation can reduce the amount of calculation without losing the amount of information. The specific principle is as follows:

[0075] Z1, Z2 = F sq (X) (2)

[0076] Where X represents the feature map output by the multi-scale convolution, F sq represents global max pooling and global average pooling, and Z1 and Z2 represent the outputs of the feature map after global max pooling and average pooling.

[0077] Step 2: Z1 and Z2 are mapped to S1 and S2 through a fully connected layer, and Z1 and Z2 pass through the same fully connected layer to reduce the amount of calculation. S1 and S2 are then added and activated to output the channel attention vector S. The dimension of the vector is equal to the channel dimension of the multi-scale feature X. The purpose of adding S1 and S2 vectors is to integrate the effects of max pooling and average pooling. The expression is as follows:

[0078] S = σ(FC(Z1) + FC(Z2)) (3)

[0079] Where FC(·) represents a shared fully connected layer, σ(·) represents a Sigmoid activation function, and S represents a channel attention vector.

[0080] Step 3: In equation (3), after the Sigmoid activation operation, all values in the S vector are compressed to 0 to 1. The specific meaning of the value represents the importance of each channel of the multi-scale feature X. Then, F sc (·) multiplies the multi-scale feature input X and the channel attention vector S to obtain the feature map Y with the fused channel attention mechanism. The principle is as follows:

[0081] Y = F sc (S, X) = S T × X (4)

[0082] Referring to Figure 4 , as an optional implementation, the graph data construction module and the graph convolution module must input the node feature matrix X and the adjacency matrix A to perform graph convolution operations. The original signal is extracted by CNN to form the node matrix X, and the extracted feature map can be described as follows:

[0083] X = CNN(X input ) (5)

[0084] where X input represents the input matrix of the mini-batch sample.

[0085] The instance graph is constructed and the adjacency matrix A is obtained by the input matrix of the mini-batch. First, the extracted feature matrix is input into a multi-layer perceptron (MLP) for dimension reduction, and the adjacency matrix is obtained by matrix multiplication of the reduced features and their transpose matrix. Finally, according to the top-k sorting mechanism, the top-k neighbors of each node are selected and their values are set to 1, indicating that there is a connection edge between the nodes. The values in the rest of the adjacency matrix are set to 0, indicating that there is no connection edge between the nodes. The MLP dimension reduction and Top-k operation are to reduce the computational complexity. The adjacency matrix can be obtained by the following formula:

[0086]

[0087] where is the output of the MLP, A is the constructed adjacency matrix, is a sparse matrix with only 0 and 1, and normalize(·) represents a normalization function.

[0088] As an optional implementation, the graph convolution module includes two feature extraction units, each of which includes a graph convolution layer and a batch normalization layer. The graph convolution operation performed is represented by equation (7):

[0089]

[0090] In the formula, H 0 and H 1 are the feature representations learned by the first and second graph convolution layers, respectively, A is the adjacency matrix, and W 0 and W 1 are the learnable parameters of the graph convolution layer.

[0091] As an optional implementation, the label classification module is composed of a fully connected layer, and the output layer of the label classification module uses a Softmax output function.

[0092] S3, training and optimizing the graph neural network according to the sample data set.

[0093] Using the data set {x, y} constructed in step S1, the parameters of the multi-scale feature extractor, the graph convolution module, and the label classifier are trained and optimized using traditional supervised learning algorithms and gradient descent methods. At the same time, the hyperparameters in the model training process are determined, such as the number of iterations, the learning rate, the sample batch, etc.

[0094] S4, in the test phase, input the test samples of different working conditions into the trained graph neural network, output the probability distribution of sample failure, and obtain the most probable failure as the predicted sample label.

[0095] In the test phase, input the test samples of different working conditions into the network model, and the Softmax output of the label classification module outputs the probability distribution of sample failure. The most probable failure is taken as the predicted sample label.

[0096] The above method will be further described in detail below in combination with the drawings and specific embodiments.

[0097] To evaluate the performance of the proposed method, a fan gearbox fault dataset with planetary gear train structure is used for experimental verification. The experimental settings are combinations of rotational speed (300r / min, 480r / min, 600r / min, 720r / min) and load (0hp, 2hp, 4hp), and the experimental data is collected by an acceleration sensor installed on the bearing seat of the gearbox robot experimental platform with a sampling frequency of 25.6kHz. In this experiment, a normal mode (Normal) and four fault modes are set. The failure modes are planetary gear broken tooth (PG_broken), planetary gear crack (PG_crack), sun gear broken tooth (SG_broken), and ring gear fault (RF_left).

[0098] The datasets of fixed rotational speed and changing load and fixed load and changing rotational speed are selected respectively for cross-condition experimental verification. Three sub-datasets with rotational speed of 300r / min and load of 0hp, 2hp and 4hp are used, respectively named Load 0, Load 2 and Load 4; three sub-datasets with load of 1hp and rotational speed of 480r / min, 600r / min and 720r / min are used, respectively named Speed 480, Speed 600 and Speed 720.

[0099] A 1024-size sliding window is used to collect samples of each category in each working condition. Each working condition includes five types of health status and five types of failure, and 500 samples are collected for each type of failure. Each working condition has a total of 2500 samples. Variable load working condition fault diagnosis: includes six fault diagnosis tasks, in each task, the model is trained on one working condition dataset and tested on another working condition dataset. The specific settings of the experimental tasks are shown in Table 1. In each task, the training set is divided into training set and validation set according to 8:2, the model is trained on the training set, and the optimal model is selected through the validation set, and finally the model accuracy is tested on the test set.

[0100] Table 1

[0101]

[0102] To verify the effectiveness of the proposed model, the WDCNN, LeNet, LSTM, ResNet, MSCNN models in the field of fault diagnosis are selected for comparison experiments. WDCNN: the first layer adopts a one-dimensional convolutional neural network fault diagnosis model with a large size of convolution kernel; LeNet: a classic convolutional neural network for image classification in computer vision, which is one-dimensional convolution to meet the one-dimensional characteristics of vibration signals. LSTM: a long short-term memory neural network is used to identify fault signals, and the specific parameter settings of the network are consistent with the reference. ResNet: a one-dimensional convolutional neural network model with residual connection; MSCNN: a neural network model with multi-scale convolution. All model algorithms in this paper are run and compared on the same computer, which has a CPU model of AMD Ryzen 7 4800H and a GPU model of Geforce RTX 2060. The model is built using the deep learning framework Pytorch, the optimizer used for training is Adam, the momentum is set to 0.9, the weight_dcay is set to 0.00001, the learning rate is 0.0001, the sample batch is set to 64, and the iteration number is 150 times to stop training.

[0103] In order to objectively and reasonably evaluate the performance of different methods, it is necessary to select appropriate evaluation indicators. In addition to selecting accuracy as an evaluation indicator, F1 score is also selected as an evaluation indicator. F1 score is the harmonic mean of precision P and recall R, which is a comprehensive indicator, and its value is between 0 and 1, the larger the value, the higher the algorithm performance. Its calculation formula is as follows:

[0104]

[0105] Among them, the precision P represents the proportion of all predicted positive samples that are actually positive samples. The recall R represents the proportion of all true positive samples that are accurately predicted as positive samples.

[0106] To verify the effectiveness of the proposed model structure and the effectiveness of the model in variable working condition fault diagnosis, ablation experiments, variable load working condition fault diagnosis and variable speed working condition fault diagnosis experiments are designed. According to Table 1, the ablation experiment and the variable load working condition fault diagnosis experiment are carried out, and to avoid accidental, 5 repeated experiments are carried out on each task.

[0107] To verify the effectiveness of each module of the proposed model, the ablation comparison of each model proposed by the proposed method is carried out, and the ablation experiment results are as follows Figure 5As shown in Table 2, the proposed model achieves the best diagnostic accuracy on all tasks. By comparing the average diagnostic accuracy of CNN and multi-scale CNN models on the 6 tasks, the CNN accuracy is only 92.63%, and the multi-scale convolution improves the diagnostic accuracy to 96.62% without significantly increasing the number of model parameters. It can be seen that the multi-scale convolution operation can improve the feature extraction capability of CNN, and the extracted features have stronger fault representation capability, which verifies the effectiveness of the multi-scale CNN module. By comparing the diagnostic accuracy of the multi-scale CNN + attention model, the multi-scale CNN + GCN model and the multi-scale CNN model on each experimental task, compared with the multi-scale CNN, adding the attention mechanism and the GCN can improve the cross-condition diagnostic ability of the model. It shows that the attention mechanism can effectively give large weight to important features in multi-scale features and small weight to redundant features to improve learning efficiency and thus improve diagnostic accuracy. It shows that GCN can extract spatial features between samples, and sample spatial features have a certain role in improving diagnostic accuracy. The proposed method integrates the advantages of multi-scale CNN, attention mechanism for extracting time features and GCN for extracting spatial features, and achieves the highest diagnostic accuracy on the 6 cross-load condition diagnostic tasks, with an average accuracy of 98.85% and an average F1 score of 0.986.

[0108] Table 2 Ablation test results under variable load conditions

[0109]

[0110] To verify the effectiveness of the proposed model in variable load condition diagnosis, variable load condition diagnosis tasks are set according to Table 2. The prediction accuracy of the proposed method and the comparison method on the 6 diagnostic tasks is shown in Table 3. Figure 6 As shown in Table 3, the proposed method and the comparison method can achieve the best diagnostic accuracy in different tasks, which verifies the effectiveness of the proposed method in variable load condition fault diagnosis. WDCNN has a diagnosis accuracy close to 100% in the same condition, but the average accuracy is only 82.29% in the variable load condition, indicating that the generalization performance of the model is poor and it is difficult to apply in the actual production environment with variable conditions. The comparison of the experimental results of MSCNN and the proposed method shows that the graph convolution layer can effectively extract sample spatial features and improve the diagnostic accuracy and generalization performance of the model. The average diagnostic accuracy of the proposed method on different tasks is 98.85%, and the average F1 score is 0.986, as shown in Table 3. It shows that the proposed method has good generalization performance on different tasks, and the model has stronger robustness.

[0111] Table 3 Comparison results under variable load conditions

[0112]

[0113] The application aims at the problem of inconsistent distribution of training data and test data when different working condition fault diagnosis is performed, takes gear box fault as the research object, and effectively improves the accuracy of fault diagnosis under different working conditions through a graph neural network driven by multi-scale space-time information fusion, and improves the fault diagnosis capability of industrial equipment.

[0114] In summary, compared with the prior art, the application has at least the following advantages and beneficial effects:

[0115] (1) The application is based on multi-dimensional time sequence information feature extraction, and different scale size convolution kernels under a single branch are constructed to extract different scale features of the original vibration signal. By increasing a small amount of calculation, the multi-scale fault features of the time sequence data can be extracted, effectively overcoming the problems of insufficient single-scale structure feature extraction and large calculation amount of multi-branch structure.

[0116] (2) The application is based on the importance of fault features, and adopts a channel attention mechanism to adaptively weight the extracted multi-scale features, effectively improving the concentration of the model on different scale fault features and improving the calculation efficiency.

[0117] (3) The application is based on space-time domain information fusion diagnosis, uses the time dimension features extracted by the convolutional neural network to construct the graph data in the space domain, and uses the graph convolutional neural network to deeply fuse the space-time domain features, effectively improving the diagnostic accuracy and generalization performance of the model.

[0118] The application also provides a graph neural network fault diagnosis device, which comprises:

[0119] At least one processor;

[0120] At least one memory for storing at least one program;

[0121] When the at least one program is executed by the at least one processor, the at least one processor implements the method. Figure 7 The method.

[0122] The graph neural network fault diagnosis device of the embodiment can execute the graph neural network fault diagnosis method provided by the method embodiment, and can execute any combination of the steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0123] The application embodiment also discloses a computer program product or a computer program, which comprises computer instructions stored in a computer readable storage medium. The processor of the computer device can read the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method. Figure 7 The method.

[0124] The embodiment also provides a storage medium storing instructions or programs for executing the method for diagnosing faults of a graph neural network provided by the embodiment of the application. When the instructions or programs are executed, any combination of the method embodiments can be executed to have the corresponding functions and advantages of the method.

[0125] In some alternative embodiments, the functions / operations mentioned in the block diagram can not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two blocks shown in succession can actually be executed substantially simultaneously or the blocks can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flowcharts of the present application are provided by way of example, with the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of larger operations are independently executed.

[0126] In addition, although the present application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the described functions and / or features can be integrated in a single physical device and / or software module, or one or more functions and / or features can be implemented in separate physical devices or software modules. It can also be understood that a detailed discussion of the actual implementation of each module is unnecessary for an understanding of the present application. Rather, given the properties, functions and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be within the routine skill of the engineer, given the benefit of this disclosure. Thus, the present application, as set forth in the claims, is capable of being practiced without resort to undue experimentation in light of the disclosure presented herein. It can also be understood that the disclosed specific concepts are merely illustrative and are not intended to limit the scope of the present application, which is defined by the full scope of the appended claims and their equivalents.

[0127] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0128] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logic functions, which can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus or device, such as a computer-based system, a system including a processor or other system that can fetch the instructions from the instruction execution system, apparatus or device and execute the instructions, or in conjunction with these instructions execution systems, apparatus or devices. For the purpose of this specification, the "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by or in connection with an instruction execution system, apparatus or device, or in conjunction with these instruction execution systems, apparatus or devices.

[0129] More specific examples (non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium can even be paper or other suitable medium on which the program can be printed, because the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpreting or otherwise processing, if necessary, in other suitable ways, to be electronically obtained and then stored in the computer memory.

[0130] It should be understood that various parts of the present application can be implemented in hardware, software, firmware or a combination thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, known in the art, or their equivalents, can be employed for implementation: discrete logic circuitry having logic gates for implementing logic functions upon data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0131] In the above description of the present specification, the description of the terms "one embodiment / one example", "another embodiment / another example", or "certain embodiments / certain examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0132] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the claims and their equivalents.

[0133] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the above-described embodiments, and those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present application, and these equivalent modifications or substitutions are included in the scope defined by the claims of the present application.

Claims

1. A graph neural network fault diagnosis method, characterized in that, Includes the following steps: Vibration acceleration signals of rotating machinery under different experimental conditions are collected, samples are obtained based on the vibration acceleration signals, and the samples are labeled to obtain a sample dataset. Construct a graph neural network driven by multi-scale spatiotemporal information fusion for identifying faults under different operating conditions; The graph neural network is trained and optimized based on the sample dataset; During the testing phase, test samples under different working conditions are input into the trained graph neural network, which outputs the probability distribution of sample failures and obtains the failure with the highest probability as the predicted sample label. The graph neural network includes a multi-scale feature extraction module, a channel attention module, a graph data construction module, a graph convolution module, and a label classification module. The multi-scale feature extraction module is used to extract information at different scales from the input signal; The channel attention module is used to model multi-scale information of different channels. The channel attention mechanism is added to the convolutional layer so that the network can selectively enhance features with large information content, so that subsequent processing can make full use of these features and suppress useless features. The graph data construction module is used to introduce spatial information between samples. The graph convolution module is used to model spatial information and extract high-dimensional features that fuse spatiotemporal information. The label classification module is used to output the category probability of known fault types; The multi-scale feature extraction module includes multiple feature extraction units; The first feature extraction unit includes convolutional layers with multiple one-dimensional convolutional kernels of different sizes, channel attention layers, batch normalization layers, and max pooling layers with one-dimensional pooling kernels. Subsequent feature extraction units include convolutional layers with one-dimensional convolutional kernels of a single scale, batch normalization layers, and max pooling layers with one-dimensional pooling kernels.

2. The graph neural network fault diagnosis method according to claim 1, characterized in that, The process of obtaining samples based on vibration acceleration signals includes: Samples are formed by extracting data segments containing at least one cycle length of the fault characteristic frequency from all fault categories in the vibration acceleration signal; The samples include training samples and test samples. The training samples and test samples are samples under different working conditions. The data feature distributions of the two are different, but their class spaces are the same and they contain the same number of fault categories.

3. The graph neural network fault diagnosis method according to claim 1, characterized in that, The expression for the multi-scale feature extraction module is: in For the first The output of the layer, For the first The output of the layer, For the first Layer bias, Represents convolution operations at different scales. It is the Sigmoid activation function. This represents the merging operation of features output by convolution operations at different scales along the channel dimension.

4. The graph neural network fault diagnosis method according to claim 1, characterized in that, The channel attention module specifically performs the following steps: A1. Feature map output by the multi-scale feature extraction module Perform global max pooling and global mean pooling operations; the pooling operation compresses the H*1 feature map of each channel into a 1*1 value, which includes the global information of that channel; the resulting vector... The global feature information is represented by the following expression: In the formula, The feature map representing the output of multi-scale convolution. This indicates that global max pooling and global mean pooling operations are performed on the features; A2, Transform the vector Feature mapping is performed through a fully connected layer. Then The output channel attention vector is summed and activated. Channel attention vector Dimensionality and multi-scale features The channel dimensions are equal, and the expression is: In the formula, This represents a shared, fully connected layer. This represents the Sigmoid activation function. Represents the channel attention vector; A3. Channel attention vector after Sigmoid activation operation All values ​​are compressed to 0 to 1, and multi-scale features are input. With channel attention vector Multiplying these together yields the feature map of the fusion channel attention mechanism. The expression is: In the formula, Indicates to Perform a scale operation.

5. The graph neural network fault diagnosis method according to claim 1, characterized in that, The graph data construction module includes a fully connected layer and a feature matrix and its transpose matrix multiplied to obtain an adjacency matrix. The adjacency matrix selects the top-k nearest neighbors of each node according to the top-k sorting mechanism and sets their values ​​to 1, indicating that there are connecting edges between the nodes. The values ​​in the remaining adjacency matrix are set to 0, indicating that there are no connecting edges between the nodes.

6. The graph neural network fault diagnosis method according to claim 1, characterized in that, The graph convolution module includes two feature extraction units, each of which includes a graph convolution layer and a batch normalization layer. The expression for the graph convolution module is: In the formula, and These are the feature representations learned from the first and second layer graph convolutions, respectively. It is an adjacency matrix. and These are the learnable parameters of the graph convolutional layer.

7. A graph neural network fault diagnosis device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1-6.

8. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the method as described in any one of claims 1-6.

Images