Rolling bearing remaining useful life prediction method based on double graph attention and gru

Abstract

The application discloses a rolling bearing residual service life prediction method based on double-graph attention and GRU. The method comprises the following steps: extracting time domain features and frequency domain features of preprocessed rolling bearing vibration signal data and constructing a feature full connection graph, constructing and training a rolling bearing residual service life prediction model based on double-graph attention and a gated recurrent unit, mining spatial correlation information between graph node features at different times and time correlation information between graph structures, inputting test set data into the trained rolling bearing residual service life prediction model to predict the residual service life of the rolling bearing to be measured. The application can adaptively mine the rolling bearing degradation information contained in the graph structure, effectively improve the prediction accuracy of the residual service life of the rolling bearing, and is beneficial to guarantee the safe and reliable operation of mechanical equipment and improve the operation cycle of the equipment.

Description

Technical Field

[0001] This invention relates to the field of equipment prediction and health management, and specifically to a method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU. Background Technology

[0002] Rolling bearings are critical components of aero engines, frequently operating under harsh conditions and prone to failure, potentially leading to engine malfunction. Remaining Useful Life (RUL) is defined as the expected continuous operating time of a device from the present until a potential failure occurs. RUL prediction technology can reveal the failure time of rolling bearings, providing information on their health status, thereby preventing significant losses and improving engine reliability. Therefore, accurately predicting the RUL of rolling bearings is a crucial research area.

[0003] Rolling bearing RUL prediction methods can be broadly categorized into physical model-based methods and data-driven methods. For multi-component composite structures like rolling bearings, it's difficult to accurately establish physical models that fit actual operating conditions; therefore, data-driven prediction methods can be used. Data-driven methods use collected time-series data to establish health indicators and predictive models characterizing the degree of equipment degradation, thus predicting future time-series values. In recent years, deep learning models, due to their powerful automatic feature extraction capabilities, have been widely adopted in rolling bearing RUL prediction. However, time-series data inherently possesses spatiotemporal characteristics; the distance between high-dimensional data is non-Euclidean distance. However, prediction methods based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs) do not explicitly model the spatial relationships between time series in non-Euclidean space, which disrupts the topological structure of the data and limits their expressive power.

[0004] Currently, graph neural networks (GNNs) have become a powerful tool for learning non-Euclidean data representations. By treating time points or variables as nodes and their relationships as edges, and modeling them with a graph structure, the network can capture diverse and complex relationships, including both spatial connections between variables and temporal dependencies. The spatial structure of a graph contains rich feature information and can represent the operating state of a rolling bearing over a period of time. However, the degradation of a rolling bearing is a dynamic process, and the spatial correlation between features of each node differs at different times of degradation. Although existing graph network-based methods are effective in predicting the RUL of rolling bearings, these prediction methods use a constant pre-constructed graph and cannot identify changes in the spatial correlation of input features at different times, resulting in low prediction accuracy. Summary of the Invention

[0005] To address the problems in the background art, this invention provides a method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU. This method effectively solves the problem that existing rolling bearing remaining service life prediction methods based on graph neural networks use a constant pre-constructed graph, which cannot identify changes in the correlation of the input feature space at different times, resulting in low accuracy in predicting the remaining service life of rolling bearings.

[0006] The technical solution adopted in this invention is:

[0007] 1) Obtain the periodic vibration signal data of the rolling bearing throughout its entire lifespan as a training set and perform preprocessing;

[0008] 2) Fully connected graph construction: For the preprocessed training set, construct fully connected graphs for time-domain features and frequency-domain features respectively;

[0009] 3) Construct a rolling bearing remaining service life prediction model based on dual-graph attention and gated recurrent unit network for fully connected graphs with time-domain and frequency-domain features;

[0010] 4) Train a model to predict the remaining service life of rolling bearings;

[0011] 5) Rolling bearing remaining service life prediction model prediction: Input the periodic vibration signal data of the rolling bearing under test throughout its entire life into the trained rolling bearing remaining service life prediction model to predict the remaining service life of the rolling bearing under test.

[0012] Step 2) specifically refers to:

[0013] 2.1) Feature extraction: Extract multiple time-domain features and multiple frequency-domain features from the periodic vibration signal data after preprocessing in step 1);

[0014] 2.2) Graph construction: Take each of the multiple time-domain features and multiple frequency-domain features obtained in step 2.1) as a node of the graph structure, and construct the graph structure to obtain the time-domain feature fully connected graph and the frequency-domain feature fully connected graph.

[0015] The rolling bearing remaining service life prediction model in step 3) includes a dual-graph attention network and a gated cyclic unit network. The output of the dual-graph attention network is input into the gated cyclic unit network. The time correlation information between graph structures is obtained by the gated cyclic unit network and used as the characteristic information of rolling bearing degradation trend.

[0016] The dual-graph attention network processes data in the following manner:

[0017] Extracting spatial correlation information between graph node features: The dual-graph attention network consists of two GAT layers, two fully connected layers, two ReLU layers, and a horzcat function. The temporal and frequency domain fully connected graphs constructed in step 2.2) are input into the dual-graph attention network. The two GAT layers process the temporal and frequency domain fully connected graphs respectively. After processing, the corresponding fully connected layers transform the temporal and frequency domain fully connected graphs into the same dimension before fusing them to generate new features. This yields deep-level fused features with extracted spatial correlation information. The calculation process is as follows:

[0018] H time =Relu(ω3·GAT1(F) time ,ω1)+b1)

[0019] H fre =Relu(ω4·GAT2(F) fre ,ω2)+b2)

[0020] H = horzcat(H) time H fre )

[0021] Among them, GAT1 and GAT2 are two GAT layers; F time F fre Here, ω1, ω2, ω3, and ω4 represent the time-domain and frequency-domain fully connected graphs of the input GAT layer, respectively; ω1, ω2, ω3, and ω4 are the first, second, third, and fourth weight matrices, respectively; b1 and b2 are the first and second bias matrices, respectively; ReLU is the activation function; H time H represents the time-domain features after activation. fre is the frequency domain feature after activation function; horzcat is the connection function; H is the deep fusion feature after extracting spatial correlation information;

[0022] The gated recurrent unit network is processed in the following way: extracting the temporal correlation information between graph structures: the gated recurrent unit network includes a reset gate and an update gate at each time node. The deep fusion features output by the dual graph attention network are input into the reset gate and update gate at each time node, and the output of the reset gate and update gate is calculated by the sigmoid function. As the time node is updated, multiple reset gate outputs and multiple update gate outputs are obtained. The temporal correlation information between graph structures is updated by using the multiple reset gate outputs and multiple update gate outputs, and the temporal correlation information between graph structures is used as the characteristic information of the degradation trend of rolling bearings.

[0023] The training of the rolling bearing remaining service life prediction model in step 4) is specifically as follows: the training set after preprocessing in step 1) is used to construct a time-domain feature fully connected graph and a frequency-domain feature fully connected graph in step 2), and then the rolling bearing remaining service life prediction model constructed in step 3) is used to obtain the rolling bearing degradation trend feature information. Finally, the linear layer of the fully connected neural network is used to fit the rolling bearing degradation trend feature information. The MSE function is selected as the loss function of the rolling bearing remaining service life prediction model. The Adam algorithm is used to optimize the hyperparameters in the dual-graph attention network, the gated recurrent unit network and the fully connected neural network, and finally the trained rolling bearing remaining service life prediction model is obtained.

[0024] The specific steps of the rolling bearing remaining service life prediction model test in step 5) are as follows: the periodic vibration signal data of the rolling bearing under test throughout its entire lifespan is used as a test set for normalization processing, and then input into the trained rolling bearing remaining service life prediction model to achieve the prediction of the remaining service life of the rolling bearing under test.

[0025] Step 1) specifically involves: normalizing the periodic vibration signal data of the rolling bearing throughout its entire lifespan using the Z-score function, as follows:

[0026]

[0027] Where x is the vibration signal data before normalization, x i The normalized vibration signal data is represented by μ, which is the mean of the vibration signal data, and σ is the standard deviation of the vibration signal data.

[0028] In step 2.1), multiple time-domain features include a variety of features such as maximum value, minimum value, peak-to-peak value, mean, variance, standard deviation, root square value, root square amplitude, kurtosis factor, skewness factor, waveform factor, and margin factor; multiple frequency-domain features include a variety of features such as mean frequency, centroid frequency, root mean square frequency, and standard deviation of frequency.

[0029] The extraction of multiple frequency domain features from the periodic vibration signal data in step 2.1) is specifically as follows:

[0030] After converting time-domain periodic vibration signal data into frequency-domain vibration signal data using the Fast Fourier Transform algorithm, multiple frequency-domain features of the frequency-domain vibration signal data are extracted.

[0031] This invention constructs a rolling bearing remaining service life prediction model. Using a constructed fully connected graph as input, a dual-graph attention network adaptively mines spatial correlation information between graph node features at different time points to obtain deep-level fused features. These fused features are then used as input to a gated recurrent unit network (GRU), which mines temporal correlation information between graph structures to construct rolling bearing degradation trend features. An optimization algorithm is then used to optimize the hyperparameters of all networks to obtain the rolling bearing remaining service life prediction model. This invention utilizes the constructed rolling bearing remaining service life prediction model based on dual-graph attention and gated recurrent unit networks to adaptively mine rolling bearing degradation information hidden in the graph structure, effectively improving the prediction accuracy of rolling bearing remaining service life.

[0032] The beneficial effects of this invention are:

[0033] This invention can effectively solve the problem that existing methods for predicting the remaining service life of rolling bearings based on graph neural networks use a constant pre-constructed graph, which cannot identify changes in the correlation of the input feature space at different times, resulting in low accuracy in predicting the remaining service life of rolling bearings.

[0034] This invention uses graph structures to map the degradation process of rolling bearings. By constructing a dual-graph attention network, it adaptively extracts spatial correlation information between graph node features at different times. By introducing a gated recurrent unit network into the field of rolling bearing remaining service life prediction, it further extracts temporal correlation information between graph structures. This invention achieves end-to-end rolling bearing remaining service life prediction based on deep learning, improves the prediction accuracy of rolling bearing remaining service life, and helps ensure the safe and reliable operation of mechanical equipment and improve the equipment's operating cycle. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the method of the present invention.

[0036] Figure 2 This is the time-domain fully connected graph of the method of the present invention.

[0037] Figure 3 This is the frequency domain feature fully connected graph of the method of the present invention.

[0038] Figure 4This is a structural diagram of the dual-graph attention network module of the method of the present invention.

[0039] Figure 5 This is a network structure diagram of the gated loop unit of the method of the present invention.

[0040] Figure 6 This is a schematic diagram illustrating the training and testing of the rolling bearing remaining service life prediction model according to the method of this invention.

[0041] Figure 7 This is the data acquisition test platform for the method of this invention.

[0042] Figure 8 This is a graph showing the predicted remaining service life of rolling bearings 1-3 in the test set using the method of the present invention.

[0043] Figure 9 This is a graph showing the predicted remaining service life of rolling bearings 2-4 in the test set using the method of the present invention.

[0044] Figure 10 This is a graph showing the predicted remaining service life of rolling bearing 3-3 in the test set of the method of this invention. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown, an embodiment of the rolling bearing remaining service prediction method based on dual-graph attention and GRU of the present invention is as follows:

[0047] 1) Data Preprocessing: Full-life-cycle vibration signal data of rolling bearings under different operating conditions were selected as the training set. In this example, the full-life-cycle vibration signal data used came from the PRONOSTIA test bench. The test bench structure is as follows: Figure 7As shown, this platform can accelerate the degradation process of rolling bearings, allowing degradation failures to be observed within hours. Data is acquired using an accelerometer mounted on the rolling bearing housing, sampling every 10 seconds at a frequency of 25600Hz, with each sampling session lasting 0.1 seconds, resulting in 2560 data points collected per session. A rolling bearing is considered to have failed when the acceleration amplitude of a data point exceeds 20g. The dataset includes three operating conditions: Condition 1 (1800 rpm, 4000 N load); Condition 2 (1650 rpm, 4200 N load); and Condition 3 (1500 rpm, 5000 N load). A total of 17 sets of rolling bearing life-cycle degradation data were obtained: seven sets under Condition 1, seven sets under Condition 2, and three sets under Condition 3. In this embodiment, rolling bearings 1-1, 1-2, 2-1, 2-2, 3-1, and 3-2 are selected as the training set data for model training, and rolling bearings 1-3, 2-4, and 3-3 are selected as the test set data.

[0048] Normalize the training set data;

[0049] Specifically, the full-life-cycle vibration signal data of rolling bearings under different working conditions obtained in step 1) are processed by Z-Score normalization to calculate the percentage of the true remaining service life (RUL) of the rolling bearing, and this percentage is used as the label value for model training. The normalization formula is as follows:

[0050]

[0051] Where x is the vibration signal data before normalization, x i The vibration signal data is normalized, μ is the mean of the vibration signal data, and σ is the standard deviation of the vibration signal data.

[0052] Assuming that the health of rolling bearings deteriorates linearly with operating time, the formula for calculating the percentage label of rolling bearings over their full lifespan is as follows:

[0053]

[0054] Where t0 is the start time of the rolling bearing operation, t is the failure time of the rolling bearing operation, and tt i Let y be the RUL value of the i-th sampling point. iLet y be the RUL percentage label for the i-th sampling point, and N be the total number of sampling points for the rolling bearing. For example, in the PRONOSTIA platform dataset used in this invention, there are 871 samples of rolling bearings under operating conditions, with a sampling interval of 10s, meaning the total lifespan of the rolling bearing is 8710s. Therefore, the current running time of the 300th sampling point (i.e., the 300th sample) is 3000s, corresponding to a RUL value of 5710s, and the RUL percentage label value is y. 300 It is 0.65557.

[0055] 2) Construction of fully connected graphs: Extract time-domain features from the vibration signal data preprocessed in step 1), including: mean, maximum, minimum, peak-to-peak value, variance, standard deviation, root mean square value, root mean square amplitude, kurtosis factor, skewness factor, waveform factor, and margin factor; then perform a fast Fourier transform on the vibration signal data and extract frequency-domain features, including: mean frequency, centroid frequency, root mean square frequency, and standard deviation of frequency; construct a time-domain feature fully connected graph and a frequency-domain feature fully connected graph using the extracted time-domain features and frequency-domain features respectively, and set the initial weight of the edges in the graph to 1;

[0056] The calculation methods for time-domain and frequency-domain features are shown in Table 1:

[0057] Table 1 Feature Calculation Methods

[0058]

[0059] Where, x i These are the normalized vibration signal data, i = 1, 2, 3, ..., n, where n is the number of sampling points, P(f k f is the spectrum of signal x(n), k = 1, 2, 3, ..., m, where m is the number of spectral lines; k It is the frequency value of the k-th spectral line;

[0060] The 12 extracted temporal features are used as 12 nodes to construct a temporal feature graph. The nodes are fully connected, and the initial weights of the edges are set to 1. The specific form of the fully connected temporal feature graph is as follows: Figure 2 As shown; the four extracted frequency domain features are used as four nodes to construct a frequency domain feature graph. The nodes are fully connected, and the initial weights of the edges are set to 1. The specific form of the fully connected frequency domain feature graph is as follows: Figure 3 As shown.

[0061] 3) Construct a rolling bearing remaining service life prediction model based on bi-graph attention and gated cyclic unit network for fully connected graphs;

[0062] 3.1) Constructing a dual-graph attention network based on the PyTorch deep learning framework:

[0063] like Figure 4 As shown, the dual-graph attention network includes two graph attention GAT layers, two fully connected FC layers, two ReLU layers, and a horzcat function. The temporal feature fully connected graph and the frequency domain feature fully connected graph constructed in step 2) are input into GAT1 and GAT2 layers, respectively. The input-output dimension of GAT1 layer is 12×64, and the input-output dimension of GAT2 layer is 4×64. At different training times, the weight coefficients of the edges of the fully connected graph in each GAT layer will change, thereby adaptively changing the structure of the fully connected graph in each GAT layer during training. This allows the dual-graph attention network to mine the spatial correlation information between graph node features at different times.

[0064] Because the number of nodes in the temporal feature fully connected graph and the frequency feature fully connected graph are different, resulting in different feature dimensions, they cannot be directly fused. Therefore, after each GAT layer, a fully connected layer is connected. The first fully connected layer has an input-output dimension of 768×64, and the second fully connected layer has an input-output dimension of 256×64. Through these two fully connected layers, the node feature matrix is ​​transformed into a graph feature matrix in graph units. Then, a ReLU activation function is connected to avoid the gradient vanishing problem. Finally, the horzcat function is used to fuse the temporal and frequency domain features that have been mined for spatial correlation, obtaining a deep fused feature that has extracted spatial correlation information. The calculation process of the dual-graph attention network is shown in the following formula:

[0065] H time =Relu(ω3·GAT1(F) time ,ω1)+b1)

[0066] H fre =Relu(ω4·GAT2(F) fre ,ω2)+b2)

[0067] H = horzcat(H) time H fre )

[0068] Among them, GAT1 and GAT2 are two GAT layers; F time F fre Here, ω1, ω2, ω3, and ω4 represent the time-domain and frequency-domain fully connected graphs of the input GAT layer, respectively; ω1, ω2, ω3, and ω4 are the first, second, third, and fourth weight matrices, respectively; b1 and b2 are the first and second bias matrices, respectively; ReLU is the activation function; H time H represents the time-domain features after activation. fre is the frequency domain feature after activation function; horzcat is the connection function; H is the deep fusion feature after extracting spatial correlation information;

[0069] 3.2) Constructing a gated recurrent unit network based on the PyTorch deep learning framework:

[0070] For deep fusion features that have extracted spatial correlation information, a gated recurrent unit (GRU) network is used to control information flow and capture the temporal dependency information between feature data, thereby constructing the degradation trend features of rolling bearings. The GRU network uses two gate controllers: a reset gate and an update gate, to perform forgetting and selective memorization of the current input information and the memory state information from previous time steps. The basic GRU unit structure at time t is as follows: Figure 5 As shown, the calculation and update steps are as follows:

[0071] 3.2.1) Calculate the value z of the update gate. t Update Gate Z t Control how to change the network cell state value h from the previous time step t-1 t-1 Compared with the input x at the current time t To combine;

[0072] z t =σ(W z ·[h t-1 x t ]+b z )

[0073] 3.2.2) Calculate the value r of the reset gate. t Reset door r t The state value h of the network element at the previous time t-1 t-1 Selectively forget historical information existing in the process;

[0074] r t =σ(W r •[h t-1 x t ]+b r )

[0075] 3.23) Calculate the candidate hidden state value c at the current time t. t ,

[0076] c t =tanh(W c ·[(r t ×h t-1 ), x t ]+b c )

[0077] 3.2.4) Calculate the hidden state output h at the current time t. t :

[0078] h t =(1-z)t )×h t-1 +z t ×c t

[0079] In the formula, W z W r W c z at time t t r t c t The weight value, b z b r b c z at time t t r t c t The bias, x t h is the input at the current time t. t-1 Let be the hidden state output at time t-1 on the network, σ be the sigmoid activation function with a value of (0, 1), tanh be the tanh activation function with a value of (-1, 1), and h be the hidden state output at time t-1 on the network. t The hidden state output at the current time t is used as the characteristic information of the degradation trend of the rolling bearing.

[0080] 4) Train a model to predict the remaining service life of rolling bearings;

[0081] 4.1) As Figure 6 As shown, the hyperparameters of the dual-graph attention network, gated recurrent unit network, and fully connected neural network, as well as the number of hidden layer nodes, are set; the learning rate of the model is specified as 0.0005, and the number of iterations during model training is 500.

[0082] 4.2) The training set preprocessed in step 1) is used to construct a fully connected graph of time-domain features and a fully connected graph of frequency-domain features in step 2). Then, the rolling bearing residual service life prediction model constructed in step 3) is used to obtain the rolling bearing degradation trend feature information. Finally, the linear layer of the fully connected neural network is used to fit the performance degradation trend of the rolling bearing, and the MSE function is selected as the loss function.

[0083]

[0084] Where j is the number of data samples, v i and These are the actual RUL value and the predicted RUL value, respectively.

[0085] The Adam optimization algorithm was used to optimize the hyperparameters of the dual-graph attention network, the gated recurrent unit network, and the fully connected neural network to predict the RUL of rolling bearings on the training set, thus obtaining a rolling bearing remaining service prediction model based on dual-graph attention and gated recurrent unit networks.

[0086] To evaluate the prediction accuracy of the model, this invention uses mean absolute error (MAE) and root mean square error (RMSE) as performance evaluation metrics, calculated as follows:

[0087]

[0088]

[0089] Where j is the number of data samples, v i and These are the actual RUL value and the predicted RUL value, respectively. The smaller the MAE and RMSE values, the higher the prediction accuracy of the model.

[0090] 5) Rolling bearing remaining service life prediction model prediction: A test set was constructed by selecting full-life-cycle vibration signal data of the rolling bearing under different working conditions. The accuracy of the prediction model was tested on the test set. The rolling bearing vibration signal data under different working conditions in the test set were normalized and input into the rolling bearing remaining service life prediction model obtained in step 4). The experimental results are shown in Table 2. Figure 8 , Figure 9 and Figure 10 As shown in Table 2, the MAE and RMSE values ​​of the experimental results for two rolling bearing RUL prediction methods are listed. The left column shows the results of the rolling bearing RUL prediction methods based on graph convolutional neural networks (GCN) and deep GRU networks, while the right column shows the results of the rolling bearing remaining service prediction method based on dual graph attention and gated recurrent unit networks proposed in this invention. Figure 8 , Figure 9 and Figure 10 The RUL prediction results obtained by applying the method proposed in this invention to test sets rolling bearings 1-3, 2-4, and 3-3 are shown respectively. (See Table 2 for details.) Figure 8 , Figure 9 and Figure 10 It can be seen that the method proposed in this invention can predict the RUL value of rolling bearings very well and has high prediction accuracy.

[0091] Table 2. RUL Test Results of Rolling Bearings

[0092]

[0093] This invention addresses the problem that existing graph neural network-based rolling bearing RUL prediction methods use a constant pre-constructed graph, which fails to identify changes in the spatial correlation of input features at different times, resulting in low accuracy in rolling bearing RUL prediction. It proposes a rolling bearing RUL prediction method based on dual-graph attention and gated recurrent unit networks. This method adaptively identifies the spatial correlation between input graph node features at different times, extracts deep-level features reflecting the rolling bearing degradation process, and predicts the rolling bearing RUL value more accurately. This ensures the safe and reliable operation of mechanical equipment, reduces equipment maintenance costs, and extends the equipment's operating cycle.

Claims

1. A method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU, characterized in that, Includes the following steps: 1) Obtain the periodic vibration signal data of the rolling bearing throughout its entire lifespan as a training set and perform preprocessing; 2) Fully connected graph construction: For the preprocessed training set, construct fully connected graphs for time-domain features and frequency-domain features respectively; 3) Construct a rolling bearing remaining service life prediction model based on dual-graph attention and gated recurrent unit network for fully connected graphs with time-domain and frequency-domain features; 4) Train a model to predict the remaining service life of rolling bearings; 5) Rolling bearing remaining service life prediction model prediction: Input the periodic vibration signal data of the rolling bearing under test throughout its entire life into the trained rolling bearing remaining service life prediction model to predict the remaining service life of the rolling bearing under test. The rolling bearing remaining service life prediction model in step 3) includes a dual-graph attention network and a gated cyclic unit network. The output of the dual-graph attention network is input into the gated cyclic unit network. The time correlation information between graph structures is obtained by the gated cyclic unit network and used as the characteristic information of rolling bearing degradation trend. The dual-graph attention network processes data in the following manner: Extracting spatial correlation information between graph node features: The dual-graph attention network consists of two GAT layers, two fully connected layers, two ReLU layers, and a horzcat function. The temporal and frequency domain fully connected graphs constructed in step 2.2) are input into the dual-graph attention network. The two GAT layers process the temporal and frequency domain fully connected graphs respectively. After processing, the corresponding fully connected layers transform the temporal and frequency domain fully connected graphs into the same dimension before fusing them to generate new features. This yields deep-level fused features with extracted spatial correlation information. The calculation process is as follows: ； ； ;in, These are two GAT layers respectively; These are the time-domain and frequency-domain fully connected graphs of the input GAT layer, respectively. These are the first, second, third, and fourth weight matrices, respectively. These are the first and second bias matrices, respectively; Relu For activation functions; The time-domain features are those after activation. The frequency domain features are those after activation. For connection functions; To extract deep fusion features that have undergone spatial correlation information; The gated recurrent unit network is processed in the following way: extracting the temporal correlation information between graph structures: the gated recurrent unit network includes a reset gate and an update gate at each time node. The deep fusion features output by the dual graph attention network are input into the reset gate and update gate at each time node, and the output of the reset gate and update gate is calculated by the sigmoid function. As the time node is updated, multiple reset gate outputs and multiple update gate outputs are obtained. The temporal correlation information between graph structures is updated by using the multiple reset gate outputs and multiple update gate outputs, and the temporal correlation information between graph structures is used as the characteristic information of the degradation trend of rolling bearings.

2. The method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU according to claim 1, characterized in that: Step 2) specifically refers to: 2.1) Feature extraction: Extract multiple time-domain features and multiple frequency-domain features from the periodic vibration signal data after preprocessing in step 1); 2.2) Graph construction: Take each of the multiple time-domain features and multiple frequency-domain features obtained in step 2.1) as a node of the graph structure, and construct the time-domain feature fully connected graph and the frequency-domain feature fully connected graph respectively.

3. The method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU according to claim 2, characterized in that: The training of the rolling bearing remaining service life prediction model in step 4) is specifically as follows: the training set after preprocessing in step 1) is used to construct a time-domain feature fully connected graph and a frequency-domain feature fully connected graph in step 2), and then the rolling bearing remaining service life prediction model constructed in step 3) is used to obtain the rolling bearing degradation trend feature information. Finally, the linear layer of the fully connected neural network is used to fit the rolling bearing degradation trend feature information. The MSE function is selected as the loss function of the rolling bearing remaining service life prediction model. The Adam algorithm is used to optimize the hyperparameters in the dual-graph attention network, the gated recurrent unit network and the fully connected neural network, and finally the trained rolling bearing remaining service life prediction model is obtained.

4. The method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU according to claim 1, characterized in that: The specific steps of the rolling bearing remaining service life prediction model test in step 5) are as follows: the periodic vibration signal data of the rolling bearing under test throughout its entire lifespan is used as a test set for normalization processing, and then input into the trained rolling bearing remaining service life prediction model to achieve the prediction of the remaining service life of the rolling bearing under test.

5. The method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU according to claim 1, characterized in that: Step 1) specifically involves: normalizing the periodic vibration signal data of the rolling bearing throughout its entire lifespan using the Z-score function, as follows: ; in, The vibration signal data before normalization. The vibration signal data is normalized. The mean of the vibration signal data. This represents the standard deviation of the vibration signal data.

6. The method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU according to claim 2, characterized in that: In step 2.1), multiple time-domain features include a variety of features such as maximum value, minimum value, peak-to-peak value, mean, variance, standard deviation, root mean square value, root mean square amplitude, kurtosis factor, skewness factor, waveform factor, and margin factor. Multiple frequency domain features include mean frequency, centroid frequency, root mean square frequency, and frequency standard deviation.

7. A method for predicting the remaining service life of rolling bearings based on dual-graph attention and GRU according to claim 2 or 6, characterized in that: In step 2.1), the extraction of multiple frequency domain features specifically involves: After converting time-domain periodic vibration signal data into frequency-domain vibration signal data using the Fast Fourier Transform algorithm, multiple frequency-domain features of the frequency-domain vibration signal data are extracted.

Images