Rotating equipment fault diagnosis method fusing CNN (Convolutional Neural Network) and graph attention network

Abstract

The invention provides a rotating equipment fault diagnosis method fusing a CNN (Convolutional Neural Network) and a graph attention network. A plurality of sensors respectively capture key operation signals of each part of a motor, a bearing, a gear box, a coupler and a rotating load, the key operation signals are input to a multi-channel attention module through a GRU network and then feature sequences are output, the feature sequences are spliced into a two-dimensional feature matrix, the two-dimensional feature matrix is input to a CNN and mixed attention mechanism module, and fault features of each part are obtained; the fault features are windowed according to a time sequence and recombined after being dynamically weighted according to importance; constructing a global fault diagnosis graph by taking the recombination features of different parts as nodes and the significant interaction relationship between the nodes as edges; learning interaction information among the nodes through a graph attention network and aggregating the interaction information to obtain global fault features; and constructing a graph network joint optimization loss function, and analyzing and optimizing model parameters. According to the model, fault information is comprehensively mined through a multi-stage fusion mode, and the accuracy of fault diagnosis is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of hardware equipment fault diagnosis based on deep network models driven by big data, and specifically relates to a rotating equipment fault diagnosis method that integrates CNN and graph attention network. Background Art

[0002] As the "heart" of industrial production systems, rotating equipment is widely used in key areas such as aerospace, energy and electricity, and petrochemicals. Its stable operation is a core element to ensure production continuity, safety, and economy. However, rotating equipment operates under long-term high loads and complex working conditions, and is prone to various types of faults such as bearing wear, gear failure, and rotor imbalance. Once a fault occurs, it will not only cause equipment shutdown and production stagnation, but may also trigger a chain reaction, causing immeasurable economic losses and safety hazards. The fault characteristics of rotating equipment are complex and changeable, and the fault signals often show nonlinear and non-stationary characteristics. There is a coupling relationship between different fault modes, which significantly increases the difficulty of fault diagnosis. At the same time, during the operation of rotating equipment, multi-source heterogeneous data such as vibration, temperature, and current will be generated, and there will be complex spatial correlations and dynamic dependencies between the data. The rotating equipment fault diagnosis method that integrates CNN and graph attention network proposed in this application can comprehensively and deeply mine fault characteristics and realize accurate fault diagnosis of rotating equipment.

[0003] Currently, traditional fault diagnosis methods, such as those based on vibration spectrum analysis and threshold judgment, rely heavily on manual experience, resulting in low diagnostic efficiency and accuracy, making them difficult to meet diagnostic requirements under complex operating conditions. While convolutional neural networks (CNNs) in deep learning can automatically extract local features from fault data, they have limitations when processing the complex spatial relationships between multi-source data on rotating equipment. While graph attention networks can effectively capture correlations between data nodes, they struggle to deeply explore the local details of the data. Therefore, a fault diagnosis method that integrates CNNs and graph attention networks is urgently needed, leveraging the strengths of both to achieve rapid and accurate identification of rotating equipment faults. This has important practical implications for reducing equipment operation and maintenance costs and improving the level of intelligent industrial production. Summary of the Invention

[0004] The purpose of the present invention is to provide a more accurate and reliable method for diagnosing faults of rotating equipment, which can accurately discover faults during the operation of the rotating equipment and ensure the safe and stable operation of the equipment.

[0005] To solve the above technical problems, the present invention proposes a rotating equipment fault diagnosis method that integrates CNN and graph attention network, including: Use multiple sensors to collect key operating signals from different parts: ① Collect current, voltage, temperature and vibration signals from the motor; ② Collect vibration and temperature signals from the bearing; ③ Collect vibration and temperature signals from the gearbox; ④ Collect vibration, temperature and sound signals from the coupling; ⑤ Collect speed, temperature and vibration signals from the rotating load. Each signal is input into its own GRU network, and the hidden layer state of the GRU network at each time step is used as the input of a channel of the channel attention module. The multi-channel attention module weights different channels to generate feature vectors. The feature vectors generated by each channel are spliced ​​along the channel dimension, so that each row of the two-dimensional feature matrix corresponds to the feature vector of a channel, and each column corresponds to a feature point in the feature vector. The two-dimensional feature matrix is ​​input into the CNN and hybrid attention modules to deeply extract features and their weights. The four-layer CNN network uses multi-scale convolution kernels of 7*7, 3*3, 3*3 and 1*1 respectively, and a hybrid attention mechanism is connected after each CNN layer for weighting and local adjustment, finally obtaining refined fault features of different signals from various parts of the rotating equipment; The refined fault features of different signals of various parts of the rotating equipment are divided into N time windows according to their time characteristics at fixed time intervals, as shown in formula (1), where T is the total number of time steps, W is the window size, and S is the step size. For the i-th time window, the similarity between the features of two different time steps is calculated according to formula (2): , and then get the similarity matrix of the i-th time window , where j and k are two different time steps, f is the time step feature, and the weight matrix of the time window is obtained after normalization using the softmax function , the features in the time window are calculated according to the weight matrix according to formula (3) After weighting, local aggregation features are obtained ,in is the feature of the jth time step in the i-th window, and the Splicing in chronological order, and finally obtaining the time window splicing features of different signals in each part , where N is the number of windows and d is the feature dimension; 1 (1) (2) (3) The electric drive system is regarded as a graph topology, with its internal motors, bearings, gearboxes, couplings and rotating loads as nodes of the graph network, and the connections or coupling relationships between them as edges. The time window splicing features of different signals in each part are used as the node feature information of the graph network. , where: ① the motor node includes four signal features: current harmonic distortion rate, stator modal frequency response, voltage fluctuation, and temperature; ② the bearing node includes two signal features: inner and outer race fault characteristic frequencies and temperature; ③ the gearbox node includes two signal features: gear meshing frequency and temperature; ④ the coupling node includes three signal features: vibration, temperature, and sound; ⑤ the rotating load node includes three signal features: speed, temperature, and vibration. The degree of fault propagation between two nodes is used as side information to construct a global fault diagnosis graph; The features of each node are mapped into query value and key value vectors through a learnable linear transformation matrix. The shared attention mechanism function LeakyReLU is used to calculate the attention coefficient between different nodes, and the attention coefficient is modulated in combination with the side information. Then all the attention coefficients are normalized to obtain the attention weight. , and weight the node features, learn the weighted features of different nodes in different subspaces through the multi-head attention mechanism, and finally integrate the output of each head to obtain the global fault feature Z; Constructing a graph network to jointly optimize the loss function , analyze and adjust model parameters.

[0006] Optionally, the construction graph network jointly optimizes the loss function , analyze and adjust model parameters, including: ① Constructing a graph network to jointly optimize the loss function As shown in formula (4): (4) in is the classification loss function, is the global-local structure contrast loss function, ② calculate the global-local structure contrast loss function , as shown in formula (5): (5) Where Z is the global fault characteristic, is the node failure characteristic, is the weight coefficient, Reflects global fault characteristics and node characteristics The difference between the means, ③ calculate the classification loss function , as shown in formula (6): (6) Where C is the total fault category, For samples The encoding in dataset c, The probability that the model predicts the fault is category c, ④ According to the joint optimization loss function Update model parameters, including: update linear change matrix and scaling factor ,by For example, as shown in formula (7): (7) Where t is the number of model iterations, is the learning rate.

[0007] This application proposes a method for fault diagnosis of rotating equipment that integrates CNN and graph attention network. Multiple sensors capture key operating signals of motors, bearings, gearboxes, couplings and rotating loads respectively, and output feature sequences after being input into the multi-channel attention module through the GRU network. The feature sequences are spliced ​​into a two-dimensional feature matrix and input into the CNN and hybrid attention mechanism module to obtain the fault features of each part; the fault features are divided into windows according to the time series and reorganized after dynamic weighting according to the importance; the reorganized features of different parts are used as nodes and the significant interaction relationships between nodes are used as edges to construct a global fault diagnosis graph; the interaction information between nodes is learned and aggregated through the graph attention network to obtain the global fault features; the graph network is constructed to jointly optimize the loss function , analyze and optimize model parameters. This model comprehensively mines fault information through multi-stage fusion, greatly improving the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0009] Figure 1 A schematic flow chart of a rotating equipment fault diagnosis method integrating CNN and graph attention network provided in an embodiment of the present application; Figure 2 A model structure diagram provided for an embodiment of the present invention; Figure 3 A diagram of the graph attention algorithm framework provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0010] In order to enable those skilled in the art to better understand the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0011] like Figure 1 As shown, Figure 1 The present invention provides a flowchart of a method for fault diagnosis of rotating equipment that integrates CNN and graph attention network, which specifically includes five contents.

[0012] S11: Data is collected from multiple parts, and the refined fault features of different signals at each part are obtained after inputting the weighted multi-channel attention mechanism through the GRU network and then extracted by the CNN and hybrid attention mechanisms.

[0013] It's important to note that the raw data collected from different locations has varying scales and feature importance. Directly extracting fault features can lead to biased results. Therefore, prior to feature extraction, the raw data must be normalized to unify the scale. Next, the operational data from each device collected at each location is fed into a multi-channel attention module via a GRU network, with the hidden layer state of each GRU network time step serving as the input for each channel. Finally, the weighted concatenation of the features from each channel is used to mine the data's deepest layers using a CNN and hybrid attention module, thereby obtaining refined fault features for each location's distinct signals. The entire process consists of three steps.

[0014] Step 11: Normalize the collected raw data using formula (1): (1) Where t represents the time step, i represents the part, The maximum and minimum values ​​of all time points of the i part are input into the normalized data into the GRU network. The important data features are extracted and retained from the historical data through the reset gate and update gate, and the update gate state at the current moment is calculated with the input of the current state and the output of the hidden layer of the previous state. and reset the door state , as shown in formulas (2) and (3): (2) (3) According to the current status and Calculate the candidate hidden state output at this moment, as shown in formula (4): (4) Finally, combined with the hidden state of the previous time step and candidate hidden states , calculate the hidden state of the current time step, as shown in formula (5): (5) In the above formula Represents the weight matrix of model training, 、 、 represents the bias matrix, , tanh represent sigmoid and tanh activation functions respectively. After obtaining the hidden layer state of each GRU network at each time step Then use the multi-channel attention module to weight it to highlight the importance of data at different times. Calculate its attention score by formula (6) , all attention scores are normalized by the softmax function to obtain the attention weight , as shown in formula (7), the attention weight is finally combined with the corresponding hidden layer state Weighted according to formula (8) to obtain the weighted feature vector of each channel . (6) (7) (8)

[0015] Step 12: Concatenate the feature vectors generated by each channel along the channel dimension so that each row of the two-dimensional feature matrix corresponds to a feature vector of a channel, and each column corresponds to a feature point in the feature vector.

[0016] Step 13: Input the concatenated two-dimensional feature matrix into the module where the CNN and hybrid attention mechanism appear four times in a cycle. The first layer of the CNN network uses a 7*7 convolution kernel for convolution operation to obtain the feature map F. Then, the activation function Swish(x) = xsigmoid(x) is used to introduce nonlinearity to enhance the expressiveness of the model. After the convolution layer, the hybrid attention module is connected to weight and locally adjust the feature map through the parallel arrangement of channel attention and spatial attention. As shown below, the channel attention weight is calculated according to formulas (9), (10) and (11): , calculate the spatial attention weight according to formula (12) : (9) (10) (11) (12) in, 、 is the weight matrix, c is the channel label, The formula for calculating the feature map F1' is as follows (13): (13) In the second, third, and fourth CNN and hybrid attention mechanism modules, the above steps are repeated with 3*3, 3*3, and 1*1 convolution kernels respectively to obtain the refined fault feature sequences of different signals in each part. , where T=1000.

[0017] Based on the above theory, in an optional embodiment of the present application, 4 layers of CNN and hybrid attention modules are arranged in series. From front to back, each layer of CNN uses convolution kernels of 7*7, 3*3, 3*3, and 1*1 sizes, and finally obtains a refined fault feature sequence of 1000 time steps. .

[0018] S12: Divide the fault features into windows according to the time series and reorganize them after weighting based on dynamics.

[0019] It should be noted that CNN and the hybrid attention module do not focus on the temporal characteristics of the data, but the regular movement of the convolution kernel retains the temporal order of the features to a certain extent. After the windowed and weighted reorganization, it is convenient to capture all kinds of feature information in the subsequent all-round way. The whole process mainly has three steps.

[0020] Step 21: For the obtained refined fault feature sequence , divided into N time windows of size W and step size S=50 according to formula (14), where the i-th time window can be expressed as . 1 (14)

[0021] Step 22: For each time window The similarity between the features of two different time steps is calculated by formula (15) , and then get the similarity matrix of the i-th time window , where j and k are two different time steps, f is the time step feature, and the similarity matrix Each element in is normalized using the softmax function as shown in formula (16), and then the weight matrix of different time windows is obtained. , and calculate the mean of the corresponding weights according to formula (17) to get the weight of each time step , where i is the time window number and j is the time step. (15) (16) (17)

[0022] Step 23: For each feature in the time window, the corresponding weight matrix After weighting according to formula (18), the local feature is obtained , and all time windows Splice them in chronological order and finally get the time window splicing features . (18)

[0023] Based on the above theory, in an optional embodiment of the present application, each refined fault feature is divided into 20 time windows, and the time window splicing feature is obtained after weighting and reorganization. .

[0024] S13: Construct a global fault diagnosis graph using weighted reorganized fault features as nodes and significant interaction relationships between nodes as edges.

[0025] It should be noted that there are direct or indirect connections or coupling relationships between the motor, bearings, gearbox, coupling and rotating load of the point drive system. Constructing them into a graph network is conducive to analyzing the relationship between the signal characteristics of each part and facilitating the global analysis of fault characteristics. The whole process mainly has two steps.

[0026] Step 31: Consider the electric drive system as a graph topology, with its internal motors, bearings, gearboxes, couplings, and rotating loads as nodes of the graph network, and the connections or coupling relationships between them as edges. The time window splicing features of different signals in each part are used as the node feature information of the graph network. , among which: ① the four signal features of current harmonic distortion rate, stator modal frequency response, voltage fluctuation and temperature are used as the characteristic information of the motor node; ② the two signal features of inner and outer race fault characteristic frequency and temperature are used as the characteristic information of the bearing node; ③ the two signal features of gear meshing frequency and temperature are used as the characteristic information of the gearbox node; ④ the three signal features of vibration, temperature and sound are used as the characteristic information of the coupling node; ⑤ the three signal features of speed, temperature and vibration are used as the characteristic information of the rotating load node.

[0027] Step 32: Based on the historical operating data of the equipment and the operating conditions of the system, define the degree of fault propagation between the five nodes of motor, bearing, gearbox, coupling and rotating load. , as the side information of the graph network.

[0028] Based on the above theory, in an optional embodiment of the present application, the number of graph nodes is 5, among which the motor node includes 4 signal characteristics of current harmonic distortion rate, stator modal frequency response, voltage fluctuation and temperature; the bearing node includes 2 signal characteristics of inner and outer ring fault characteristic frequency and temperature; the gearbox node includes 2 signal characteristics of gear meshing frequency and temperature; the coupling node includes 3 signal characteristics of vibration, temperature and sound; the rotating load node includes 3 signal characteristics of speed, temperature and vibration, and the degree of fault propagation influence between two nodes with direct physical connection is calculated. =0.7, the impact of fault propagation between two nodes without direct physical connection =0.15.

[0029] S14: The fault information between different nodes is learned and aggregated through the graph attention network to obtain the global fault features.

[0030] It should be noted that the weighted fault features of each node may have one or more correlations. This application selects a multi-head attention mechanism to capture the various connections between different node features. The whole process mainly includes three steps.

[0031] Step 41: Obtain the node feature set V from the global fault diagnosis graph as the input matrix , where N is the number of nodes, D is the feature dimension of each node, and a shared linear change matrix is ​​used , map the node features into Query and Key vectors according to formulas (19) and (20): (19) (20) And as shown in formula (21): (twenty one) Calculate the similarity between nodes as the original attention coefficient , where i and j represent two different nodes, It is a scaling factor to prevent the dot product result from being too large and causing the gradient to disappear.

[0032] Step 42: Use the degree of fault propagation between two nodes as the original attention coefficient The modulation factor is calculated according to formula (22) to obtain the modulated attention coefficient . (twenty two)

[0033] Step 43: Use the LeakyRelu function to perform nonlinear transformation on the original coefficients to alleviate the gradient disappearance, as shown in formula (23), where is the weight coefficient: (twenty three) And normalize the rows to convert the original attention coefficient into the attention weight in the form of probability distribution, and get , represents the normalized attention weight of node i to node j, and finally the attention weight matrix is ​​obtained .

[0034] Step 44: Use the attention weight matrix B to weight the node features to obtain the weighted features , and the weighted feature Repeat steps 41 and 42 by performing parallel processing through h independent heads, where each head corresponds to an independent linear transformation matrix , obtain the independent Query value, Key value and Value value of each head, and then obtain different weight matrices. Different heads learn different fault information and connections between nodes. Finally, the outputs of each head are spliced ​​and reduced in dimension through linear transformation to obtain the final global fault feature Z.

[0035] Based on the above discussion, in an optional embodiment of the present application, the multi-head attention mechanism has a total of 6 heads, and the coefficients in the LeakyRelu function are The value of is 0.01.

[0036] S15: Construct a graph network to jointly optimize the loss function, analyze and adjust the model parameters.

[0037] It should be noted that the global fault features learned by the graph attention network may differ significantly from the fault features of each node, and there may be certain errors in the fault feature classification. Because the graph network is constructed to jointly optimize the loss function, it is possible to calculate the loss size and optimize the model parameters based on it, thereby improving the performance of the model fault diagnosis and classification. The whole process mainly consists of two steps:

[0038] Step 51: Constructing a graph network joint optimization loss function , as shown in formula (24): (twenty four) in, As the classification loss function, the binary cross entropy is extended to the multi-category scenario, which can measure the difference between the predicted probability distribution and the true distribution. The calculation formula is shown in formula (25): (25) Where C is the total fault category, For samples The encoding in dataset c, The probability that the model predicts the fault to be category c; is a global-local structural loss function, which can calculate whether the mean of the global fault representation and the local node feature are consistent, avoiding the loss of fine-grained fault information during feature aggregation. The calculation formula is shown in Equation (26): (26) Where Z is the global fault characteristic, is the node failure characteristic, is the weight coefficient.

[0039] Step 52: Jointly optimize the loss function based on the calculated graph network , first calculate according to formula (27) and (28) Linear change matrix (by for example) and the scaling factor (The initial value is )’s gradient: (27) (28) in is the weight coefficient of the classification loss, according to the joint optimization loss function Update the model parameters according to formula (29) for the gradient of each parameter 、 (with linear change matrix for example). (29)

[0040] Based on the above theory, calculate the global-local structure loss function middle The general value of is between 0.5 and 1.0. Linear change matrix In the gradient The value of is 0.05.

[0041] Specific examples are used in this application to illustrate the principles and implementation methods of the present invention. The description of the above examples is only intended to help explain the method and core concept of the present invention. It should be noted that for ordinary people in this technical field, various improvements and modifications can be made to the present invention without departing from the principles of the present invention, and such improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A rotating equipment fault diagnosis method integrating CNN and graph attention network, characterized by: include: Use multiple sensors to collect key operating signals from different parts: ① Collect current, voltage, temperature and vibration signals from the motor; ② Collect vibration and temperature signals from the bearing; ③ Collect vibration and temperature signals from the gearbox; ④ Collect vibration, temperature and sound signals from the coupling; ⑤ Collect speed, temperature and vibration signals from the rotating load. Each signal is input into its own GRU network, and the hidden layer state of the GRU network at each time step is used as the input of a channel of the channel attention module. The multi-channel attention module weights different channels to generate feature vectors. The feature vectors generated by each channel are spliced ​​along the channel dimension, so that each row of the two-dimensional feature matrix corresponds to the feature vector of a channel, and each column corresponds to a feature point in the feature vector. The two-dimensional feature matrix is ​​input into the CNN and hybrid attention modules to deeply extract features and their weights. The four-layer CNN network uses multi-scale convolution kernels of 7*7, 3*3, 3*3 and 1*1 respectively, and a hybrid attention mechanism is connected after each CNN layer for weighting and local adjustment, finally obtaining refined fault features of different signals from various parts of the rotating equipment; The refined fault features of different signals of various parts of the rotating equipment are divided into N time windows according to their time characteristics at fixed time intervals, as shown in formula (1), where T is the total number of time steps, W is the window size, and S is the step size. For the i-th time window, the similarity between the features of two different time steps is calculated according to formula (2): , and then get the similarity matrix of the i-th time window , where j and k are two different time steps, f is the time step feature, and the weight matrix of the time window is obtained after normalization using the softmax function , the features in the time window are calculated according to the weight matrix according to formula (3) After weighting, local aggregation features are obtained ,in is the feature of the jth time step in the i-th window, and the Splicing in chronological order, and finally obtaining the time window splicing features of different signals in each part , where N is the number of windows and d is the feature dimension; 1（1） （2） （3） The electric drive system is regarded as a graph topology, with its internal motor, bearings, gearbox, coupling, and rotating load as nodes of the graph network, and the connections or coupling relationships between them as edges. The time window splicing features of different signals in each part are used as the node feature information of the graph network, among which: ① The motor node includes four signal features: current harmonic distortion rate, stator modal frequency response, voltage fluctuation, and temperature; ② The bearing node includes two signal features: inner and outer race fault characteristic frequency and temperature; ③ The gearbox node includes two signal features: gear meshing frequency and temperature; ④ The coupling node includes three signal features: vibration, temperature, and sound; ⑤ The rotating load node includes three signal features: speed, temperature, and vibration. The degree of fault propagation influence between two nodes is used as edge information to construct a global fault diagnosis graph; The features of each node are mapped into query value and key value vectors through a learnable linear transformation matrix. The shared attention mechanism function LeakyReLU is used to calculate the attention coefficient between different nodes, and the attention coefficient is modulated in combination with the side information. Then all the attention coefficients are normalized to obtain the attention weight. , and weight the node features, learn the weighted features of different nodes in different subspaces through the multi-head attention mechanism, and finally integrate the output of each head to obtain the global fault feature Z; Constructing a graph network to jointly optimize the loss function , analyze and adjust model parameters.

2. The rotating equipment fault diagnosis method integrating CNN and graph attention network according to claim 1, characterized in that: Constructing a graph network to jointly optimize the loss function , analyze and adjust model parameters, including: ① Constructing a graph network to jointly optimize the loss function As shown in formula (4): （4） in is the classification loss function, is the global-local structure contrast loss function, ② calculate the global-local structure contrast loss function , as shown in formula (5): （5） Where Z is the global fault characteristic, is the node failure characteristic, is the weight coefficient, Reflects global fault characteristics and node characteristics The difference between the means, ③ calculate the classification loss function , as shown in formula (6): （6） Where C is the total fault category, For samples The encoding in dataset c, The probability that the model predicts the fault is category c, ④ According to the joint optimization loss function Update model parameters, including: update linear change matrix and scaling factor ,by For example, as shown in formula (7): （7） Where t is the number of model iterations, is the learning rate.

Images