Intelligent fault diagnosis and life prediction method based on multi-task graph neural network

Abstract

The application discloses a method for intelligent fault diagnosis and life prediction based on a multi-task graph neural network. The method is as follows: collecting the full life cycle vibration signals of a rotating machine, setting a fault category label and a remaining useful life label; using a sliding window to extract multiple time-frequency domain features from the vibration signals, calculating the correlation coefficients between the features, and combining the label information to construct a sample graph; inputting the sample graph into a graph convolution module, automatically extracting deep features through a graph isomorphism layer and a global average pooling layer; inputting the features extracted by the graph convolution module into a self-attention module, and weighting the features by using a self-attention mechanism; inputting the weighted features into two fully connected layers in parallel, which are used to perform a fault diagnosis task and a life prediction task respectively, constructing a joint loss function for training, and obtaining a multi-task model; inputting a test sample graph into the trained multi-task model, and the model can simultaneously output the fault category and the remaining useful life of the sample graph.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent operation and maintenance of rotating machinery, and specifically relates to an intelligent fault diagnosis and life prediction method based on a multi-task graph neural network. Background Art

[0002] In recent years, a new round of global scientific and technological revolution and industrial transformation has deepened. Information technology has continuously made breakthroughs and is rapidly integrating with manufacturing technologies, accelerating the manufacturing industry's transition to high-end, intelligent, and digital manufacturing. Furthermore, with the rapid development of artificial intelligence (AI) technology and its accelerated integration with manufacturing, intelligent manufacturing has achieved remarkable results. This has not only improved manufacturing efficiency and economic benefits, but also promoted the digital transformation of manufacturing, comprehensively enhancing the development level of my country's manufacturing industry and product quality. However, as mechanical equipment continues to develop towards larger-scale, continuous, and automated operations, its construction has become increasingly sophisticated and complex, presenting significant challenges for its maintenance. At the same time, growing production demands have placed higher demands on the stable operation of mechanical equipment.

[0003] Rotating machinery (such as bearings, gears, and spindles) are critical components in mechanical equipment. Due to long-term operation in harsh environments with high loads and variable operating conditions, they inevitably experience various faults such as wear and tear, leading to unstable operation or even failure. If these faults are not discovered and repaired in a timely manner, they can cause significant safety hazards and economic losses. Therefore, diagnosing and predicting the health status of key components in rotating machinery and other equipment is of great research significance.

[0004] With the rapid development of computer science and big data technology, deep learning-based intelligent fault diagnosis and life prediction methods for rotating machinery have gained favor among many scholars, becoming a research hotspot in the field and producing a large number of scientific research results. Such methods have powerful feature extraction and nonlinear fitting capabilities, and can adaptively and quickly mine the knowledge contained in massive amounts of data, achieving "end-to-end" fault diagnosis and life prediction. In existing technologies, deep learning-based methods use long-short-term memory networks to construct a fault diagnosis model and a life prediction model, respectively. This effectively monitors the state of rotating machinery and reduces the operation and maintenance costs of mechanical equipment (Chen Xiang, Liu Qinming, Hu Jiarui. Bearing fault diagnosis and life prediction based on attention mechanism and long-short-term memory network under multi-source sensor data [J / OL]. Information and Control: 1-15 [2023-11-03].).

[0005] Although the above method can well solve the condition monitoring problem of rotating machinery, a fault diagnosis model and a life prediction model need to be established respectively, resulting in an increase in model development cost, reducing the practicability, and being not conducive to deployment in industrial practice. Therefore, how to train a single model to realize the fault diagnosis and life prediction of rotating machinery is undoubtedly a research topic full of engineering practical value. SUMMARY

[0006] In order to solve the problems existing in the prior art, the present application proposes an intelligent fault diagnosis and life prediction method based on multi-task graph neural network. The method is composed of a sample graph construction module, a graph convolution module, a self-attention module and a multi-task module, which simultaneously performs the fault diagnosis and life prediction of rotating machinery through multi-task joint training, can effectively reduce the development and deployment cost of the model, and provides more comprehensive information for the maintenance of mechanical equipment.

[0007] The object of the present application is achieved at least by one of the following technical solutions.

[0008] The intelligent fault diagnosis and life prediction method based on multi-task graph neural network comprises the following steps:

[0009] Step 1, data acquisition and labeling: install an acceleration sensor to collect the full life cycle vibration signal of rotating machinery during the running degradation process, set the segmented linear rotating machinery remaining useful life label for the full life cycle vibration signal according to the fault initiation point, and set the fault category label at the same time;

[0010] Step 2, sample graph construction: in the sample graph construction module, the collected full life cycle vibration signal is sampled and the multiple time-frequency domain features of the sample are extracted by using a sliding window to obtain a feature vector, the Pearson correlation coefficient between each feature vector is calculated, and connection edges are added between each feature vector according to the set threshold T to obtain a sample graph of each sample. Finally, combined with the label information provided in step 1, a sample graph dataset of rotating machinery can be constructed Wherein, x i represents the i-th sample graph in the sample graph dataset, and respectively represent the fault category label and the remaining life label corresponding to the i-th sample graph, and N is the total number of samples;

[0011] Step 3. Feature extraction: In the graph convolution module, three graph isomorphism layers and three global average pooling layers are used to extract features from the input sample graph. Each graph isomorphism layer consists of a graph isomorphism convolution layer, a batch normalization layer, an activation layer, and a graph pooling layer. The output of the previous graph pooling layer is the input of the next graph isomorphism convolution layer. When the i-th sample graph x is input, i When it comes to the graph convolution module, the outputs of the three-layer graph pooling layer are and Afterwards, and The global average pooling layer is used for further pooling, and the graph convolution module is obtained for the i-th sample graph x i The extracted features are represented as and

[0012] Step 4. Self-attention mechanism construction: In the self-attention module, the features extracted by the graph convolution module in step 3 on the sample graph are and Perform linear mapping to obtain the query matrix Q, key matrix K and value matrix V respectively, and then calculate the attention score based on the similarity principle to obtain the attention score matrix A; finally, use the attention score matrix A to weight the value matrix V to obtain the output P after the input features are weighted by the self-attention mechanism. in and Represents the input features respectively and The corresponding output features;

[0013] Step 5, multi-task joint training: In the multi-task module, two fully connected layers are constructed in parallel to realize the joint training of fault diagnosis task and life prediction task; the output features obtained in step 4 are respectively used. and As input to the fault diagnosis task and life prediction task, the two tasks are trained using the gradient descent algorithm. The joint loss is calculated through forward propagation, and the parameters of each part of the multi-task model are updated through backpropagation based on the joint loss. The multi-task model includes a sample graph construction module, a graph convolution module, a self-attention module, and a multi-task module to achieve joint training.

[0014] Step 6. Model testing: Input the test sample image into the multi-task model trained in step 5. The multi-task model simultaneously outputs a natural number and a floating-point number between [0, 1], which respectively represent the rotating machinery fault category and remaining service life of the sample image.

[0015] Furthermore, in step 1, the fault category label of the rotating machinery is represented by a natural number, and the remaining service life label is represented by a floating point number between [0, 1], where the critical values ​​1 and 0 represent the fault starting point and complete failure point of the rotating machinery, respectively.

[0016] Furthermore, in step 2, when the sliding window is used to sample the full life cycle vibration signal and extract sample features, commonly used time-frequency domain features including mean, root mean square value, variance, etc. are specifically extracted.

[0017] Furthermore, in step 2, the Pearson correlation coefficient used to calculate the similarity between each feature vector can be expressed as:

[0018]

[0019] Among them, r represents the Pearson correlation coefficient, which ranges from [0,1]. The closer it is to 1, the stronger the correlation of the eigenvectors. il and Represents the eigenvector u i The lth component of and the mean of each component in the eigenvector, u jl and Similarly, n is the dimension of the feature vector.

[0020] Furthermore, in step 3, when the three-layer graph isomorphism layer and the global average pooling layer are used to extract features from the input sample graph, the output of each layer can be expressed as:

[0021]

[0022] Among them, GINConv(·) is the graph isomorphism convolution layer, BN(·) is the batch normalization layer, σ(·) is the ReLU activation function layer, EdgePool(·) is the graph pooling layer, and the numerical subscripts of the above layers indicate the graph isomorphism layer to which they belong. In addition, x i represents the i-th sample graph of the input graph convolution module, and Represents the sample graph x i The outputs obtained after the first, second and third graph isomorphism layers, Readout1(·), Readout2(·) and Readout3(·) are global average pooling layers, which are used to and Pooling is performed, and finally the graph convolution module is obtained for the i-th sample graph x i The extracted features and

[0023] Furthermore, in step 4, the input features and When performing linear mapping, the query matrix, key matrix, and value matrix obtained are expressed as follows:

[0024]

[0025] in, and is the input feature of the self-attention module, W q 、W k and W v are three learnable parameter matrices, which can be multiplied by the input features to obtain the query matrix Q, key matrix K and value matrix V respectively.

[0026] Furthermore, in step 4, the attention score matrix is ​​determined by the similarity between the query matrix Q and the key matrix K, and its calculation expression is:

[0027]

[0028] Among them, Q, K and A represent the query matrix, key matrix and attention score matrix respectively, Softmax(·) is the normalization function, d k The dimension of the input feature in the self-attention module is used for scale transformation to prevent the input Softmax(·) value from being too large, causing the partial derivative calculated by the gradient descent algorithm to approach 0.

[0029] Furthermore, in step 4, input features and The output after weighting by the self-attention module is:

[0030]

[0031] Among them, A and V are the attention score matrix and value matrix respectively. and Represents the input features and The output features obtained after weighting by the self-attention module, P is the output feature matrix.

[0032] Furthermore, in step 5, the fully connected layer used to perform the fault diagnosis task is trained using the Softmax classifier and the cross entropy loss function, while the fully connected layer used to perform the life prediction task is trained using the mean square error loss function. The cross entropy loss function and the mean square error loss function are expressed as follows:

[0033]

[0034]

[0035] Among them, L c is the cross entropy loss function, L p is the mean square error loss function, x i 、 and represent the i-th sample image used for training and its corresponding fault category label and remaining service life label, respectively, f c (·) is the mapping function of the fault diagnosis task, f p (·) is the mapping function of the lifespan prediction task, N is the number of training sample graphs;

[0036] When performing multi-task joint training, the joint loss function is expressed as:

[0037] L u =λ c L c +λ p L p (8)

[0038] Among them, L c 、L p and L u They are the cross entropy loss function, the mean square error loss function and the joint loss function of the two, λ c and λ p Represents weights, which are used to weight the cross entropy loss and mean square error loss respectively to control the importance of the multi-task model to the fault diagnosis task and life prediction task.

[0039] Furthermore, in step 6, when testing the trained multi-task model, the accuracy and F1 score can be used as indicators to evaluate the fault diagnosis performance of the multi-task model, and the RMSE and R 2 The value is used as an indicator to evaluate the life prediction performance of the multi-task model.

[0040] Compared with the prior art, the present invention has at least the following beneficial effects:

[0041] 1. This invention constructs a multi-task model based on the parameter sharing mechanism and the joint loss function, realizing the function of one model completing two tasks. It effectively solves the problem of excessive development and deployment costs caused by training two models in traditional methods, and provides a more efficient solution for the health management and maintenance of mechanical equipment, which has certain engineering application value.

[0042] 2. The present invention proposes a sample graph construction strategy based on feature similarity, which fully utilizes the time-frequency domain feature information contained in the rotating machinery vibration signal, and provides a more reliable mapping scheme for the rotating machinery fault diagnosis and life prediction method based on graph neural network. It also verifies the feasibility of applying graph structure data to the intelligent operation and maintenance of mechanical equipment.

[0043] 3. The present invention adaptively weights the output features of the graph convolution module through the self-attention module, so that the fault diagnosis task and the life prediction task can obtain different feature representations while sharing model parameters, thereby effectively improving the performance of the model on each task and making the model have more superior performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of an intelligent fault diagnosis and life prediction method based on a multi-task graph neural network in an embodiment of the present invention;

[0045] Figure 2 This is an overall framework diagram of an intelligent fault diagnosis and life prediction method based on a multi-task graph neural network in an embodiment of the present invention;

[0046] Figure 3 is a histogram of the F1 scores obtained by each comparison method in the embodiment of the present invention;

[0047] Figure 4 is a histogram of the RMSE obtained by each comparison method in the embodiment of the present invention;

[0048] Figure 5 is the R obtained by each comparison method in the embodiment of the present invention 2 Value histogram. DETAILED DESCRIPTION

[0049] In order to more clearly illustrate the purpose and technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to better illustrate the application of the present invention, but the technical features involved in the embodiments of the present invention are not limited thereto.

[0050] Example

[0051] like Figure 1 As shown in the figure, a method for intelligent fault diagnosis and life prediction based on a multi-task graph neural network is presented. The method can simultaneously perform the two tasks of fault diagnosis and life prediction of rotating machinery through multi-task joint training, thereby effectively reducing the development and deployment costs of the model. The method includes the following steps:

[0052] Step 1: Data collection and labeling: Install an acceleration sensor to collect the full life cycle vibration signal of the rotating machinery during the operational degradation process. Set a piecewise linear remaining service life label for the rotating machinery based on the fault origin point for the full life cycle vibration signal, and also set a fault category label.

[0053] Step 2, sample graph construction: In the sample graph construction module, the collected full life cycle vibration signal is sampled using a sliding window and multiple time-frequency domain features of the sample are extracted to obtain the feature vector. By calculating the Pearson correlation coefficient between each feature vector and adding connecting edges between each feature vector according to the set threshold T, the sample graph of each sample is obtained. Finally, combined with the label information provided in step 1, the sample graph dataset of the rotating machinery can be constructed. Among them, x i represents the i-th sample graph in the sample graph dataset, and They represent the fault category label and remaining life label corresponding to the i-th sample graph, respectively, and N is the total number of samples;

[0054] Step 3. Feature extraction: In the graph convolution module, three graph isomorphism layers and three global average pooling layers are used to extract features from the input sample graph. Each graph isomorphism layer consists of a graph isomorphism convolution layer, a batch normalization layer, an activation layer, and a graph pooling layer. The output of the previous graph pooling layer is the input of the next graph isomorphism convolution layer. When the i-th sample graph x is input, i When it comes to the graph convolution module, the outputs of the three-layer graph pooling layer are and Afterwards, and The global average pooling layer is used for further pooling, and the graph convolution module is obtained for the i-th sample graph x i The extracted features are represented as and

[0055] Step 4. Self-attention mechanism construction: In the self-attention module, the features extracted by the graph convolution module in step 3 on the sample graph are and Perform linear mapping to obtain the query matrix Q, key matrix K and value matrix V respectively, and then calculate the attention score based on the similarity principle to obtain the attention score matrix A; finally, use the attention score matrix A to weight the value matrix V to obtain the output P after the input features are weighted by the self-attention mechanism. in and Represents the input features and The corresponding output features;

[0056] Step 5, multi-task joint training: In the multi-task module, two fully connected layers are constructed in parallel to realize the joint training of fault diagnosis task and life prediction task. The output features obtained in step 4 are respectively used. and As the input of the fault diagnosis task and the life prediction task, the two tasks are trained using the gradient descent algorithm. The joint loss is calculated through forward propagation, and the parameters of each part of the multi-task model are updated through backpropagation based on the joint loss to achieve joint training.

[0057] Step 6. Model testing: Input the test sample image into the multi-task model trained in step 5. The multi-task model simultaneously outputs a natural number and a floating-point number between [0, 1], which respectively represent the rotating machinery fault category and remaining service life of the sample image.

[0058] As an embodiment, in step 1, the fault category label of the rotating machinery is represented by a natural number, and the remaining service life label is represented by a floating point number between [0, 1], where the critical values ​​1 and 0 represent the fault starting point and complete failure point of the rotating machinery, respectively.

[0059] Furthermore, in step 2, when the sliding window is used to sample the full life cycle vibration signal and extract sample features, commonly used time-frequency domain features including mean, root mean square value, variance, etc. are specifically extracted.

[0060] Furthermore, in step 2, the Pearson correlation coefficient used to calculate the similarity between each feature vector can be expressed as:

[0061]

[0062] In the formula, r represents the Pearson correlation coefficient, which ranges from [0,1]. The closer it is to 1, the stronger the correlation of the eigenvectors. il and Represents the eigenvector u i The lth component of and the mean of each component in the eigenvector, u jl and Similarly, n is the dimension of the feature vector. In addition, the threshold T is set to 0.8. When the Pearson correlation coefficient r between any two features is greater than the threshold T, a connection edge in the sample graph will be generated for them.

[0063] Furthermore, when the three-layer graph isomorphism layer and the global average pooling layer are used to extract features from the input sample graph in step 3, the output of each layer can be expressed as:

[0064]

[0065] In the formula, GINConv(·) is the graph homogeneity convolution layer, BN(·) is the batch normalization layer, σ(·) is the ReLU activation function layer, EdgePool(·) is the graph pooling layer, and the numerical subscripts of the above layers indicate the graph homogeneity layer to which they belong. In addition, x i represents the i-th sample graph of the input graph convolution module, and Represents the sample graph x i The outputs obtained after the first, second and third graph isomorphism layers, Readout1(·), Readout2(·) and Readout3(·) are global average pooling layers, which are used to and Pooling is performed, and finally the graph convolution module is obtained for the i-th sample graph x i The extracted features and

[0066] Furthermore, in step 4, the input features and When performing linear mapping, the query matrix, key matrix, and value matrix obtained are expressed as follows:

[0067]

[0068] Where, and is the input feature of the self-attention module, W q 、W k and W v are three learnable parameter matrices, which can be multiplied by the input features to obtain the query matrix Q, key matrix K and value matrix V respectively.

[0069] Furthermore, in step 4, the attention score matrix is ​​determined by the similarity between the query matrix Q and the key matrix K, and its calculation expression is:

[0070]

[0071] Where Q, K and A represent the query matrix, key matrix and attention score matrix respectively, Softmax(·) is the normalization function, and d k The dimension of the input feature in the self-attention module is used for scale transformation to prevent the input Softmax(·) value from being too large, causing the partial derivative calculated by the gradient descent algorithm to approach 0.

[0072] Furthermore, in step 4, input features and The output after weighting by the self-attention module is:

[0073]

[0074] Where A and V are the attention score matrix and value matrix respectively. and Represents the input features respectively and The output features obtained after weighting by the self-attention module, P is the output feature matrix.

[0075] Furthermore, in step 5, the fully connected layer used to perform the fault diagnosis task is trained using the Softmax classifier and the cross entropy loss function, while the fully connected layer used to perform the life prediction task is trained using the mean square error loss function. The cross entropy loss function and the mean square error loss function are expressed as follows:

[0076]

[0077]

[0078] Among them, L c is the cross entropy loss function, L p is the mean square error loss function, x i 、 and represent the i-th sample image used for training and its corresponding fault category label and remaining service life label, respectively, f c (·) is the mapping function of the fault diagnosis task, f p (·) is the mapping function for the lifespan prediction task, and N is the number of training sample graphs.

[0079] Furthermore, in step 5, when performing multi-task joint training, the joint loss function is expressed as:

[0080] L u =λ c L c +λ p L p (8)

[0081] Where, L c , L p and L u They are the cross entropy loss function, the mean square error loss function and the joint loss function of the two, λ c and λ p Represents weights, which are used to weight the cross entropy loss and mean square error loss respectively to control the importance of the multi-task model to the fault diagnosis task and life prediction task.

[0082] Furthermore, in step 6, when testing the trained multi-task model, the accuracy and F1 score can be used as indicators to evaluate the fault diagnosis performance of the multi-task model, and the RMSE and R 2 The value is used as an indicator to evaluate the life prediction performance of the multi-task model.

[0083] The present invention is further described below with reference to the accompanying drawings and experimental cases:

[0084] In one embodiment, the experiments were designed using the publicly available Xi'an Jiaotong University rolling bearing dataset (XJTU-SY dataset). This dataset includes rolling bearing lifecycle vibration signal data collected under three operating conditions, with a sampling frequency of 25.6 kHz. Detailed information for the three operating conditions is shown in Table 1.

[0085] Table 1 Experimental conditions information of XJTU-SY dataset

[0086]

[0087] Table 2 Experimental bearing description and data set division table

[0088]

[0089] In one embodiment, nine bearing data sets under three operating conditions from the XJTU-SY dataset were selected to design a fault diagnosis and life prediction experiment. A detailed description of the bearings used and the dataset partitioning are shown in Table 2. The training set includes data from two bearings under each operating condition, while the test set includes data from one bearing under each operating condition. The bearing fault categories are normal, inner race fault, and outer race fault, with labels 0, 1, and 2, respectively. A sliding window of length 2048 was used to capture the vibration signal to obtain the raw samples. The 16 time-frequency domain features shown in Table 3 were extracted from the raw samples, and the Pearson correlation coefficients between the features were calculated to construct a sample graph.

[0090] Table 3 Extracted time-frequency domain features

[0091]

[0092] Table 4 Model structure of the method proposed in this embodiment

[0093]

[0094]

[0095] In one embodiment, in order to verify the effectiveness and advancement of the proposed method, a first layer wide convolutional deep neural network (WDCNN) and a long short-term memory network (LSTM) are adopted as comparison, wherein the WDCNN is composed of 5 convolutional layers and 1 fully connected layer, and the LSTM is composed of four fully connected layers with memory units stacked. For the proposed method, it is composed of three graph isomorphic layers and two parallel double-layer fully connected layers, and the detailed structure of the model is shown in Table 4. In addition, the hyperparameter settings of the proposed method are shown in Table 5.

[0096] Table 5 Hyperparameter settings table of the proposed method in the embodiment

[0097]

[0098] In one embodiment, due to the extremely unbalanced number of samples in each category, the accuracy used to evaluate the fault diagnosis performance of the model will have a large deviation. In order to ensure the reliability of the experimental results, the F1 score is used as the evaluation index of the fault diagnosis performance of the model in this embodiment, and the RMSE and R 2 value are used as the evaluation indexes of the life prediction performance of the model. In order to reduce randomness, a total of 5 repeated experiments are carried out, and the average value of each evaluation index is taken as the final experimental result.

[0099] Figure 3 、 Figure 4 and Figure 5 respectively show the F1 score, RMSE and R 2 value obtained by various comparison methods, and Table 6 summarizes these results. According to the experimental results, the average F1 score obtained by the LSTM is 0.657, which is significantly lower than the 0.747 of the WDCNN. However, for the RMSE and R 2 value, the average values obtained by the LSTM are 0.122 and 0.859 respectively, which are significantly better than the 0.14 and 0.803 of the WDCNN. This shows that the WDCNN has a deep feature extraction capability, which can effectively mine the fault discrimination information contained in the input samples, so it performs well in the fault diagnosis task; on the other hand, the LSTM can better capture time series information and long-term dependencies due to its memory unit characteristics, so it has an advantage in life prediction task. As for the proposed method, its F1 score, RMSE and R 2 value reached 0.803, 0.086 and 0.927 respectively, which are better than the WDCNN and LSTM, indicating that the method can have the ability of deep feature extraction and time series information capture, so it can achieve excellent performance in both fault diagnosis task and life prediction task, fully verifying the effectiveness and advancement of its multi-task modeling.

[0100] Table 6 Experimental results of each pair of comparison methods in the examples

[0101]

[0102] From the above analysis, it can be seen that the method based on the multi-task graph neural network can simultaneously perform effective fault diagnosis and life prediction for rotating machinery. Compared with the single-task method, the proposed method can significantly reduce the development and deployment cost of the model, and performs well in each task. In the health management and predictive maintenance process of mechanical equipment, the proposed method can effectively improve the operation and maintenance efficiency of mechanical equipment by providing more comprehensive fault category and degradation level information, and therefore has high engineering practical value.

[0103] It should be noted that although the embodiments of the present application have been described in detail with reference to examples, those skilled in the art should understand that any modifications, replacements and improvements made without departing from the design spirit of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent fault diagnosis and life prediction method based on a multi-task graph neural network, characterized by: The following steps are involved: Step 1: Data collection and labeling: Install an acceleration sensor to collect the full life cycle vibration signal of the rotating machinery during the operational degradation process. Set a piecewise linear remaining service life label for the rotating machinery based on the fault origin point for the full life cycle vibration signal, and also set a fault category label. Step 2, sample graph construction: In the sample graph construction module, the collected full life cycle vibration signal is sampled using a sliding window and multiple time-frequency domain features of the sample are extracted to obtain the feature vector. By calculating the Pearson correlation coefficient between each feature vector and adding connecting edges between each feature vector according to the set threshold T, the sample graph of each sample is obtained. Finally, combined with the label information provided in step 1, the sample graph dataset of the rotating machinery can be constructed. Among them, x i represents the i-th sample graph in the sample graph dataset, and They represent the fault category label and remaining life label corresponding to the i-th sample graph, respectively, and N is the total number of samples; Step 3. Feature extraction: In the graph convolution module, three graph isomorphism layers and three global average pooling layers are used to extract features from the input sample graph. Each graph isomorphism layer consists of a graph isomorphism convolution layer, a batch normalization layer, an activation layer, and a graph pooling layer. The output of the previous graph pooling layer is the input of the next graph isomorphism convolution layer. When the i-th sample graph x is input, i When it comes to the graph convolution module, the outputs of the three-layer graph pooling layer are and Afterwards, and The global average pooling layer is used for further pooling, and the graph convolution module is obtained for the i-th sample graph x i The extracted features are represented as and Step 4. Self-attention mechanism construction: In the self-attention module, the features extracted by the graph convolution module in step 3 on the sample graph are and Perform linear mapping to obtain the query matrix Q, key matrix K and value matrix V respectively, and then calculate the attention score based on the similarity principle to obtain the attention score matrix A; finally, use the attention score matrix A to weight the value matrix V to obtain the output P after the input features are weighted by the self-attention mechanism. in and Represents the input features respectively and The corresponding output features; Step 5, multi-task joint training: In the multi-task module, two fully connected layers are constructed in parallel to realize the joint training of fault diagnosis task and life prediction task; the output features obtained in step 4 are respectively used. and As input to the fault diagnosis task and life prediction task, the two tasks are trained using the gradient descent algorithm. The joint loss is calculated through forward propagation, and the parameters of each part of the multi-task model are updated through backpropagation based on the joint loss. The multi-task model includes a sample graph construction module, a graph convolution module, a self-attention module, and a multi-task module to achieve joint training. Step 6. Model testing: Input the test sample image into the multi-task model trained in step 5. The multi-task model simultaneously outputs a natural number and a floating-point number between [0, 1], which respectively represent the rotating machinery fault category and remaining service life of the sample image.

2. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 1, the fault category label of the rotating machinery is represented by a natural number, and the remaining service life label is represented by a floating point number between [0, 1], where the critical values ​​1 and 0 represent the fault initiation point and complete failure point of the rotating machinery, respectively.

3. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 2, when the sliding window is used to sample the full life cycle vibration signal and extract sample features, the time-frequency domain features including the mean, root mean square value, and variance are specifically extracted.

4. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 2, the Pearson correlation coefficient used to calculate the similarity between each feature vector can be expressed as: Among them, r represents the Pearson correlation coefficient, which ranges from [0,1]. The closer it is to 1, the stronger the correlation of the eigenvectors. il and Represents the eigenvector u i The lth component of and the mean of each component in the eigenvector, u jl and Similarly, n is the dimension of the feature vector.

5. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 3, when the three-layer graph isomorphism layer and the global average pooling layer are used to extract features from the input sample graph, the output of each layer can be expressed as: Among them, GINConv(·) is the graph isomorphism convolution layer, BN(·) is the batch normalization layer, σ(·) is the ReLU activation function layer, EdgePool(·) is the graph pooling layer, and the numerical subscripts of the above layers indicate the graph isomorphism layer to which they belong. In addition, x i represents the i-th sample graph of the input graph convolution module, and Represents the sample graph x i The outputs obtained after the first, second and third graph isomorphism layers, Readout1(·), Readout2(·) and Readout3(·) are global average pooling layers, which are used to and Pooling is performed, and finally the graph convolution module is obtained for the i-th sample graph x i The extracted features and 6. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 4, the input features and When performing linear mapping, the query matrix, key matrix, and value matrix obtained are expressed as follows: in, and is the input feature of the self-attention module, W q 、W k and W v are three learnable parameter matrices, which can be multiplied by the input features to obtain the query matrix Q, key matrix K and value matrix V respectively.

7. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 4, the attention score matrix is ​​determined by the similarity between the query matrix Q and the key matrix K, and its calculation expression is: Among them, Q, K and A represent the query matrix, key matrix and attention score matrix respectively, Softmax(·) is the normalization function, d k The dimension of the input feature in the self-attention module is used for scale transformation to prevent the input Softmax(·) value from being too large, causing the partial derivative calculated by the gradient descent algorithm to approach 0.

8. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 4, input features and The output after weighting by the self-attention module is: Among them, A and V are the attention score matrix and value matrix respectively. and Represents the input features respectively and The output features obtained after weighting by the self-attention module, P is the output feature matrix.

9. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 5, the fully connected layer used to perform the fault diagnosis task is trained using the Softmax classifier and the cross entropy loss function, while the fully connected layer used to perform the life prediction task is trained using the mean square error loss function. The cross entropy loss function and the mean square error loss function are expressed as follows: Among them, L c is the cross entropy loss function, L p is the mean square error loss function, x i 、 and represent the i-th sample image used for training and its corresponding fault category label and remaining service life label, respectively, f c (·) is the mapping function of the fault diagnosis task, f p (·) is the mapping function of the lifespan prediction task, N is the number of training sample graphs; When performing multi-task joint training, the joint loss function is expressed as: L u =λ c L c +λ p L p (8) Among them, L c 、L p and L u They are the cross entropy loss function, the mean square error loss function and the joint loss function of the two, λ c and λ p Represents weights, which are used to weight the cross entropy loss and mean square error loss respectively to control the importance of the multi-task model to the fault diagnosis task and life prediction task.

10. The intelligent fault diagnosis and life prediction method based on multi-task graph neural network according to claim 1 is characterized in that: In step 6, when testing the trained multi-task model, the accuracy and F1 score can be used as indicators to evaluate the fault diagnosis performance of the multi-task model, and the RMSE and R 2 The value is used as an indicator to evaluate the life prediction performance of the multi-task model.

Images