Multi-sensor-based graph attention time sequence fusion network fault diagnosis method

Abstract

The invention discloses a multi-sensor-based graph attention time sequence fusion network fault diagnosis method, relates to the technical field of fault diagnosis, and aims to solve the problem of low fault diagnosis accuracy caused by the fact that a single-sensor signal can only identify abnormality of a time dimension and cannot diagnose abnormality of a space dimension in the prior art. According to the method, the multi-sensor signal nodes are constructed by using the graph neural network to connect data from different sensors, then the attention mechanism is embedded into the graph neural network to perform effective screening of data weights, and finally data fusion is performed by using the adaptive weighting strategy to improve the efficiency of multi-sensor information interaction. According to the method, the time sequence information and the space correlation information are effectively extracted by fusing the time sequence information and the space correlation information and inputting the fused new data into the LSTM to extract the time domain information and then performing adaptive weighting, and the time sequence information and the space correlation information are subjected to data fusion again by using an adaptive weighting strategy, so that the anomaly of the space dimension can be diagnosed, and the accuracy of data fusion is improved. And the accuracy of fault diagnosis is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a multi-sensor based graph attention temporal fusion network fault diagnosis method. Background Art

[0002] Rotating machinery such as gears and bearings plays a very important role in modern industrial production. Since rotating machinery is generally exposed to extreme working environments, various types of faults will occur frequently. Condition monitoring of rotating machinery has become an important task in equipment operation and maintenance. In order to improve equipment reliability and reduce maintenance costs, it is necessary to develop intelligent fault diagnosis technology for rotating machinery.

[0003] Most of the current intelligent fault diagnosis technologies are based on artificial intelligence algorithms such as machine learning and deep learning, and are trained on vibration signal data collected by sensors to obtain fault diagnosis models. However, most methods only use single sensor signals, which can only identify anomalies in the time dimension but cannot diagnose anomalies in the spatial dimension, resulting in low fault diagnosis accuracy. Summary of the invention

[0004] The purpose of the present invention is to propose a multi-sensor based graph attention temporal fusion network fault diagnosis method to address the problem that a single sensor signal in the prior art can only identify anomalies in the time dimension but cannot diagnose anomalies in the spatial dimension, thus resulting in low fault diagnosis accuracy.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] A multi-sensor based graph attention temporal fusion network fault diagnosis method comprises the following steps:

[0007] The vibration data of the gearbox is collected by multiple sensors, and the vibration data of the gearbox collected by multiple sensors is input into the trained graph attention time series fusion network to obtain the output fault classification result;

[0008] The graph attention temporal fusion network specifically performs the following steps:

[0009] Step 1: Use multiple sensors to collect the vibration data of the gearbox, and use the sensors as nodes and the vibration data as node attributes to input the graph attention network to obtain the output feature h of the graph attention network at the current moment i ', i = 1, 2, ... t;

[0010] Step 2: Output features h of the graph attention network at all times i 'Input the SE module and LSTM respectively to obtain the adaptively weighted feature h output by the SE module at each moment'all And the value h" of LSTM output at each moment i ;

[0011] Step 3: h' at each moment all and h” i Add the values, then input the sum of all the values ​​at all times into the SE module to obtain the adaptively weighted feature h' at each moment a ' ll , and finally h' a ' ll After entering the fully connected layer, the classification result is obtained through softmax.

[0012] Furthermore, the specific steps of step one are:

[0013] Step 1: Take each sensor as a node to construct the graph G t , the graph G t It is expressed as:

[0014] G t =(V,E,X) t

[0015] Wherein, V represents the set of node numbers, E represents the set of boundary numbers, X represents the time point data of the node, and t represents the time sum of the vibration data.

[0016] Step 1 and 2: Use multiple sensors to collect vibration data of the gearbox, and normalize the vibration data by maximum and minimum values;

[0017] Step 13: Apply the ε radius criterion to determine the neighbor nodes of each node through the cosine similarity between two nodes, and connect the node with the neighbor node;

[0018] Step 14: Calculate the similarity coefficient between the node and each of its corresponding neighbor nodes respectively, and based on the similarity coefficient, obtain the attention coefficient between the nodes. Then use the time point data of the node as a feature, linearly weight the attention coefficient, and obtain the output feature h' of the graph attention network from the current moment i .

[0019] Further, the similarity coefficient is expressed as:

[0020] e ij =a([Wn i ||Wn j ]),j∈N i

[0021] Among them, || represents the feature connection operation, a represents the mapping, W represents the weight matrix, and n iRepresents node features, n j represents the neighbor node characteristics, N i represents the total number of neighbor nodes, j represents the sequence number of the neighbor node, e ij Represents the similarity coefficient between the central node and another node.

[0022] Furthermore, the attention coefficient is expressed as:

[0023]

[0024] Among them, k represents each node, e ik Represents the similarity coefficient between the central node and each other node.

[0025] Furthermore, the output feature h of the graph attention network i ' is expressed as:

[0026]

[0027] Among them, σ represents the sigmoid activation function.

[0028] Furthermore, the fusion feature h a ' ll It is expressed as:

[0029]

[0030] s n =σ(W2δ(W1z n ))

[0031]

[0032] Among them, R represents the number of elements in each graph attention output matrix, δ represents the relu function, W1 and W2 represent the coefficients of the fully connected layer, Represents an operation symbol, which means that the data from each graph attention network is multiplied by the weight obtained by adaptive weighting. n represents the weight matrix, z n Represents the aggregation of graph attention outputs at different moments.

[0033] Furthermore, the fusion feature h' a ' ll It is expressed as:

[0034]

[0035] s n =σ(W2δ(W1z n ))

[0036]

[0037] Furthermore, the value h" output by the LSTM i It is expressed as:

[0038] h” i =o t *tanh(C t )

[0039] o t =σ(W o [h t-1 ,x t ]+b o )

[0040]

[0041] i t =σ(W i ·[h t-1 ,x t ]+b t )

[0042] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0043] Among them, W f is the weight of the forget gate, h t-1 is the input of the previous neuron, x t is the data input, b f is the bias value of the forget gate, f t is the output of the forget gate, W i is the weight of the input gate, h t-1 is the input of the previous neuron, x t is the data input, b t is the bias value of the input gate, i t is the output of the input gate, W C is the weight of the candidate value, h t-1 is the input of the previous neuron, x t is the data input, b C is the deviation value of the candidate value, is the output of the candidate value, C t-1 is the output of the previous candidate value, C t is the updated value of the candidate value, W o is the weight of the output gate, h t-1 is the input of the previous neuron, x t is the data input, b o is the deviation value of the output gate, ot is the output of the output gate, h t is the value passed to the next neuron.

[0044] The beneficial effects of the present invention are:

[0045] This application uses graph neural networks to construct multi-sensor signal nodes to connect data from different sensors, then embeds an attention mechanism into the graph neural network to effectively screen data weights, and finally uses an adaptive weighting strategy to fuse data to improve the efficiency of multi-sensor information interaction. The fused new data is then input into the LSTM to extract time domain information, and then adaptive weighting is performed to effectively extract the timing information and spatial correlation information. This application uses an adaptive weighting strategy to fuse the timing information and spatial correlation information again, so that the application can diagnose anomalies in the spatial dimension, further improving the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a diagram of the neural network structure in this application;

[0047] Figure 2 Calculate the attention coefficient graph in this application;

[0048] Figure 3 For the attention structure diagram in this application;

[0049] Figure 4 This is the LSTM structure diagram in this application;

[0050] Figure 5 Schematic diagram of LSTM input for GAT in this application;

[0051] Figure 6 Schematic diagram of GAT adaptive weighted fusion at different times in this application;

[0052] Figure 7 Schematic diagram of the adaptive graph attention temporal fusion algorithm in this application;

[0053] Figure 8 This is a flow chart of the adaptive graph attention temporal fusion network diagnosis method in this application;

[0054] Fig. 9 This is the confusion matrix diagram in this application;

[0055] Fig.10 This is a schematic diagram of the overall process of this application. DETAILED DESCRIPTION

[0056] It should be particularly noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.

[0057] Specific implementation method 1: A multi-sensor based graph attention temporal fusion network fault diagnosis method described in this implementation method includes the following steps:

[0058] The vibration data of the gearbox is collected by multiple sensors, and the vibration data of the gearbox collected by multiple sensors is input into the trained graph attention time series fusion network to obtain the output fault classification result;

[0059] The graph attention temporal fusion network specifically performs the following steps:

[0060] Step 1: Use multiple sensors to collect the vibration data of the gearbox, and use the sensors as nodes and the vibration data as node attributes to input the graph attention network to obtain the output feature h of the graph attention network at the current moment i ', i = 1, 2, ... t;

[0061] Step 2: Output features h of the graph attention network at all times i 'Input the SE module and LSTM respectively to obtain the adaptively weighted feature h output by the SE module at each moment' all And the value h" of LSTM output at each moment i ;

[0062] Step 3: h' at each moment all and h” i Add the values, then input the sum of all the values ​​at all times into the SE module to obtain the adaptively weighted feature h' at each moment a ' ll , and finally h' a ' ll After entering the fully connected layer, the classification result is obtained through softmax.

[0063] This application aims to use graph neural networks to construct multi-sensor signal nodes to connect data from different sensors, then embed an attention mechanism into the graph neural network to effectively screen data weights, and finally use an adaptive weighting strategy to fuse data to improve the efficiency of multi-sensor information interaction. The fused new data is then input into LSTM to extract time domain information, and then adaptive weighting is performed to further improve the accuracy of multi-sensor signal fault diagnosis.

[0064] Specific implementation method 2: This implementation method is a further description of the specific implementation method 1. The difference between this implementation method and the specific implementation method 1 is that the specific steps of step 1 are:

[0065] Step 1: Take each sensor as a node to construct the graph G t , the graph G t It is expressed as:

[0066] G t =(V,E,X) t

[0067] Wherein, V represents the set of node numbers, E represents the set of boundary numbers, X represents the time point data of the node, and t represents the time sum of the vibration data.

[0068] Step 1 and 2: Use multiple sensors to collect vibration data of the gearbox, and normalize the vibration data by maximum and minimum values;

[0069] Step 13: Apply the ε radius criterion to determine the neighbor nodes of each node through the cosine similarity between two nodes, and connect the node with the neighbor node;

[0070] Step 14: Calculate the similarity coefficient between the node and each of its corresponding neighbor nodes respectively, and based on the similarity coefficient, obtain the attention coefficient between the nodes. Then, use the time point data of the node as a feature, perform linear weighting on the attention coefficient, and obtain the output feature h from the current moment to the graph attention network. i '.

[0071] Specific implementation method three: This implementation method is a further description of specific implementation method two. The difference between this implementation method and specific implementation method two is that the similarity coefficient is expressed as:

[0072] e ij =a([Wn i ||Wn j ]),j∈N i

[0073] Among them, || represents the feature connection operation, a represents the mapping, W represents the weight matrix, and n i Represents node features, n j represents the neighbor node characteristics, N i represents the total number of neighbor nodes, j represents the sequence number of the neighbor node, e ij Represents the similarity coefficient between the central node and another node.

[0074] Specific implementation method 4: This implementation method is a further explanation of specific implementation method 3. The difference between this implementation method and specific implementation method 3 is that the attention coefficient is expressed as:

[0075]

[0076] Among them, k represents each node, e ik Represents the similarity coefficient between the central node and each other node.

[0077] Specific implementation mode 5: This implementation mode is a further explanation of specific implementation mode 4. The difference between this implementation mode and specific implementation mode 4 is that the output feature h of the graph attention network i ' is expressed as:

[0078]

[0079] Among them, σ represents the sigmoid activation function.

[0080] Specific implementation method 6: This implementation method is a further explanation of specific implementation method 5. The difference between this implementation method and specific implementation method 5 is that the fusion feature h a ' ll It is expressed as:

[0081]

[0082] s n =σ(W2δ(W1z n ))

[0083]

[0084] Among them, R represents the number of elements in each graph attention output matrix, δ represents the relu function, W1 and W2 represent the coefficients of the fully connected layer, Represents an operation symbol, which means that the data from each graph attention network is multiplied by the weights obtained by adaptive weighting. n represents the weight matrix, z n Represents the aggregation of graph attention outputs at different moments.

[0085] Specific implementation method 7: This implementation method is a further explanation of specific implementation method 6. The difference between this implementation method and specific implementation method 6 is that the fusion feature h' a ' ll It is expressed as:

[0086]

[0087] s n =σ(W2δ(W1z n ))

[0088]

[0089] Specific implementation method eight: This implementation method is a further explanation of specific implementation method seven. The difference between this implementation method and specific implementation method seven is that the value h output by the LSTM is i It is expressed as:

[0090] h” i =ot *tanh(C t )

[0091] o t =σ(W o [h t-1 ,x t ]+b o )

[0092]

[0093] i t =σ(W i ·[h t-1 ,x t ]+b t )

[0094] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0095] Among them, W f is the weight of the forget gate, h t-1 is the input of the previous neuron, x t is the data input, b f is the bias value of the forget gate, f t is the output of the forget gate, W i is the weight of the input gate, h t-1 is the input of the previous neuron, x t is the data input, b t is the bias value of the input gate, i t is the output of the input gate, W C is the weight of the candidate value, h t-1 is the input of the previous neuron, x t is the data input, b C is the deviation value of the candidate value, is the output of the candidate value, C t-1 is the output of the previous candidate value, C t is the updated value of the candidate value, W o is the weight of the output gate, h t-1 is the input of the previous neuron, x t is the data input, b o is the deviation value of the output gate, o t is the output of the output gate, h t is the value passed to the next neuron.

[0096] In order to more comprehensively understand the operating status of rotating machinery, various sensors are deployed at various locations of rotating machinery. These sensors can form a network that integrates multi-source data through graph neural networks, from which more structured graph data can be obtained. Each sensor is a node, and the graph can be represented as:

[0097] U=(V,E) (1)

[0098] Where V represents the set of node numbers, and E represents the set of edge numbers.

[0099] The number of nodes can be expressed as n = |V|. For fault diagnosis, data from different sensors can be viewed as multivariate time series. At this point, the sensor can be constructed as a single node of the graph neural network. However, since the sensor data contains certain time information, the properties of the node will change dynamically along the time scale. So the graph is further expressed as:

[0100] G t =(V,E,X) t (2)

[0101] Where X∈R n×d The time point data of the node.

[0102] The detailed graph construction process is as follows Figure 1 As shown, the data output by the sensor is first normalized by the maximum and minimum values. Then the signal data from the sensor forms a graph node in the form of a fixed time. After obtaining a single node, it is necessary to determine the neighbors of each node. At this time, the ε radius criterion can be applied, which can be measured by the cosine similarity between two nodes. The selection of neighbor points is expressed as follows:

[0103]

[0104] is the ε radius criterion. Among them, Ne(n i ) represents the neighbors of a single node n, γ is the set of nodes, sim is the cosine similarity, and ε represents the threshold.

[0105] Applying this criterion, we can form a boundary by counting the correlations from different sensors. The correlation weight W between two nodes i,j It can be expressed by Gaussian kernel weight function, as shown in formula 4:

[0106]

[0107] Where β is the bandwidth variance of the Gaussian function, n i is the node feature, n jis the node feature of the neighbor. By using the node neighbors and boundary weights, we can get the n×n neighbor matrix. So far, the GNN layer is built, and then we can calculate the attention coefficient to assign the corresponding weights to get the GAT layer.

[0108] Construct the correlation matrix W using all correlation weights

[0109] Before calculating the attention coefficient, the similarity coefficient between each node and its neighboring nodes is counted, as shown in Formula 5.

[0110] e ij =a([Wn i ||Wn j ]),j∈N i (5)

[0111] Among them, || represents the feature connection operation, and the mapping a and weight matrix W are learnable parameters.

[0112] Subsequently, the activation functions required to calculate the attention coefficient are LeakyReLu and Softmax. The LeakyRelu function is a variant of the standard ReLU function. It contains a new parameter n that allows negative values ​​to pass a small amount instead of being set to zero. Therefore, the LeakyRelu function is easier to activate neurons than the traditional relu function, allowing the model to converge faster, thereby improving the performance of the model. The Softmax function standardizes the input data and sets the sum of all its outputs to 1, which is more beneficial to the calculation of attention weights during model training. The calculation of the attention coefficient is shown in Equation 6. The entire calculation process of calculating the attention coefficient between nodes is shown in Equation 6. Figure 2 shown.

[0113]

[0114] After calculating the attention coefficient, the time point data of each node is used as a feature, and the attention coefficient is linearly weighted. At this time, the new feature of the central node is expressed as:

[0115]

[0116] Among them, h i ' is the output feature of the graph attention network (GAT), and σ represents the sigmoid activation function.

[0117] At this point, the traditional graph structure network has introduced the attention coefficient to reasonably and effectively allocate the weights between different nodes representing different sensors, and further realized the screening of related information between different sensors, thereby forming GAT.

[0118] After processing the output value of GAT, let’s consider how to obtain the time series information in the multi-sensor signal data. The signal in the time series establishes the node attributes in the graph structure. Since the LSTM network has a strong representation ability for the analysis of time series, this paper uses the LSTM network to model the time dependency.

[0119] h i 'Input LSTM and get LSTM output h t

[0120] In this application, the signals from multiple sensors are sliced ​​simultaneously and fed into the LSTM unit to calculate the implicit state. The difference between the LSTM network and the traditional recurrent neural network lies in the three "gates" of a single neuron. The "gate" consists of a sigmoid function and a dot product operation, and the sigmoid function outputs a value between 0 and 1, describing how many components can pass through, such as Figure 4 As shown, the three "gates" are the forget gate, input gate, and output gate, corresponding to 1, 2, and 3 respectively.

[0121] The first step of LSTM is to decide what information to extract from the neuron. This process is completed by forget gate 1, which is activated by the sigmoid function to obtain a value output between 0 and 1, as shown in the formula.

[0122] f t =σ(W f ·[h t-1 ,x t ]+b f ) (8)

[0123] Where: W f is the weight of the forget gate; h t-1 is the input from the previous neuron; x t is data input; b f is the bias value of the forget gate; f t Output of the forget gate.

[0124] The second step is to decide what information to store in the neuron. First, the input gate 2 decides which values ​​to update, as shown in the formula. On the other hand, the vector of new candidate values ​​is created through the tanh activation function. Will and i t Perform multiplication as shown.

[0125] i t =σ(W i ·[h t-1 ,x t ]+b t ) (9)

[0126] Where: Wi is the weight of the input gate; h t-1 is the input from the previous neuron; x t is data input; b t is the bias value of the input gate; i t is the output of the input gate.

[0127]

[0128] Where: W C is the weight of the candidate value; h t-1 is the input from the previous neuron; x t is data input; b C is the deviation value of the candidate value; The output of the candidate value.

[0129] Next, the scaling is performed according to the scaling degree of the decision to update each state value. It can be understood that the operation is where some old information is deleted and new information is added, as shown in the formula.

[0130]

[0131] Where: C t-1 is the output of the previous candidate value; C t The updated value of the candidate value.

[0132] In the third step, the output gate 3 determines which part of the neuron state is output, as shown in the formula. Then, the neuron state is converted into a value between -1 and 1 through the tanh function, and then multiplied by the output of the output gate, as shown in the formula.

[0133] o t =σ(W o [h t-1 ,x t ]+b o ) (12)

[0134] h t =o t *tanh(C t ) (13)

[0135] Where: W o is the weight of the output gate; h t-1 is the input from the previous neuron; x t is data input; b o is the deviation value of the output gate; o t is the output of the output gate; h t is the value passed to the next neuron.

[0136] This application uses a layer of LSTM network. First, the continuous time series signal is cut into samples of fixed length, and then the samples are embedded into nodes using a graph attention structure. The time series signal from a single sensor is processed by an LSTM network containing a single hidden layer. This LSTM architecture captures the time dependency of a single sensor signal and retains the time series information in a single source signal. Then the LSTM hidden states of each time step are connected to form a structured LSTM network. In this way, the graph attention structure mentioned above can be input into the LSTM, and each sensor in the graph attention structure embedded in each time step of the LSTM is represented as a node. As a result, the time features derived from the LSTM are associated with each node in the graph attention structure, and the signal data from different sensors are screened for valid data and information extraction of time series through graph attention and LSTM.

[0137] So far, the graph attention structure and LSTM have only been used to extract time series information. Although the number of nodes represented by sensors in the graph attention structure at different times is fixed, the weight distribution between nodes may not be consistent due to different time scales. Therefore, LSTM can only obtain time-related information, and cannot simultaneously obtain the spatial information output by each graph attention structure at different time scales (i.e., the amplitude characteristic information of different sensor data itself). Regarding how to fuse different output data in the graph attention network at different time scales, this paper uses an adaptive weighting strategy to process it.

[0138] In order to improve the quality of fused data, a squeeze and excitation SE module is constructed to further assign weights to the output data of the graph attention network at different time scales, such as Figure 6 As shown. Now we get the graph attention output h′ at a single time t it , use average pooling to formally aggregate the graph attention outputs at different times:

[0139]

[0140] Among them, R is the number of elements in each graph attention output matrix. Then, the weight matrix is:

[0141] s n =σ(W2δ(W1z n )) (15)

[0142] Among them, δ is the relu function; σ is the sigmoid function, W1 and W2 are the coefficients of the two fully connected layers, which are learnable parameters in the process of training the model.

[0143] Finally, the new data fusion obtained through the adaptive weighting strategy is:

[0144]

[0145] in, It is an operation symbol, which means that the data from each graph attention network is multiplied by their weights obtained through adaptive weighting.

[0146] The graph attention temporal fusion network trained in this application is trained through a loss function, which is expressed as:

[0147] y is the true category and p is the probability of predicting it to be correct.

[0148] Through the adaptive weighting strategy, the output values ​​of the graph attention from different time scales are deeply fused, and the new data obtained incorporates the spatial correlation information of the sensor data at different times, and the time series information can be obtained through the LSTM operation mentioned above. In order to extract more effective information from the time series information and spatial correlation information, make data fusion more efficient and easier to calculate, it is necessary to assign different weights to the two types of information, which requires the time series information and spatial correlation information to be fused again using the adaptive weighting strategy, such as Figure 7 As shown, the final fused data can be obtained.

[0149] Based on the above ideas, the idea of ​​the adaptive graph attention temporal fusion algorithm proposed in this application is as follows Figure 8 shown.

[0150] This application first uses the graph attention network structure to automatically discover and understand key features in large-scale sensor data. By establishing a graph model, using sensor data as nodes and the relationships between different sensors as edges, graph attention can model the interactions between different sensor data. In this way, the importance weights between each sensor node can be automatically learned using the graph attention mechanism, so as to better capture fault-related features. Secondly, graph attention can also help reduce data dimensions and reduce computational complexity. In large-scale sensor data, there is often a large amount of redundant and irrelevant information. By applying graph attention, you can selectively focus on sensor data related to the fault, thereby reducing data dimensions and reducing computational complexity. This can not only improve the efficiency of fault diagnosis, but also reduce the demand for storage and computing resources.

[0151] Secondly, LSTM is a recurrent neural network that can process sequence data. Since there is temporal correlation between different sensor data, this algorithm combines graph attention with LSTM to enable the model to better understand the correlation between nodes when processing time series data, and to better capture time information, and further discover more patterns and trends in time series data.

[0152] Furthermore, since the output values ​​of the graph attention network at different times not only have temporal correlation, but also have spatial correlation with different weights assigned to nodes at different times, the adaptive weighting strategy is used to weight the graph attention output at each time scale, so as to obtain spatial correlation data between multiple sensors at different times. After obtaining the spatial correlation data and temporal correlation data, the two are further weighted using the adaptive weighting strategy, and finally the highest quality fusion data based on this algorithm is obtained, which further improves the accuracy of fault diagnosis. The results are shown in Figure 2. Fig. 9 shown.

[0153] Table 1 Algorithm evaluation

[0154]

[0155] refer to Fig.10 The implementation steps of the present invention are described in detail:

[0156] (1) Sample collection: Use multiple sensors to collect gearbox vibration signals.

[0157] (2) Dataset division: 80% of the 3280 samples are used as training sets to train the neural network parameters, 50% of the remaining 20% ​​are used as validation sets to verify the feasibility of the training model, and the remaining 10% are used as test sets for the estimation test of the final model. Finally, 2624 training sets, 328 validation sets, and 328 test sets are obtained.

[0158] (3) Dataset standardization: One of the conditions that deep learning must meet is that the training set, validation set, and test set must be under the same distribution. Therefore, the mean and variance of the training set must be extracted to achieve training set standardization, and the mean and variance of the training set must be used to standardize the validation set and test set. In this way, the three data sets satisfy the same distribution.

[0159] (4) Model training: Build a graph attention network and adaptively weight the GAT at different times. On the other hand, use LSTM to extract temporal features of the GAT at different times. The output value and the adaptively weighted output result are adaptively weighted again, and the output model is trained iteratively.

[0160] (5) Model evaluation: The final model is put into the test set for evaluation, and the loss graph and accuracy graph of the training set and validation set are output. The degree of fault classification of the test set is observed through the confusion matrix.

[0161] It should be noted that the specific implementation is only an explanation and description of the technical solution of the present invention, and cannot be used to limit the scope of protection of the rights. Any partial changes made according to the claims and description of the present invention should still fall within the scope of protection of the present invention.

Claims

1. A multi-sensor based graph attention temporal fusion network fault diagnosis method, characterized by The following steps are involved: The vibration data of the gearbox is collected by multiple sensors, and the vibration data of the gearbox collected by multiple sensors is input into the trained graph attention time series fusion network to obtain the output fault classification result; The graph attention temporal fusion network specifically performs the following steps: Step 1: Use multiple sensors to collect the vibration data of the gearbox, and use the sensors as nodes and the vibration data as node attributes to input the graph attention network to obtain the output feature h of the graph attention network at the current moment i ', i = 1, 2, ... t; Step 2: Output features h of the graph attention network at all times i 'Input the SE module and LSTM respectively to obtain the adaptively weighted feature h output by the SE module at each moment' all And the value h" of LSTM output at each moment i ; Step 3: h' at each moment all and h” i Add the values, then input the sum of all the values ​​at all times into the SE module to obtain the adaptively weighted feature h' at each moment a ' ll , and finally h' a ' ll After entering the fully connected layer, the classification result is obtained through softmax.

2. According to claim 1, a multi-sensor based graph attention temporal fusion network fault diagnosis method is characterized in that The specific steps of step one are: Step 1: Take each sensor as a node to construct the graph G t , the graph G t It is expressed as: G t =(V,E,X) t Wherein, V represents the set of node numbers, E represents the set of boundary numbers, X represents the time point data of the node, and t represents the time sum of the vibration data. Step 1 and 2: Use multiple sensors to collect vibration data of the gearbox, and normalize the vibration data by maximum and minimum values; Step 13: Apply the ε radius criterion to determine the neighbor nodes of each node through the cosine similarity between two nodes, and connect the node with the neighbor node; Step 14: Calculate the similarity coefficient between the node and each of its corresponding neighbor nodes respectively, and based on the similarity coefficient, obtain the attention coefficient between the nodes. Then, use the time point data of the node as a feature, perform linear weighting on the attention coefficient, and obtain the output feature h from the current moment to the graph attention network. i '.

3. According to claim 2, a multi-sensor based graph attention time series fusion network fault diagnosis method is characterized in that The similarity coefficient is expressed as: have been ij =a([Wn i ||Wn j ]),j∈N i Among them, || represents the feature connection operation, a represents the mapping, W represents the weight matrix, and n i Represents node features, n j represents the neighbor node characteristics, N i represents the total number of neighbor nodes, j represents the sequence number of the neighbor node, e ij Represents the similarity coefficient between the central node and another node.

4. According to claim 3, a multi-sensor based graph attention time series fusion network fault diagnosis method is characterized in that The attention coefficient is expressed as: Among them, k represents each node, e ik Represents the similarity coefficient between the central node and each other node.

5. According to claim 4, a multi-sensor based graph attention temporal fusion network fault diagnosis method is characterized in that The output feature h of the graph attention network i ' is expressed as: Among them, σ represents the sigmoid activation function.

6. According to claim 5, a multi-sensor based graph attention time series fusion network fault diagnosis method is characterized in that The fusion feature h a ' ll It is expressed as: s n =σ(W2δ(W1z n )) Among them, R represents the number of elements in each graph attention output matrix, δ represents the relu function, W1 and W2 represent the coefficients of the fully connected layer, Represents an operation symbol, which means that the data from each graph attention network is multiplied by the weight obtained by adaptive weighting. n represents the weight matrix, z n Represents the aggregation of graph attention outputs at different moments.

7. A multi-sensor based graph attention temporal fusion network fault diagnosis method according to claim 6, characterized in that The fusion feature h' a ' ll It is expressed as: s n =σ(W2δ(W1z n )) 8. According to claim 7, a multi-sensor based graph attention time series fusion network fault diagnosis method is characterized in that The LSTM outputs the value h" i It is expressed as: h” i =o t *tanh(C t ) the t =σ(W o [h t-1 ,x t ]+b o ) i t =σ(W i ·[h t-1 ,x t ]+b t ) f t =σ(W f ·[h t-1 ,x t ]+b f ) Among them, W f is the weight of the forget gate, h t-1 is the input of the previous neuron, x t is the data input, b f is the bias value of the forget gate, f t is the output of the forget gate, W i is the weight of the input gate, h t-1 is the input of the previous neuron, x t is the data input, b t is the bias value of the input gate, i t is the output of the input gate, W C is the weight of the candidate value, h t-1 is the input of the previous neuron, x t is the data input, b C is the deviation value of the candidate value, is the output of the candidate value, C t-1 is the output of the previous candidate value, C t is the updated value of the candidate value, W o is the weight of the output gate, h t-1 is the input of the previous neuron, x t is the data input, b o is the deviation value of the output gate, o t is the output of the output gate, h t is the value passed to the next neuron.

Images