A bearing residual service life prediction method based on stage-aware dynamic graph attention network

Abstract

This invention belongs to the field of mechanical engineering, and specifically relates to a bearing remaining service life (RUL) prediction method based on a stage-aware dynamic graph attention network. The purpose of this invention is to address the problem of low accuracy in existing bearing RUL prediction methods. A bearing RUL prediction method based on a fused graph attention network, a dynamic graph modeling mechanism, and multi-channel feature fusion using multi-head attention features is proposed. This method constructs a time-series graph structure using a sliding window approach, representing each vibration signal segment as a node in the graph. Local spatiotemporal degradation features are learned through graph convolution, and a learnable projection matrix is ​​introduced to achieve adaptive structural adjustment, thereby enhancing the modeling ability of structural changes during the degradation process and solving the problem of low accuracy in existing bearing RUL prediction methods.

Description

Technical Field

[0001] This invention belongs to the field of mechanics, and in particular relates to a method for predicting the remaining service life of bearings based on a stage-aware dynamic graph attention network. Background Technology

[0002] Existing methods for predicting the remaining useful life (RUL) of bearings, while employing deep models such as LSTM or TCN to improve prediction accuracy, rely solely on short-term or long-term temporal features, making it difficult to comprehensively capture the dynamic spatiotemporal patterns of bearing degradation. Furthermore, the success of GNNs in fault detection has not been fully extended to RUL prediction, particularly in handling full lifecycle data and domain adaptation. In addition, while transfer learning has made progress, insufficient optimization of feature mapping mechanisms limits the model's generalization performance and accuracy. Summary of the Invention

[0003] The purpose of this invention is to address the problem of low accuracy in predicting the remaining service life of bearings in existing methods. A method for predicting the remaining service life of bearings based on a stage-aware dynamic graph attention network is provided, comprising:

[0004] Step 1: Real-time acquisition of vibration signals during the operation of the bearing under test, wherein the real-time acquisition of vibration signals during the operation of the bearing under test is a time-series signal;

[0005] The vibration signals collected during the operation of the bearing under test include: horizontal vibration signals and vertical vibration signals during the operation of the bearing under test.

[0006] According to the preset time length, the horizontal vibration signal of the bearing to be tested during its operation is divided into T time periods of horizontal vibration signal to be tested.

[0007] According to the preset time length, the vertical vibration signal of the bearing under test during its operation is divided into T time periods of vertical vibration signal under test.

[0008] The time length of each segment of the vibration signal to be detected after segmentation is the same as the segment length in the training set;

[0009] In addition, the total time for real-time acquisition of vibration signals during the operation of the bearing under test is fixed. After the latest vibration signal is acquired, the earliest recorded signal is deleted, and the vibration signals during the operation of the bearing under test are updated in a rolling manner; only the vibration signals of the most recent T time periods are retained.

[0010] Step 2: Based on the vibration signals in the horizontal direction of the bearing to be tested over T time periods, extract B1 sensitive feature vectors in the horizontal direction of the bearing to be tested.

[0011] Based on the vibration signals in the vertical direction to be detected over T time periods, B2 sensitive feature vectors in the vertical direction of the bearing to be detected are extracted.

[0012] Based on the B1 sensitive feature vectors in the horizontal direction of the bearing to be tested, construct T initial dynamic graph structure data in the horizontal direction of the bearing to be tested.

[0013] Based on the B2 sensitive feature vectors in the vertical directions of the bearing to be tested, construct T initial dynamic graph structure data in the vertical directions of the bearing to be tested.

[0014] The one-dimensional health index value of the bearing under test is obtained based on the vibration signals of the B1 horizontal sensitive feature vectors and the B2 vertical sensitive feature vectors of the bearing under test.

[0015] The stage label of the bearing to be tested can be obtained based on the one-dimensional health index value of the bearing to be tested;

[0016] In step two, the initial graph of the dynamic graph is constructed at the beginning. It is called a dynamic graph because the graph structure changes at each layer when it is passed into the network in step three.

[0017] The process of constructing graph structure data for the bearing to be tested from the data of the t-th time segment includes:

[0018] First, feature vectors are extracted from the vibration signals at each time step in the data of the t-th time segment; and the feature vectors of the vibration signals at each time step in the data of the t-th time segment are defined as nodes of the graph.

[0019] Among them, the features of the initial dynamic graph structure data in the horizontal direction are B1 sensitive features in the horizontal direction determined during the training process; the features of the initial dynamic graph structure data in the vertical direction are B2 sensitive features in the vertical direction determined during the training process.

[0020] Then, the cosine similarity (temporal distance) between nodes is calculated. For each node in the graph, the K nodes with the highest similarity are selected to establish connection edges, generating the initial adjacency matrix A corresponding to that time segment. The above process is repeated to construct the corresponding node feature matrix and adjacency matrix for each outer time segment, completing the construction of the t-th vertical temporal feature map sequence; the connection relationships and weights between all nodes are quantized into the initial adjacency matrix.

[0021] Finally, the feature matrix containing node feature information and the initial adjacency matrix representing the topology are used to construct graph structure data, which is then used as input to the SAGAT model.

[0022] Step 3: Input the initial dynamic graph structure data of the horizontal direction of T bearings to be tested, the initial dynamic graph structure data of the vertical direction of T bearings to be tested, and the stage labels of the bearings to be tested into the pre-trained bearing service life prediction model based on the stage-aware dynamic graph attention network for processing, and obtain the predicted value of the remaining service life of the bearings to be tested.

[0023] The stage label of the bearing to be tested is used to determine which SAGAT model the initial dynamic graph structure data is input into;

[0024] When the stage label data is 0, input the first SAGAT network module;

[0025] When the stage label data is 1, input it into the second SAGAT network module;

[0026] When the stage label data is 2, input it into the third SAGAT network module;

[0027] The training process of the bearing life prediction model based on a stage-aware dynamic graph attention network pre-trained in step three is as follows:

[0028] S1. Collect horizontal vibration signal data of the bearing throughout its entire life cycle, vertical vibration signal data of the bearing throughout its entire life cycle, and RUL life tag data.

[0029] A training set is constructed based on the horizontal vibration signal data of the bearing throughout its entire life cycle, the vertical vibration signal data of the bearing throughout its entire life cycle, and the RUL life tag data.

[0030] S2. Construct a bearing life prediction model based on a stage-aware dynamic graph attention network; including: a first SAGAT network module, a second SAGAT network module, a third SAGAT network module, a weighted fusion module based on a multi-head attention mechanism, and a regression prediction layer;

[0031] S3. Train the bearing life prediction model based on the stage-aware dynamic graph attention network according to the training set to obtain the trained bearing life prediction model based on the stage-aware dynamic graph attention network.

[0032] The beneficial effects of this invention are as follows:

[0033] A bearing RUL prediction method is proposed, which integrates Graph Attention Network (GAT), dynamic graph modeling mechanism, and multi-head attention features to fuse multi-channel features. This method constructs a time-series graph structure using a sliding window approach, representing each vibration signal segment as a node in the graph. It learns local spatiotemporal degradation features through graph convolution and introduces a learnable projection matrix to achieve adaptive structural adjustment, thereby enhancing the modeling ability of structural changes during the degradation process.

[0034] Furthermore, this paper designs a multi-head attention fusion module to capture the differences in importance of features across different channels in the graph, enabling explicit modeling of cross-channel semantic information and improving the model's global expressive power.

[0035] Furthermore, this paper introduces a stage-aware mechanism based on Geometric Graphs (GMM) before the model training phase. By performing soft clustering on the constructed health indicator sequences, the degradation stages of the bearing operation process are automatically classified, and the dynamic construction of the dynamic graph structure is guided, enabling the model to have stage-aware capabilities and enhancing its adaptability to modeling node adjacency relationships under different degradation states. To verify the effectiveness of the proposed method, a systematic experiment was conducted on a publicly available bearing life dataset. The results show that compared with existing mainstream methods, the proposed model exhibits significant advantages in prediction accuracy, robustness, and generalization ability. In particular, it can maintain stable performance under varying operating conditions, verifying the application potential of the method in practical engineering. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall architecture of the method of the present invention. Detailed Implementation

[0037] Specific implementation method one: Combining Figure 1 This invention is described;

[0038] Step 1: Real-time acquisition of vibration signals during the operation of the bearing under test, wherein the real-time acquisition of vibration signals during the operation of the bearing under test is a time-series signal;

[0039] The vibration signals collected during the operation of the bearing under test include: horizontal vibration signals and vertical vibration signals during the operation of the bearing under test.

[0040] According to the preset time length, the horizontal vibration signal of the bearing to be tested during its operation is divided into T time periods of horizontal vibration signal to be tested.

[0041] According to the preset time length, the vertical vibration signal of the bearing under test during its operation is divided into T time periods of vertical vibration signal under test.

[0042] The time length of each segment of the vibration signal to be detected after segmentation is the same as the segment length in the training set;

[0043] In addition, the total time for real-time acquisition of vibration signals during the operation of the bearing under test is fixed. After the latest vibration signal is acquired, the earliest recorded signal is deleted, and the vibration signals during the operation of the bearing under test are updated in a rolling manner; only the vibration signals of the most recent T time periods are retained.

[0044] Step 2: Based on the vibration signals in the horizontal direction of the bearing to be tested over T time periods, extract B1 sensitive feature vectors in the horizontal direction of the bearing to be tested.

[0045] Based on the vibration signals in the vertical direction to be detected over T time periods, B2 sensitive feature vectors in the vertical direction of the bearing to be detected are extracted.

[0046] Based on the B1 sensitive feature vectors in the horizontal direction of the bearing to be tested, construct T initial dynamic graph structure data in the horizontal direction of the bearing to be tested.

[0047] Based on the B2 sensitive feature vectors in the vertical directions of the bearing to be tested, construct T initial dynamic graph structure data in the vertical directions of the bearing to be tested.

[0048] The one-dimensional health index value of the bearing under test is obtained based on the vibration signals of the B1 horizontal sensitive feature vectors and the B2 vertical sensitive feature vectors of the bearing under test.

[0049] The stage label of the bearing to be tested can be obtained based on the one-dimensional health index value of the bearing to be tested;

[0050] In step two, the initial graph of the dynamic graph is constructed at the beginning. It is called a dynamic graph because the graph structure changes at each layer when it is passed into the network in step three.

[0051] The process of constructing graph structure data for the bearing to be tested from the data of the t-th time segment includes:

[0052] First, feature vectors are extracted from the vibration signals at each time step in the data of the t-th time segment; and the feature vectors of the vibration signals at each time step in the data of the t-th time segment are defined as nodes of the graph.

[0053] Among them, the features of the initial dynamic graph structure data in the horizontal direction are B1 sensitive features in the horizontal direction determined during the training process; the features of the initial dynamic graph structure data in the vertical direction are B2 sensitive features in the vertical direction determined during the training process.

[0054] Then, the cosine similarity (temporal distance) between nodes is calculated. For each node in the graph, the K nodes with the highest similarity are selected to establish connection edges, generating the initial adjacency matrix A corresponding to that time segment. The above process is repeated to construct the corresponding node feature matrix and adjacency matrix for each outer time segment, completing the construction of the t-th vertical temporal feature map sequence; the connection relationships and weights between all nodes are quantized into the initial adjacency matrix.

[0055] Finally, the feature matrix containing node feature information and the initial adjacency matrix representing the topology are used to construct graph structure data, which is then used as input to the SAGAT model.

[0056] Furthermore, the process of obtaining the one-dimensional health index value of the bearing under test based on the vibration signals of the B1 horizontal sensitive feature vectors and the B2 vertical sensitive feature vectors of the bearing under test is the same as the training process.

[0057] Step 3: Input the initial dynamic graph structure data of the horizontal direction of T bearings to be tested, the initial dynamic graph structure data of the vertical direction of T bearings to be tested, and the stage labels of the bearings to be tested into the pre-trained bearing service life prediction model based on the stage-aware dynamic graph attention network for processing, and obtain the predicted value of the remaining service life of the bearings to be tested.

[0058] The stage label of the bearing to be tested is used to determine which SAGAT model the initial dynamic graph structure data is input into;

[0059] When the stage label data is 0, input the first SAGAT network module;

[0060] When the stage label data is 1, input it into the second SAGAT network module;

[0061] When the stage label data is 2, input the third SAGAT network module.

[0062] Specific Implementation Method Two: The difference between this implementation method and Specific Implementation Method One is that:

[0063] The training process of the bearing life prediction model based on a stage-aware dynamic graph attention network pre-trained in step three is as follows:

[0064] S1. Collect horizontal vibration signal data of the bearing throughout its entire life cycle, vertical vibration signal data of the bearing throughout its entire life cycle, and RUL life tag data.

[0065] A training set is constructed based on the horizontal vibration signal data of the bearing throughout its entire life cycle, the vertical vibration signal data of the bearing throughout its entire life cycle, and the RUL life tag data.

[0066] S2. Construct a bearing life prediction model based on a stage-aware dynamic graph attention network; including: a first SAGAT network module, a second SAGAT network module, a third SAGAT network module, a weighted fusion module based on a multi-head attention mechanism, and a regression prediction layer;

[0067] S3. Train the bearing life prediction model based on the stage-aware dynamic graph attention network according to the training set to obtain the trained bearing life prediction model based on the stage-aware dynamic graph attention network; other steps and parameters are the same as in Specific Implementation Method 1.

[0068] Specific Implementation Method Three: The difference between this implementation method and Specific Implementation Methods One and Two is that:

[0069] In step S1, a training set is constructed based on the horizontal vibration signal data and the vertical vibration signal data of the bearing throughout its entire lifespan; the specific process is as follows:

[0070] S1.1: Perform data preprocessing on the horizontal vibration signal data of the bearing throughout its entire life cycle to obtain preprocessed horizontal vibration signal data;

[0071] The preprocessed horizontal vibration signal data includes: preprocessed horizontal vibration signal data for T time periods;

[0072] The vertical vibration signal data of the bearing throughout its entire life cycle is preprocessed to obtain the preprocessed vertical vibration signal data.

[0073] The preprocessed vertical vibration signal data includes: preprocessed vertical vibration signal data for T time periods;

[0074] The specific process is as follows:

[0075] First, the vibration signal sequence is divided into segments using a sliding time window of length U and step size 1 to obtain T segments of vibration signal to be detected, which serve as the preprocessed horizontal vibration signal data; the processing steps for the horizontal and vertical directions are the same.

[0076] S1.2: Extract A1 horizontal time-domain features and A2 horizontal frequency-domain features from the preprocessed horizontal vibration signal data;

[0077] Extract A3 vertical time-domain features and A4 vertical frequency-domain features from the preprocessed vertical vibration signal data;

[0078] A1, A2, A3, and A4 are all positive integers;

[0079] The time-domain features include: mean, effective value (RMS), variance, peak value, peak-to-peak value, kurtosis, skewness, waveform factor, peak factor, impulse factor, margin factor, approximate entropy, etc.

[0080] The frequency domain features include: frequency center, root mean square frequency, frequency variance, spectral kurtosis, spectral entropy, characteristic frequency amplitude, and frequency band energy proportion, etc.

[0081] S1.3: Perform correlation analysis on the time-domain features of A1 horizontal directions and the frequency-domain features of A2 horizontal directions to screen out the sensitive features of B1 horizontal directions;

[0082] Correlation analysis was performed on the time-domain features of A3 vertical directions and the frequency-domain features of A4 vertical directions to obtain B2 sensitive features in the vertical directions; B1 and B2 are both positive integers.

[0083] This invention first extracts candidate health indicators from sliding window features, and then uses Pearson and Spearman correlation to jointly screen redundant features.

[0084] S1.4: Kernel principal component analysis (KPC) is used to reduce the dimensionality of the B1 horizontal and B2 vertical sensitive features, resulting in a one-dimensional health indicator sequence HI; expressed by the formula:

[0085]

[0086] This represents the health indicator values ​​for the first time period. This represents the health indicator values ​​for the second period. This represents the health indicator value for the T-th time period.

[0087] The sensitive features in the B1 horizontal direction and the B2 vertical direction were used for dimensionality reduction using kernel principal component analysis (KPCA), and then fused using an autoencoder to obtain a unified HI sequence. ;

[0088] This represents the health indicator values ​​for the first time period. This represents the health indicator values ​​for the second period. This represents the health indicator value for the T-th time period.

[0089] Input the filtered sensitive features into KPCA.

[0090] Since equipment degradation is non-linear, linear PCA may not be effective. KPCA, by mapping to a high-dimensional space through a kernel function and then reducing the dimensionality, can better extract non-linear degradation information and finally fuse it into a one-dimensional health indicator (HI) sequence. This sequence should clearly decrease or increase monotonically from 1 (health) to 0 (failure).

[0091] S1.5: Perform clustering processing on the one-dimensional health indicator sequence HI to obtain the clustering results of the one-dimensional health indicator sequence HI.

[0092] Stage label data is constructed based on the clustering results of the one-dimensional health indicator sequence HI;

[0093] This invention divides the entire life cycle of bearings into a healthy stage, a mild degradation stage, and a severe degradation stage based on the clustering results of the one-dimensional health index sequence HI.

[0094] Preferably, one method of clustering the one-dimensional health index sequence HI according to the present invention is to use a moving average filtering algorithm to perform time-series smoothing on the fused one-dimensional health index sequence HI to eliminate high-frequency local noise and enhance the monotonicity of the degradation trend, use a Gaussian mixture model (GMM) to perform clustering modeling on the smoothed HI sequence, use the expectation-maximization (EM) algorithm to estimate the model parameters, and finally obtain the clustering results.

[0095] The overall probability density function of the Gaussian mixture model is:

[0096] (0.2)

[0097] Where x represents health indicator data, corresponding to the one-dimensional health indicator sequence HI of this invention, and K represents the number of Gaussian components. In this embodiment, K=3. This represents the mixing weight of the k-th Gaussian component. Let the probability density function of the k-th Gaussian component be denoted as . and The mean and covariance of the k-th Gaussian component are respectively.

[0098] The parameters of the Gaussian mixture model are obtained using the expectation-maximization algorithm. , and Perform iterative estimation, and based on the estimated parameters, divide the HI sequence into three stages:

[0099] Phase 0 (Healthy Phase): The HI value is high and remains stable in this phase;

[0100] Stage 1 (mild degradation stage): The HI value in this stage shows a slow downward trend;

[0101] Phase 2 (Severe Degradation Phase): In this phase, the HI index drops rapidly, approaching the failure point;

[0102] S1.6: Based on the preprocessed horizontal vibration signal data and B1 sensitive features in the horizontal direction obtained from S1.1, construct a sequence of time-series feature maps in the horizontal direction;

[0103] The sequence of temporal feature maps in the horizontal direction includes: T temporal feature maps in the horizontal direction;

[0104] Specifically, based on the preprocessed horizontal vibration signal data of the t-th time period and B1 sensitive features of the horizontal direction, a time-series feature map of the t-th horizontal direction is constructed.

[0105] Based on the preprocessed vertical vibration signal data obtained from S1.1 and the B2 sensitive features in the vertical direction, a sequence of time-series feature maps in the vertical direction is constructed.

[0106] The sequence of temporal feature maps in the vertical direction includes: T temporal feature maps in the vertical direction;

[0107] Specifically, based on the preprocessed vertical vibration signal data of the t-th time period and the B2 vertical sensitive features, a time-series feature map of the t-th vertical direction is constructed.

[0108] S1.7: Construct a training set based on the horizontal temporal feature map sequence, the vertical temporal feature map sequence, stage label data, and RUL lifetime labels;

[0109] The temporal feature map sequences in the horizontal and vertical directions of the training set are used as input data for the subsequent SAGAT network; the lifetime RUL lifetime label is used to calculate the loss and update the network parameters; the stage label data is used to determine which SAGAT network the temporal feature map is input into; other steps and parameters are the same as in one of the specific implementation methods one or two.

[0110] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that:

[0111] In step S1.6, a time-series feature map of the t-th horizontal direction is constructed based on the preprocessed horizontal vibration signal data of the t-th time period and B1 sensitive features of the horizontal direction; the specific process is as follows:

[0112] S1.6.1.1: Use a sliding window to segment the preprocessed horizontal vibration signal data in the t-th time period to obtain S time slices of horizontal vibration signal data;

[0113] S1.6.1.2: Define the S time slices as the S nodes of the time series feature map in the t-th horizontal direction;

[0114] S1.6.1.3: Construct the feature vectors of the S nodes of the temporal feature map in the t-th horizontal direction based on the sensitive features in the B1 horizontal directions;

[0115] Construct the node feature matrix of the time series feature map in the t-th horizontal direction based on the feature vectors of the S nodes of the time series feature map in the t-th horizontal direction; where d is the feature dimension;

[0116] The t-th time period of the preprocessed horizontal vibration signal data Within a time window, the extracted node feature vector can be defined as Node feature matrix

[0117] The node feature matrix of the time series feature map in the t-th horizontal direction is expressed by the formula:

[0118] (3)

[0119] S is the number of sampling segments within the sliding window, i.e., the number of nodes in the graph, and d is the feature dimension of each node;

[0120] S1.6.1.4: Based on the feature vectors of the S nodes of the time series feature map in the t-th horizontal direction, calculate the cosine similarity between the feature vectors of any two nodes in the time series feature map in the t-th horizontal direction.

[0121] Based on the cosine similarity between the feature vectors of any two nodes in the temporal feature map of the t-th horizontal direction, the edges of the graph are constructed using the K-Nearest Neighbor (KNN) strategy to obtain the temporal feature map of the t-th horizontal direction.

[0122] For each node in the graph, select the K nodes with the highest similarity to establish connecting edges, and generate the initial adjacency matrix A corresponding to that time segment. Repeat the above process to construct the corresponding node feature matrix and adjacency matrix for each outer time segment, and complete the construction of the time series feature map sequence in the t-th horizontal direction; other steps and parameters are the same as in one of the specific implementation methods one to three.

[0123] Specific Implementation Method Five: The difference between this implementation method and Specific Implementation Methods One to Four is that:

[0124] In step S1.6, based on the preprocessed vertical vibration signal data of the t-th time period and the B2 vertical sensitive features, a time-series feature map of the t-th vertical direction is constructed; the specific process is as follows:

[0125] S1.6.2.1: Use a sliding window to segment the preprocessed vertical vibration signal data of the t-th time period to obtain S time slices of vertical vibration signal data;

[0126] S1.6.2.2: Define the S time slices as the S nodes of the time series feature map in the t-th vertical direction;

[0127] S1.6.2.3: Construct the feature vectors of the S nodes of the temporal feature map in the t-th vertical direction based on the sensitive features in the B1 vertical directions;

[0128] Construct the node feature matrix of the temporal feature map in the t-th vertical direction based on the feature vectors of the S nodes in the t-th vertical direction; where d is the feature dimension.

[0129] The t-th time period of the preprocessed vertical vibration signal data Within a time window, the extracted node feature vector can be defined as ,

[0130] The node feature matrix of the temporal feature map in the t-th vertical direction is expressed by the formula:

[0131] (3)

[0132] S is the number of sampling segments within the sliding window, i.e., the number of nodes in the graph, and d is the feature dimension of each node;

[0133] S1.6.2.4: Based on the feature vectors of the S nodes of the time series feature map in the t-th vertical direction, calculate the cosine similarity between the feature vectors of any two nodes in the time series feature map in the t-th vertical direction.

[0134] Based on the cosine similarity between the feature vectors of any two nodes in the temporal feature map of the t-th vertical direction, the edges of the graph are constructed using the K-Nearest Neighbor (KNN) strategy to obtain the temporal feature map of the t-th vertical direction.

[0135] For each node in the graph, select the K nodes with the highest similarity to establish connection edges, generating the initial adjacency matrix A corresponding to that time segment. Repeat the above process to construct the corresponding node feature matrix and adjacency matrix for each outer time segment, completing the construction of the t-th vertical time series feature map sequence;

[0136] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0137] Specific Implementation Method Six: The difference between this implementation method and Specific Implementation Methods One through Five is that:

[0138] In step S3, the bearing life prediction model based on the stage-aware dynamic graph attention network is trained using the training set to obtain the trained bearing life prediction model based on the stage-aware dynamic graph attention network. The specific process is as follows:

[0139] S3.1: Input the temporal feature map of the horizontal direction in the training set into the corresponding SAGAT network module according to the stage label data for feature extraction processing to obtain the deep degradation feature map of the horizontal channel;

[0140] The temporal feature maps in the vertical direction of the training set are input into the corresponding SAGAT network module for feature extraction based on the stage label data, so as to obtain the deep degradation feature maps of the vertical channel.

[0141] When the stage label data is 0, input the first SAGAT network module;

[0142] When the stage label data is 1, input it into the second SAGAT network module;

[0143] When the stage label data is 2, input it into the third SAGAT network module;

[0144] S3.2: The deep degradation feature maps of the horizontal channel and the deep degradation feature maps of the vertical channel are concatenated to obtain a fused feature matrix;

[0145] The fused feature matrix is ​​input into a weighted fusion module based on a multi-head attention mechanism for weighted fusion processing to obtain the fused global spatiotemporal feature map;

[0146] S3.3: Input the fused global spatiotemporal feature map into the regression prediction layer and output the predicted value of the remaining service life of the bearing;

[0147] The specific process is as follows: the fused global spatiotemporal features Z are input into the regression prediction layer (fully connected layer), and the predicted remaining service life of the bearing at the current moment is output through linear mapping. To ensure the continuity of the prediction results, no nonlinear activation function is used after the fully connected layer;

[0148] S3.4: Calculate the loss function based on the predicted remaining service life of the bearing and the RUL life labels in the training set. Train the bearing service life prediction model based on the stage-aware dynamic graph attention network according to the loss function. Stop training when the loss function converges to obtain the trained bearing service life prediction model based on the stage-aware dynamic graph attention network. Other steps and parameters are the same as in one of the specific implementation methods one to five.

[0149] Specific Implementation Method Seven: The difference between this implementation method and Specific Implementation Methods One through Six is ​​that:

[0150] In step S3.1, the temporal feature map of the horizontal direction in the training set is input into the SAGAT network module for feature extraction to obtain the deep degradation feature map of the horizontal channel; the specific process is expressed by the formula:

[0151] ;

[0152] In the formula, This represents the node feature matrix of the time-series feature map in the t-th horizontal direction; Represents the attention weighting coefficient of the l-th layer. This represents the attention weighting coefficient for the second layer. This represents the attention weighting coefficient for the third layer. This represents the activation function. This represents the intermediate feature map at the first level. This represents the intermediate feature map at the second level. This represents a deep degradation feature map of the horizontal channel; Denotes the parameter matrix of the zeroth filter. This represents the parameter matrix of the first filter. This represents the parameter matrix of the second filter; Indicates the zeroth bias term; Indicates the first bias term. Indicates the second bias term. Denotes the first adaptive projection matrix. Represents the second adaptive projection matrix

[0153] The element in the j-th row and k-th column of the attention weighting coefficients The representation is the attention weight between node j and its neighbor node k; The calculation formula is:

[0154]

[0155] In the formula, The unnormalized attention coefficient between node j and its neighbor node k is represented by exp; exp represents the exponential function. This represents the set of neighboring nodes of node j; softmax represents the softmax function.

[0156] The unnormalized attention coefficient between node j and its neighbor node k The calculation formula is:

[0157]

[0158] In the formula, This indicates a splicing operation. express transpose, The learnable weight vector represents the attention mechanism; LeakyReLU represents the non-linear activation function.

[0159] The intermediate mapping variable representing node j; The intermediate mapping variable representing node k;

[0160] The intermediate mapping variable for node j is obtained based on the characteristics of node j, and the intermediate mapping variable for node k is obtained based on the characteristics of node k; expressed by the formula:

[0161]

[0162]

[0163] In the formula, Represents the linear mapping weight matrix. The feature vector representing node j in the (l-1)th horizontal (vertical) intermediate feature map; The feature vector representing node k in the (l-1)th horizontal (vertical) intermediate feature map;

[0164] The process involves inputting the temporal feature map of the vertical direction in the training set into the SAGAT network module for feature extraction to obtain the deep degradation feature map of the vertical channel. The specific process is expressed by the following formula:

[0165] ;

[0166] In the formula, Let represent the node feature matrix of the time series feature map in the t-th vertical direction. This represents the intermediate feature map at the first level. This represents the intermediate feature map at the second level. This represents a deep degradation feature map of the horizontal channel;

[0167] The other steps and parameters are the same as those in any of the specific implementation methods one to six.

[0168] Specific Implementation Method Eight: The difference between this implementation method and Specific Implementation Methods One through Seven is that:

[0169] The weighted fusion module based on the multi-head attention mechanism in S3.2 includes H attention heads;

[0170] The fused feature matrix is ​​input into a weighted fusion module based on a multi-head attention mechanism for weighted fusion processing to obtain a fused global spatiotemporal feature map; the specific process is as follows:

[0171] S3.2.1: Input the fused feature map into H attention heads for processing to obtain the intermediate attention output map of H attention heads;

[0172] Among them, the feature maps will be fused. The h-th attention head is processed to obtain the intermediate attention output graph of the h-th attention head. ;

[0173] S3.2.2: The intermediate attention output maps of the H attention heads are concatenated to obtain the concatenated features; expressed by the formula:

[0174]

[0175] In the formula, Indicates the features after splicing. This represents the intermediate attention output diagram of the first attention head; This represents the intermediate attention output diagram of the first attention head; This represents the intermediate attention output graph of the Hth attention head;

[0176] S3.2.3: Process the concatenated features from S3.2.2 using a linear transformation matrix to obtain a global spatiotemporal feature map; other steps and parameters are the same as in one of the specific implementation methods one to seven.

[0177] Specific Implementation Method Nine: The difference between this implementation method and Specific Implementation Methods One through Eight is that:

[0178] In S3.2.1, the feature maps will be fused. The h-th attention head is processed to obtain the intermediate attention output graph of the h-th attention head. The formula is as follows:

[0179] ;

[0180] ;

[0181] ;

[0182] In the formula, This represents the query projection matrix of the h-th attention head. This represents the key projection matrix of the h-th attention head. The projection matrix represents the value of the h-th attention head. This represents the query vector for the h-th attention head. This represents the query vector for the h-th attention head. This represents the query vector for the h-th attention head. This represents the attention distribution matrix of the h-th attention head. Indicated Transpose This represents the feature dimension of the h-th attention head;

[0183] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0184] Specific Implementation Method Ten: The difference between this implementation method and Specific Implementation Methods One through Nine is that:

[0185] The loss function in S3.4 is the mean squared error function; the formula for the mean squared error function is:

[0186]

[0187] Where M is the total number of samples; This represents the actual remaining useful life (RUL) value. Predict the remaining useful life (RUL) value for the model; update all parameters of the network end-to-end by minimizing the loss function.

[0188] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0189] To verify the effectiveness of the proposed method, experiments were conducted on a publicly available dual-channel bearing life dataset. During the experiments, vibration signal sequences from this dataset were used to construct the model input. The model training phase involved dividing the dataset into training and test sets, and the prediction error was calculated using the true bearing life (RUL) as the label. Commonly used metrics such as root mean square error (RMSE) and mean absolute error (MAE) were employed to compare the performance of the proposed model with existing methods.

[0190] Case 1: RUL Prediction on the XJTU-SY Dataset

[0191] Dataset description:

[0192] The XJTU-SY dataset was provided by Xi'an Jiaotong University and Zhejiang Changxing Shuangyang Technology Co., Ltd. This dataset contains the full-life degradation process of 15 bearings, run under three different operating conditions (C1, C2, C3). Each operating condition consists of a different combination of speed and load.

[0193] Each bearing was continuously operated under constant conditions until failure, during which vibration signals were acquired at fixed time intervals. The signal sampling frequency was 32.768 kHz, with one sampling point per minute, and each signal group had a length of 40,960 points, including outputs from acceleration sensors in both the horizontal and vertical directions. This dataset provides the full-cycle degradation trajectory of the vibration signal, which helps to capture subtle features during the failure evolution process.

[0194] Parameter settings and evaluation indicators:

[0195] The original vibration signal is segmented using a sliding window W. A set of time-frequency domain feature vectors is extracted for each signal segment as input features for graph nodes. The sliding window length is set to 512, the step size to 256, and the maximum number of nodes is limited to 50 to avoid excessive graph expansion. For the graph structure constructed for each signal segment, the cosine similarity between nodes is first calculated based on the time-domain and frequency-domain features. Then, a K-nearest neighbor strategy is used to construct graph edges. Adjacent node pairs are explicitly connected to preserve time-series features. Simultaneously, for each node, K=5 of the most similar non-adjacent nodes are selected for edge connection, thereby enhancing the local semantic expressiveness of the graph structure.

[0196] For the vibration data at each moment, independent graph structures are constructed from both horizontal and vertical channels, and each is assigned a channel identifier to achieve multi-source information fusion. Each graph is associated with a corresponding degradation label, with the label value decreasing from 1.0 to 0.0 using a linear function to characterize the degree of degradation of the equipment from initial operation to complete failure. To ensure the rigor of the model evaluation, the graph data partitioning strategy strictly groups data according to the individual bearing dimension. For bearing samples under each operating condition, each bearing is selected as the test set, and the two bearings with the closest lifespan length to the bearings in the test set are selected as the training set to ensure that the test data does not overlap with the training set at the individual bearing level.

[0197] Evaluation indicators:

[0198] This invention employs four evaluation metrics to quantify the model's performance in the Remaining Life (RUL) prediction task: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), r², and MSE. RMSE simultaneously measures the dispersion of early and late-stage remaining lifetime predictions; MAE accurately reflects the magnitude of the actual RUL prediction error; and r² intuitively represents the fitting accuracy of the RUL prediction model.

[0199] Experimental results

[0200] (1) Necessity analysis of stage classification

[0201] To verify the importance of GMM-based degradation stage segmentation for RUL prediction, this invention designed a comparative experiment before and after stage segmentation. While maintaining a consistent model structure, RUL predictions were performed under two modes: without stage segmentation and with stage segmentation, and the prediction curves and error indices were compared.

[0202] In the early healthy stage (stage 0) and the intermediate mild degradation stage (stage 1), due to the gradual changes in signal characteristics, the model has difficulty in effectively extracting the stage degradation patterns, which makes it difficult for the model to predict the correct RUL, resulting in large errors and insufficient sensitivity to mutation features.

[0203] In contrast, by dividing the entire degradation process into three stages—healthy, mildly degraded, and severely degraded—the model performs local modeling within each stage, enabling it to better capture the characteristic evolution patterns of that stage. Compared to not using stage classification, the model significantly improves its ability to capture features in the early and middle stages of degradation. Especially in stage 1, the model more accurately captures the rapid decline trend of RUL, and the predicted curve is highly consistent with the actual degradation trajectory, demonstrating stronger stage adaptability.

[0204] From the perspective of error metrics, after introducing stage segmentation, the RMSE decreased from 0.158 to 0.084, and the MAE decreased from 0.129 to 0.0612, significantly outperforming the un-staged model. This indicates that the signal distribution and characteristic changes differ significantly across different degradation stages, making it difficult for a uniform model to adapt to multi-stage characteristics and prone to prediction bias. However, through staged modeling, the model can dynamically adjust parameters according to the characteristics of each stage, improving its ability to model abrupt changes in degradation patterns, thereby significantly improving prediction accuracy and stability.

[0205] (2) Ablation Experiment Analysis

[0206] To verify the effectiveness of each component of the SAGAT model, a systematic ablation experiment was designed. The experiment evaluated the contribution of each part to the RUL prediction performance by progressively removing or replacing key components of the model. The models were defined as follows: SAGCN: No GAT attention mechanism was introduced; self-attention was used instead of multi-head attention; SAGAT-Attention: GAT was introduced, but channel fusion used self-attention instead of multi-head attention; GCN-MHA: GAT was used but without a dynamic graph structure, but a multi-head attention fusion mechanism was introduced; SAGAT: Both GAT and MHA were introduced.

[0207] The comparison results show that:

[0208] The proposed SAGAT model outperforms mainstream deep learning methods in prediction performance under various operating conditions. For example, the average RMSE of SAGAT reaches 0.114, significantly better than SAGCN (0.146), CNN-LSTM (0.161), CNN (0.221), DANN (0.216), and traditional LSTM (0.327). Specifically, for each individual bearing tested, SAGAT achieved the lowest error in 8 out of 11 samples, demonstrating good stability and adaptability.

[0209] SAGAT also demonstrates the ability to capture weak early features, effectively alleviating the problem of inaccurate predictions in the early stages of traditional methods. Thanks to the weighted aggregation of key neighbors and the multi-channel information fusion strategy of the graph attention mechanism, SAGAT can extract degenerate collaborative information between channels while preserving local sequence features, thereby improving the overall prediction accuracy.

[0210] (4) Generalization performance under different working conditions

[0211] To verify the generalization ability of the model, experiments were conducted under three different operating conditions. The SAGAT model was able to accurately fit the actual degradation trend under all three operating conditions (C1, C2, and C3).

[0212] Case Study 2: RUL Prediction on the IMS Dataset

[0213] Dataset Description

[0214] The IMS dataset is one of the important benchmark datasets for RUL prediction research on bearing-related equipment. The test platform for this dataset is equipped with four Rexnord ZA-2115 double-row bearings, rotating at a speed of 2000 r / min, bearing a load of 6000 lbs, and sampling at a frequency of 20 kHz.

[0215] The experiment used Case 1 from the IMS dataset. Case 1 contained four bearings, each with two channels. At the end of the test-failure experiment, bearing 3 showed an inner ring defect, and bearing 4 showed a roller element defect. Since the signals collected for each bearing represented its entire lifespan, data was collected every 5 minutes, with 20,480 data points collected each time. The graph structure was built using the data from each collection, and to prevent the model from becoming overly complex, the maximum number of nodes in each graph was set to 50. The dataset collected acceleration signals at a total of 2156 time points, thus dividing the entire lifespan into 2156 RUL prediction points.

[0216] Experimental Analysis

[0217] In this experiment, the IMS dataset Case 1 described in section 4.2.1 was used. Bearing 3 was used as the training set, and bearing 4 as the test set to verify the cross-bearing prediction capability of the proposed SAGAT model. The experimental procedure was consistent with the XJTU-SY dataset case: first, a sliding window (window length 256, step size 128) was used to extract the temporal and frequency domain features at each sampling time, constructing the corresponding dual-channel graph structure; then, feature modeling was performed in the horizontal and vertical directions respectively, and a multi-head attention mechanism was introduced in the fusion stage, finally outputting the RUL prediction value at that time. During training, the maximum number of nodes was limited to 50 to avoid excessive graph structure complexity.

[0218] To compare and verify the model performance, four model variants (SAGCN-MHA, SAGAT-Attention, GAT-MHA, and SAGAT) were set up, and the RMSE and MAE of the prediction results on test bearing 4 were calculated. The results are the average of 5 experiments with different random seeds. The experimental results show that the SAGAT model still maintains the best performance in this cross-bearing prediction task, and its RMSE and MAE are better than the other three variants, indicating that dynamic graph structure modeling and multi-head attention fusion have significant advantages in improving the cross-bearing generalization ability.

[0219] To further validate the advantages of the proposed SAGAT model in the cross-bearing RUL prediction task, this section compares its performance with several commonly used baseline methods, including the traditional temporal modeling method LSTM, convolutional neural networks (CNN), a hybrid convolutional and recurrent structure CNN-LSTM, and graph-based modeling SAGCN. All methods were trained and tested on the same IMS dataset Case 1 (bearing 3 as the training set and bearing 4 as the test set), and the input features and data preprocessing procedures were consistent with SAGAT to ensure fairness in the comparison. SAGAT demonstrates excellent generalization ability and stability in the cross-bearing prediction scenario, accurately predicting remaining life even when the bearing failure modes differ between training and testing (inner ring defect for bearing 3, roller element defect for bearing 4), demonstrating good cross-fault type transfer capability.

[0220] This invention addresses the problems of insufficient channel fusion, feature redundancy, and lack of stage awareness in traditional deep learning methods for bearing RUL prediction. It proposes an adaptive graph convolutional network (SAGAT) that integrates a dynamic graph attention mechanism and a dual-channel structure. This method constructs an initial graph structure through a sliding window, introduces dynamic adjacency updates and a graph attention mechanism to achieve adaptive weighted aggregation between neighboring nodes; the dual-channel graph structure and multi-head attention mechanism further enhance the multi-source feature fusion capability; combined with the automatic degradation stage segmentation strategy of the GMM, the model can dynamically adjust its modeling strategy according to the feature differences at different degradation stages.

[0221] Experimental results on the XJTU-SY and IMS public bearing life-cycle datasets demonstrate that SAGAT outperforms mainstream comparative methods under various operating conditions and cross-bearing prediction scenarios, achieving the best performance in metrics such as RMSE and MAE, and exhibiting higher prediction sensitivity and stability at abrupt changes in the degradation stage. Particularly in the IMS dataset's cross-failure-type prediction, the model maintains high accuracy across bearings with different failure modes, validating the method's generalization ability and practical application potential.

[0222] The above description is merely of preferred embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention, and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of bearings based on a stage-aware dynamic graph attention network, characterized in that, include: Step 1: Real-time acquisition of vibration signals during the operation of the bearing under test, wherein the real-time acquisition of vibration signals during the operation of the bearing under test is a time-series signal; The vibration signals collected during the operation of the bearing under test include: horizontal vibration signals and vertical vibration signals during the operation of the bearing under test. According to the preset time length, the horizontal vibration signal of the bearing under test during its operation is divided into T time periods of horizontal vibration signal under test. According to the preset time length, the vertical vibration signal of the bearing under test during its operation is divided into T time periods of vertical vibration signal under test. Step 2: Based on the vibration signals in the horizontal direction of the bearing to be tested over T time periods, extract B1 sensitive feature vectors in the horizontal direction of the bearing to be tested. Based on the vibration signals in the vertical direction to be detected over T time periods, B2 sensitive feature vectors in the vertical direction of the bearing to be detected are extracted. Based on the B1 sensitive feature vectors in the horizontal direction of the bearing to be tested, construct T initial dynamic graph structure data in the horizontal direction of the bearing to be tested. Based on the B2 sensitive feature vectors in the vertical directions of the bearing to be tested, construct T initial dynamic graph structure data in the vertical directions of the bearing to be tested. The one-dimensional health index value of the bearing under test is obtained based on the vibration signals of the B1 horizontal sensitive feature vectors and the B2 vertical sensitive feature vectors of the bearing under test. The stage label of the bearing to be tested can be obtained based on the one-dimensional health index value of the bearing to be tested; Step 3: Input the initial horizontal dynamic graph structure data of T bearings to be tested, the initial vertical dynamic graph structure data of T bearings to be tested, and the stage labels of the bearings to be tested into the pre-trained bearing life prediction model based on stage-aware dynamic graph attention network for processing, and obtain the predicted value of the remaining life of the bearings to be tested.

2. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 1, characterized in that, The training process of the bearing life prediction model based on a stage-aware dynamic graph attention network pre-trained in step three is as follows: S1. Collect horizontal vibration signal data of the bearing throughout its entire life cycle, vertical vibration signal data of the bearing throughout its entire life cycle, and RUL life tag data. A training set is constructed based on the horizontal vibration signal data of the bearing throughout its entire life cycle, the vertical vibration signal data of the bearing throughout its entire life cycle, and the RUL life tag data. S2. Construct a bearing life prediction model based on a stage-aware dynamic graph attention network; including: a first SAGAT network module, a second SAGAT network module, a third SAGAT network module, a weighted fusion module based on a multi-head attention mechanism, and a regression prediction layer; S3. Train the bearing life prediction model based on the stage-aware dynamic graph attention network according to the training set to obtain the trained bearing life prediction model based on the stage-aware dynamic graph attention network.

3. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 2, characterized in that, In step S1, a training set is constructed based on the horizontal vibration signal data and the vertical vibration signal data of the bearing throughout its entire lifespan; the specific process is as follows: S1.1: Perform data preprocessing on the horizontal vibration signal data of the bearing throughout its entire life cycle to obtain preprocessed horizontal vibration signal data; The preprocessed horizontal vibration signal data includes: preprocessed horizontal vibration signal data for T time periods; The vertical vibration signal data of the bearing throughout its entire life cycle is preprocessed to obtain the preprocessed vertical vibration signal data. The preprocessed vertical vibration signal data includes: preprocessed vertical vibration signal data for T time periods; S1.2: Extract A1 horizontal time-domain features and A2 horizontal frequency-domain features from the preprocessed horizontal vibration signal data; Extract A3 vertical time-domain features and A4 vertical frequency-domain features from the preprocessed vertical vibration signal data; A1, A2, A3, and A4 are all positive integers; S1.3: Perform correlation analysis on the time-domain features of A1 horizontal directions and the frequency-domain features of A2 horizontal directions to screen out the sensitive features of B1 horizontal directions; Correlation analysis was performed on the time-domain features of A3 vertical directions and the frequency-domain features of A4 vertical directions to obtain B2 sensitive features in the vertical directions; B1 and B2 are both positive integers. S1.4: Kernel principal component analysis (KPC) is used to reduce the dimensionality of the B1 horizontal and B2 vertical sensitive features, resulting in a one-dimensional health indicator sequence HI; expressed by the formula: This represents the health indicator values ​​for the first time period. This represents the health indicator values ​​for the second period. This represents the health indicator value for the T-th time period. S1.5: Perform clustering processing on the one-dimensional health indicator sequence HI to obtain the clustering results of the one-dimensional health indicator sequence HI. Stage label data is constructed based on the clustering results of the one-dimensional health indicator sequence HI; S1.6: Based on the preprocessed horizontal vibration signal data and B1 sensitive features in the horizontal direction obtained from S1.1, construct a sequence of time-series feature maps in the horizontal direction; The sequence of temporal feature maps in the horizontal direction includes: T temporal feature maps in the horizontal direction; Specifically, based on the preprocessed horizontal vibration signal data of the t-th time period and the B1 sensitive features of the horizontal direction, a time series feature map of the t-th horizontal direction is constructed. Based on the preprocessed vertical vibration signal data obtained from S1.1 and the B2 sensitive features in the vertical direction, a sequence of time-series feature maps in the vertical direction is constructed. The sequence of temporal feature maps in the vertical direction includes: T temporal feature maps in the vertical direction; Specifically, based on the preprocessed vertical vibration signal data of the t-th time period and the B2 vertical sensitive features, a time-series feature map of the t-th vertical direction is constructed. S1.7: Construct a training set based on the temporal feature map sequence in the horizontal direction, the temporal feature map sequence in the vertical direction, the stage label data, and the RUL lifetime label.

4. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 3, characterized in that, In step S1.6, a time-series feature map of the t-th horizontal direction is constructed based on the preprocessed horizontal vibration signal data of the t-th time period and B1 sensitive features of the horizontal direction; the specific process is as follows: S1.6.1.1: Use a sliding window to segment the preprocessed horizontal vibration signal data in the t-th time period to obtain S time slices of horizontal vibration signal data; S1.6.1.2: Define the S time slices as the S nodes of the time series feature map in the t-th horizontal direction; S1.6.1.3: Construct the feature vectors of the S nodes of the temporal feature map in the t-th horizontal direction based on the sensitive features in the B1 horizontal directions; Construct the node feature matrix of the time series feature map in the t-th horizontal direction based on the feature vectors of the S nodes of the time series feature map in the t-th horizontal direction; Where d is the feature dimension; S1.6.1.4: Based on the feature vectors of the S nodes of the time series feature map in the t-th horizontal direction, calculate the cosine similarity between the feature vectors of any two nodes in the time series feature map in the t-th horizontal direction. Based on the cosine similarity between the feature vectors of any two nodes in the temporal feature map of the t-th horizontal direction, the edges of the graph are constructed using the K-Nearest Neighbor (KNN) strategy to obtain the temporal feature map of the t-th horizontal direction.

5. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 4, characterized in that, In step S1.6, based on the preprocessed vertical vibration signal data of the t-th time period and the B2 vertical sensitive features, a time-series feature map of the t-th vertical direction is constructed; the specific process is as follows: S1.6.2.1: Use a sliding window to segment the preprocessed vertical vibration signal data of the t-th time period to obtain S time slices of vertical vibration signal data; S1.6.2.2: Define the S time slices as the S nodes of the time series feature map in the t-th vertical direction; S1.6.2.3: Construct the feature vectors of the S nodes of the temporal feature map in the t-th vertical direction based on the sensitive features in the B1 vertical directions; Construct the node feature matrix of the temporal feature map in the t-th vertical direction based on the feature vectors of the S nodes in the t-th vertical direction; where d is the feature dimension. S1.6.2.4: Based on the feature vectors of the S nodes of the time series feature map in the t-th vertical direction, calculate the cosine similarity between the feature vectors of any two nodes in the time series feature map in the t-th vertical direction. Based on the cosine similarity between the feature vectors of any two nodes in the temporal feature map of the t-th vertical direction, the edges of the graph are constructed using the K-Nearest Neighbor (KNN) strategy to obtain the temporal feature map of the t-th vertical direction.

6. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 5, characterized in that, In step S3, the bearing life prediction model based on the stage-aware dynamic graph attention network is trained using the training set to obtain the trained bearing life prediction model based on the stage-aware dynamic graph attention network. The specific process is as follows: S3.1: Input the temporal feature map of the horizontal direction in the training set into the corresponding SAGAT network module according to the stage label data for feature extraction processing to obtain the deep degradation feature map of the horizontal channel; The temporal feature maps in the vertical direction of the training set are input into the corresponding SAGAT network module for feature extraction based on the stage label data, so as to obtain the deep degradation feature maps of the vertical channel. S3.2: The deep degradation feature maps of the horizontal channel and the deep degradation feature maps of the vertical channel are concatenated to obtain a fused feature matrix; The fused feature matrix is ​​input into a weighted fusion module based on a multi-head attention mechanism for weighted fusion processing to obtain the fused global spatiotemporal feature map; S3.3: Input the fused global spatiotemporal feature map into the regression prediction layer and output the predicted value of the remaining service life of the bearing; S3.4: Calculate the loss function based on the predicted remaining service life of the bearing and the RUL life labels in the training set. Train the bearing service life prediction model based on the stage-aware dynamic graph attention network according to the loss function. Stop training when the loss function converges to obtain the trained bearing service life prediction model based on the stage-aware dynamic graph attention network.

7. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 6, wherein in step S3.1, the temporal feature map of the horizontal direction in the training set is input into the SAGAT network module for feature extraction processing to obtain the deep degradation feature map of the horizontal channel; the specific process is expressed by the formula as follows: ； In the formula, This represents the node feature matrix of the time-series feature map in the t-th horizontal direction; This represents the attention weighting coefficient of the l-th layer. This represents the attention weighting coefficient of the second layer. This represents the attention weighting coefficient for the third layer. This represents the activation function. This represents the intermediate feature map at the first level. This represents the intermediate feature map at the second level. This represents a deep degradation feature map of the horizontal channel; Denotes the parameter matrix of the zeroth filter. This represents the parameter matrix of the first filter. This represents the parameter matrix of the second filter; Indicates the zeroth bias term; Indicates the first bias term. Indicates the second bias term. Denotes the first adaptive projection matrix. Represents the second adaptive projection matrix The process involves inputting the temporal feature map of the vertical direction in the training set into the SAGAT network module for feature extraction to obtain the deep degradation feature map of the vertical channel. The specific process is expressed by the following formula: ； In the formula, Let represent the node feature matrix of the time series feature map in the t-th vertical direction. This represents the intermediate feature map at the first level. This represents the intermediate feature map at the second level. This represents a deep degradation feature map of the horizontal channel.

8. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 7, characterized in that, The weighted fusion module based on the multi-head attention mechanism in S3.2 includes H attention heads; The fused feature matrix is ​​input into a weighted fusion module based on a multi-head attention mechanism for weighted fusion processing to obtain a fused global spatiotemporal feature map; the specific process is as follows: S3.2.1: Input the fused feature map into H attention heads for processing to obtain the intermediate attention output map of H attention heads; Among them, the feature maps will be fused. The h-th attention head is processed to obtain the intermediate attention output graph of the h-th attention head. ; S3.2.2: The intermediate attention output maps of the H attention heads are concatenated to obtain the concatenated features; expressed by the formula: ； In the formula, Indicates the features after splicing. This represents the intermediate attention output diagram of the first attention head; This represents the intermediate attention output diagram of the first attention head; This represents the intermediate attention output graph of the Hth attention head; S3.2.3: The concatenated features in S3.2.2 are processed by a linear transformation matrix to obtain the global spatiotemporal feature map.

9. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 8, characterized in that, In S3.2.1, the feature maps will be fused. The h-th attention head is processed to obtain the intermediate attention output graph of the h-th attention head. The formula is as follows: ； ； ； In the formula, This represents the query projection matrix of the h-th attention head. This represents the key projection matrix of the h-th attention head. The projection matrix represents the value of the h-th attention head. This represents the query vector for the h-th attention head. This represents the query vector for the h-th attention head. This represents the query vector for the h-th attention head. This represents the attention distribution matrix of the h-th attention head. Indicated Transpose This represents the feature dimension of the h-th attention head.

10. The bearing remaining service life prediction method based on a stage-aware dynamic graph attention network according to claim 9, characterized in that, The loss function in S3.4 is the mean square error function.

Images