Bearing life prediction method and system based on heterogeneous graph and spatio-temporal feature fusion

Abstract

The application provides a bearing life prediction method and system based on a heterogeneous graph and space-time feature fusion, which comprises the following steps: collecting vibration signals of a rolling bearing in a whole life cycle and extracting time domain features and frequency domain features as original inputs; constructing a heterogeneous graph data structure composed of multiple node types and edge connection types on a heterogeneous graph convolution branch, using different types of graph convolution networks to extract the spatial correlation of degradation features; introducing multiple residual blocks with different expansion coefficients in the time convolution branch to learn the time sequence features in the degradation process layer by layer; enhancing the expression ability of the features through a node attention mechanism and a self-attention mechanism, dynamically adjusting the feature weights by using a gating mechanism to realize space-time feature fusion, and obtaining a prediction result of the remaining service life of the rolling bearing through a full connection layer. Thanks to the introduction of the heterogeneous graph data structure and the space-time parallel model framework, the prediction model proposed by the application has great improvement in prediction accuracy and stability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a bearing life prediction method and system based on the fusion of heterogeneous graphs and spatiotemporal features. Background Technology

[0002] Rolling bearings, as core support and transmission components of rotating machinery, are widely used in important fields such as machinery manufacturing, rail transportation, and aerospace. Their operating status directly determines the reliability and efficiency of the entire equipment system. Under long-term operation, rolling bearings inevitably deteriorate due to continuously changing external loads, lubrication conditions, and other environmental factors. The goal of rolling bearing remaining service (RUL) prediction is to analyze the degradation patterns of rolling bearings and estimate the time from the current state to complete failure based on the state data detected during operation. Therefore, achieving accurate rolling bearing RUL prediction is crucial for predictive maintenance, reducing downtime, and optimizing maintenance decisions, and has significant engineering and economic value for ensuring safe equipment operation and reducing maintenance costs. Existing rolling bearing RUL prediction methods can be divided into four categories: probabilistic statistics, physical models, data-driven methods, and combined models. With the rapid development of sensor technology and the significant increase in monitoring data, in-depth research on data-driven RUL prediction methods has been directly driven. Deep learning methods, represented by neural networks, have received widespread attention in the field of RUL prediction due to their flexibility and versatility. This type of method achieves end-to-end direct prediction by automatically extracting deep features from the original data, demonstrating high prediction accuracy and nonlinear modeling capabilities.

[0003] Deep learning-based rolling bearing RUL prediction methods have branched into two distinct research directions based on different data construction methods: temporal models represented by Long Short-Term Memory (LSTM) networks and graph models represented by Graph Convolutional Networks (GCN). Temporal models utilize techniques such as sliding windows to read adjacent sequences and deeply mine temporal correlations within the time series to achieve RUL prediction, representing a typical Euclidean data prediction method. Graph models, on the other hand, read the node feature matrices and adjacency matrices in the graph data structure to learn spatial correlations and complete RUL prediction, establishing a prediction process for non-Euclidean data. Furthermore, concatenated models such as GCN-LSTM enable the joint use of temporal and graph data structures, integrating degradation features from both temporal and spatial dimensions through a feature fusion module.

[0004] However, graph-based rolling bearing RUL prediction techniques still have shortcomings, limiting their application in practical engineering. First, while sequential spatiotemporal prediction models are easy to build and understand in their overall framework, their separate learning of temporal and spatial degradation features causes them to ignore the correlation between these features. Furthermore, the order of the temporal and graph models can also influence the bias of the prediction results. Second, most graph-based prediction methods construct homogeneous graphs with single relational attributes. The fixed node and edge connection types in homogeneous graphs limit the model's ability to capture complex, multi-layered relationships in the data and its potential to utilize embedded associations and differences within nodes. Therefore, there is room for further improvement in the information utilization of graph data structures. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a bearing life prediction method and system based on heterogeneous graph and spatiotemporal feature fusion, so as to at least solve the shortcomings of the above-mentioned technology.

[0006] This invention proposes a bearing life prediction method based on the fusion of heterogeneous graphs and spatiotemporal features, comprising: The vibration signals of the rolling bearing in the horizontal and vertical directions are acquired. Each vibration signal sample is used as the smallest feature extraction unit. Its time domain features and frequency domain features are calculated respectively, and the features in the two directions are combined to form a complete feature set. Data slices are performed on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and the continuous feature sequence is then subjected to standardized data processing. Two node types are defined based on the time-domain features and the frequency-domain features. Two edge connection types are defined based on the similarity of features and the matching relationship between the same feature in different directions. The continuous feature sequence is transformed into a heterogeneous graph data structure. The heterogeneous graph data structure is input into the heterogeneous graph convolution branch. Node features are updated based on similarity edges and matching edges to obtain spatial features. Spatial feature enhancement is achieved through a node attention mechanism to obtain spatially enhanced features. The continuous feature sequence is input into the temporal convolution branch, and temporal features are learned using multiple temporal convolution residual blocks with different dilation coefficients. Temporal feature enhancement is achieved through a self-attention mechanism to obtain temporally enhanced features. The spatial enhancement features and the temporal enhancement features are input into the spatiotemporal feature fusion model. The weights of the two types of features are dynamically adjusted through a gating mechanism to calculate the spatiotemporal fusion features. The spatiotemporal fusion features are then input into the fully connected prediction layer to obtain the prediction result of the remaining service life of the rolling bearing.

[0007] Furthermore, the time-domain features include maximum value, minimum value, average value, variance, median, range, peak value, absolute mean, root mean square, standard deviation, kurtosis, skewness, waveform factor, peak factor, impulse factor, and margin factor; the frequency-domain features include total spectral energy, median spectral amplitude, maximum spectral amplitude, minimum spectral amplitude, mean spectral amplitude, spectral frequency centroid, spectral frequency standard deviation, and spectral frequency root mean square.

[0008] Furthermore, the steps of slicing the data on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and then standardizing the continuous feature sequence include: The sliding window algorithm is used to slice the data on the complete feature set according to the preset sliding window length and the sliding step size is 1, so as to obtain the feature vector of each vibration signal acquisition sample in the continuous feature sequence corresponding to each time step in the continuous feature sequence in two acquisition directions. The feature dimension of the feature vector is equal to the sum of the types of extracted features. The continuous feature sequence increases the number of training samples while retaining the performance degradation information between adjacent sampling samples. The maximum and minimum values ​​of the extracted features from the training data are used as the scale for feature standardization. The data input to the prediction model are then subjected to min-max standardization according to the following expression.

[0009] Furthermore, the continuous feature sequence is transformed into a heterogeneous graph data structure, and the heterogeneous graph data structure is input into the heterogeneous graph convolution branch. Node features are updated based on similarity edges and matching edges to obtain spatial features. Spatial feature enhancement is achieved through a node attention mechanism to obtain spatially enhanced features. The steps include: Construct the heterogeneous graph data structure ,in, Represents a set of nodes. Denotes the set of edges. Represents a set of node types. Represents a set of edge connection types. express Node type mapping, express Edge connection type mapping; The nodes obtained from time-domain features and frequency-domain features are defined as time-domain feature nodes and frequency-domain feature nodes. By calculating the absolute value of the Pearson correlation coefficient between different feature nodes, the most relevant nodes are selected to establish the similarity edge connection. The matching edge connection is established by searching for feature nodes with the same feature in the horizontal and vertical directions. In the heterogeneous graph convolution branch, the node update process of a single-layer heterogeneous graph convolutional network is as follows:

[0010] In the formula, This represents the input features of a single-layer heterogeneous graph convolutional network. The initial input of the first layer of the heterogeneous graph convolutional network is a heterogeneous graph. The node feature matrix, This represents the spatial features of the output of a single-layer heterogeneous graph convolutional network; By using a node attention mechanism and linear transformation, the spatial features obtained from the convolutional branch of the heterogeneous graph are transformed into a key matrix and a value matrix. The temporal form of the node feature matrix of the heterogeneous graph is linearly transformed into a query matrix. By traversing all feature nodes in the heterogeneous graph through the query matrix at each prediction time step, spatial enhancement features based on the node attention mechanism are obtained.

[0011] Furthermore, the steps of inputting the continuous feature sequence into the temporal convolution branch, learning temporal features using multiple temporal convolution residual blocks with different dilation coefficients, and enhancing temporal features through a self-attention mechanism to obtain enhanced temporal features include: In the temporal convolution branch, the constructed temporal convolutional network contains three layers of temporal convolutional residual blocks with different dilation coefficients. Each temporal convolutional residual block consists of a dilated causal convolutional layer, a normalization layer, a nonlinear activation layer, and a residual connection layer. The temporal features obtained from the temporal convolution branch are transformed into a query matrix, a key matrix, and a value matrix using a linear transformation through a self-attention mechanism. By calculating the correlation weights between different prediction time steps, the temporal features of each prediction time step are weighted and fused to obtain the temporal enhancement features.

[0012] Furthermore, the step of inputting the spatiotemporal fusion features into the fully connected prediction layer to obtain the remaining service life prediction result of the rolling bearing includes: The spatial enhancement features and the temporal enhancement features are input into the spatiotemporal feature fusion model, and the spatial enhancement features and temporal enhancement features are concatenated in the feature dimension to obtain the spatiotemporal joint features; The gating mechanism calculates the gating weights used to measure the correlation between the two types of features and the performance degradation state. The two types of features are then weighted and fused according to the gating weights to obtain the spatiotemporal fusion features. After the spatiotemporal fusion features are input into the fully connected prediction layer, they are mapped to the [0,1] interval through the fully connected layer and the Sigmoid activation function to obtain the prediction result of the remaining service life of the rolling bearing.

[0013] This invention also proposes a bearing life prediction system based on the fusion of heterogeneous graphs and spatiotemporal features, comprising: The signal acquisition module is used to acquire the vibration signals of the rolling bearing in the horizontal and vertical directions. It takes a single vibration signal sample as the smallest feature extraction unit, calculates its time domain features and frequency domain features respectively, and combines the features in the two directions to form a complete feature set. The data processing module is used to slice the data on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and to perform standardized data processing on the continuous feature sequence. The spatial feature enhancement module is used to define two node types based on the time-domain features and the frequency-domain features, and two edge connection types based on the similarity of features and the matching relationship between the same feature in different directions. It transforms the continuous feature sequence into a heterogeneous graph data structure, inputs the heterogeneous graph data structure into a heterogeneous graph convolution branch, updates the node features based on similarity edges and matching edges to obtain spatial features, and enhances the spatial features through a node attention mechanism to obtain spatially enhanced features. The temporal feature enhancement module is used to input the continuous feature sequence into the temporal convolution branch, learn the temporal features using multiple temporal convolution residual blocks with different dilation coefficients, and enhance the temporal features through a self-attention mechanism to obtain the temporal enhanced features. The life prediction module is used to input the spatial enhancement features and the temporal enhancement features into the spatiotemporal feature fusion model, dynamically adjust the weights of the two types of features through a gating mechanism, calculate the spatiotemporal fusion features, and input the spatiotemporal fusion features into the fully connected prediction layer to obtain the remaining life prediction result of the rolling bearing.

[0014] Furthermore, the data processing module is specifically used for: The sliding window algorithm is used to slice the data on the complete feature set according to the preset sliding window length and the sliding step size is 1, so as to obtain the feature vector of each vibration signal acquisition sample in the continuous feature sequence corresponding to each time step in the continuous feature sequence in two acquisition directions. The feature dimension of the feature vector is equal to the sum of the types of extracted features. The continuous feature sequence increases the number of training samples while retaining the performance degradation information between adjacent sampling samples. The maximum and minimum values ​​of the extracted features from the training data are used as the scale for feature standardization. The data input to the prediction model are then subjected to min-max standardization according to the following expression.

[0015] Furthermore, the spatial feature enhancement module is specifically used for: Construct the heterogeneous graph data structure ,in, Represents a set of nodes. Denotes the set of edges. Represents a set of node types. Represents a set of edge connection types. express Node type mapping, express Edge connection type mapping; The nodes obtained from time-domain features and frequency-domain features are defined as time-domain feature nodes and frequency-domain feature nodes. By calculating the absolute value of the Pearson correlation coefficient between different feature nodes, the most relevant nodes are selected to establish the similarity edge connection. The matching edge connection is established by searching for feature nodes with the same feature in the horizontal and vertical directions. In the heterogeneous graph convolution branch, the node update process of a single-layer heterogeneous graph convolutional network is as follows:

[0016] In the formula, This represents the input features of a single-layer heterogeneous graph convolutional network. The initial input of the first layer of the heterogeneous graph convolutional network is a heterogeneous graph. The node feature matrix, This represents the spatial features of the output of a single-layer heterogeneous graph convolutional network; By using a node attention mechanism and linear transformation, the spatial features obtained from the convolutional branch of the heterogeneous graph are transformed into a key matrix and a value matrix. The temporal form of the node feature matrix of the heterogeneous graph is linearly transformed into a query matrix. By traversing all feature nodes in the heterogeneous graph through the query matrix at each prediction time step, spatial enhancement features based on the node attention mechanism are obtained.

[0017] Furthermore, the temporal feature enhancement module is specifically used for: In the temporal convolution branch, the constructed temporal convolutional network contains three layers of temporal convolutional residual blocks with different dilation coefficients. Each temporal convolutional residual block consists of a dilated causal convolutional layer, a normalization layer, a nonlinear activation layer, and a residual connection layer. The temporal features obtained from the temporal convolution branch are transformed into a query matrix, a key matrix, and a value matrix using a linear transformation through a self-attention mechanism. By calculating the correlation weights between different prediction time steps, the temporal features of each prediction time step are weighted and fused to obtain the temporal enhancement features.

[0018] Furthermore, the lifetime prediction module is specifically used for: The spatial enhancement features and the temporal enhancement features are input into the spatiotemporal feature fusion model, and the spatial enhancement features and temporal enhancement features are concatenated in the feature dimension to obtain the spatiotemporal joint features; The gating mechanism calculates the gating weights used to measure the correlation between the two types of features and the performance degradation state. The two types of features are then weighted and fused according to the gating weights to obtain the spatiotemporal fusion features. After the spatiotemporal fusion features are input into the fully connected prediction layer, they are mapped to the [0,1] interval through the fully connected layer and the Sigmoid activation function to obtain the prediction result of the remaining service life of the rolling bearing.

[0019] The present invention also proposes a storage medium on which a computer program is stored, which, when executed by a processor, implements the above-described bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion.

[0020] The present invention also proposes a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion.

[0021] Compared with existing technologies, the bearing life prediction method and system based on heterogeneous graphs and spatiotemporal feature fusion in this invention constructs a spatiotemporally parallel prediction model. It extracts spatial and temporal features of the bearing performance degradation process from the heterogeneous graph convolutional branch and the temporal convolutional branch, respectively. A gating mechanism is used to achieve spatiotemporal feature fusion, outputting the model's RUL prediction result. The proposed heterogeneous graph data structure, by defining multiple node types and edge connection types, improves the utilization of degradation information in vibration signals by the graph data, enhancing the prediction model's ability to represent spatially correlated features. Compared with homogeneous graph data structures, the heterogeneous graph data structure and its corresponding heterogeneous graph convolutional network avoid homogeneous processing of heterogeneous features by the prediction model, enhancing the targeting of feature extraction. Simultaneously, node attention and self-attention mechanisms are introduced to enhance the features of key nodes and key time steps in the prediction data. The gating mechanism adaptively adjusts the contribution weights of spatiotemporal features, making the spatiotemporally fused features better reflect the performance degradation law of rolling bearings, improving the model's prediction accuracy and stability. Attached Figure Description

[0022] Figure 1 This is a flowchart of the bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the prediction model in the first embodiment of the present invention; Figure 3 This is a comparison chart of the prediction performance of Bearing1-3 in different models in the first embodiment of the present invention; Figure 4 This is a comparison chart of the prediction performance of Bearing1-6 in different models in the first embodiment of the present invention; Figure 5 This is a comparison chart of the prediction performance of Bearing2-4 in different models in the first embodiment of the present invention; Figure 6 This is a comparison chart of the prediction performance of Bearing3-3 in different models in the first embodiment of the present invention; Figure 7 This is a structural block diagram of the bearing life prediction system based on heterogeneous graph and spatiotemporal feature fusion in the second embodiment of the present invention; Figure 8 This is a structural block diagram of the computer in the third embodiment of the present invention.

[0023] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0024] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] Example 1 Please see Figure 1 The image shows a bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion in the first embodiment of the present invention. The method specifically includes steps S101 to S105: S101, acquire the vibration signals of the rolling bearing in the horizontal and vertical directions, take a single vibration signal sample as the smallest feature extraction unit, calculate its time domain features and frequency domain features respectively, and combine the features in the two directions to form a complete feature set; In practical implementation, a feature set is obtained by extracting the time-domain and frequency-domain features of the horizontal and vertical vibration signals of the rolling bearing. ,in, Indicates the number of samples collected. The feature set represents the number of feature classes. The time-domain and frequency-domain features in the feature set deeply characterize the degradation status of the rolling bearing during operation. Some of the time-domain features show obvious trends in the degradation process, while some frequency-domain features directly reflect the fault characteristics. The combined use of horizontal and vertical vibration signals improves the information integrity of the subsequently constructed heterogeneous graph data structure and time series. Furthermore, the time-domain features include maximum values. Minimum value ,average value ,variance , median Range Peak Absolute mean Root mean square Standard deviation , cliff skewness Waveform factor Peak factor Pulse factor and leeway factor ,in, This represents a single vibration signal sample. This represents the first sample in a single vibration signal sampling sample. The amplitude of each vibration signal; Furthermore, the frequency domain features include the total spectral energy. Spectral amplitude median Maximum spectral amplitude Minimum spectral amplitude Spectral amplitude mean Spectral frequency centroid Spectral frequency standard deviation Root mean square of spectral frequency ,in, This represents the spectrum of a single vibration signal sample, with the number of spectral lines being... , For the first The frequency value corresponding to each spectral line.

[0027] S102, Data slices are performed on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and the continuous feature sequence is subjected to standardized data processing. Furthermore, step 102 specifically includes steps S1021 to S1022: S1021, using the sliding window algorithm, data is sliced ​​on the complete feature set according to a preset sliding window length and a sliding step size of 1, to obtain the feature vector of a vibration signal acquisition sample in two acquisition directions corresponding to each time step in the continuous feature sequence. The feature dimension of the feature vector is equal to the sum of the types of extracted features. The continuous feature sequence increases the number of training samples while retaining the performance degradation information between adjacent sampling samples. S1022, Read the maximum and minimum values ​​of the extracted features from the training data as the scale for feature standardization, and perform min-max standardization on the data input to the prediction model according to the following expression.

[0028] In practical implementation, the complete feature set obtained from the above steps will be... Data is divided according to a fixed sliding window length and step size to obtain a continuous feature sequence of the same length as the sliding window. ,in, Indicates the number of feature sequences. This indicates the length of the sliding window; each time step of the feature sequence corresponds to the feature vector of the vibration signal acquisition sample in two acquisition directions, and its feature dimension is equal to the number of feature classes extracted; Obtain feature sequences Then read the maximum value of the training data. and minimum value As a measure of feature standardization, min-max standardization is performed on the data input to the model to reduce the impact of outliers on the prediction model and improve the convergence speed of model training.

[0029] S103, define two node types based on the time domain features and the frequency domain features, define two edge connection types based on the similarity of features and the matching relationship between the same feature in different directions, transform the continuous feature sequence into a heterogeneous graph data structure, and input the heterogeneous graph data structure into a heterogeneous graph convolution branch, update the node features based on similarity edges and matching edges respectively to obtain spatial features, and enhance the spatial features through a node attention mechanism to obtain spatially enhanced features; Furthermore, step S103 specifically includes steps S1031 to S1034: S1031, Construct the heterogeneous graph data structure ,in, Represents a set of nodes. Denotes the set of edges. Represents a set of node types. Represents a set of edge connection types. express Node type mapping, express Edge connection type mapping; S1032, the nodes obtained from time-domain features and frequency-domain features are defined as time-domain feature nodes and frequency-domain feature nodes. By calculating the absolute value of the Pearson correlation coefficient between different feature nodes, the most relevant nodes are selected to establish the similarity edge connection. The matching edge connection is established by searching for feature nodes with the same feature in the horizontal and vertical directions. S1033, in the heterogeneous graph convolution branch, the node update process of a single-layer heterogeneous graph convolutional network is as follows:

[0030] In the formula, This represents the input features of a single-layer heterogeneous graph convolutional network. The initial input of the first layer of the heterogeneous graph convolutional network is a heterogeneous graph. The node feature matrix, This represents the spatial features of the output of a single-layer heterogeneous graph convolutional network; S1034, using a node attention mechanism, the spatial features finally obtained from the convolutional branch of the heterogeneous graph are transformed into a key matrix and a value matrix through linear transformation. The temporal form of the node feature matrix of the heterogeneous graph is linearly transformed into a query matrix. By traversing all feature nodes in the heterogeneous graph through the query matrix at each prediction time step, spatial enhancement features based on the node attention mechanism are obtained.

[0031] In practical implementation, a heterogeneous graph data structure composed of heterogeneous nodes and heterogeneous edge connections is constructed. The heterogeneous graph convolutional branch input to the prediction model learns the spatial correlations of degenerate features. The node feature matrix of a single heterogeneous graph can be represented as... Heterogeneous graph data structure Spatial augmentation features are obtained through heterogeneous graph convolutional networks and node attention mechanisms. The network structure and data dimension variations of the heterogeneous graph convolutional branch are shown in Table 1. Table 1. Network structure and data dimension changes of heterogeneous graph convolution branches

[0032] In this embodiment, the calculation process of the node attention mechanism is as follows:

[0033]

[0034]

[0035]

[0036] in, The key matrix is ​​represented by the feature matrix of the heterogeneous graph nodes. The linear transformation yields the result, where, and These are the learned key weights and biases, respectively. The query matrix is ​​represented by the node feature matrix in the time-series form of the heterogeneous graph. The linear transformation is used to calculate the result, where and These are the learned query weights and biases, respectively.

[0037] S104, the continuous feature sequence is input into the temporal convolution branch, and temporal features are learned using multiple temporal convolution residual blocks with different dilation coefficients. Temporal feature enhancement is achieved through a self-attention mechanism to obtain temporal enhanced features. Furthermore, step S104 specifically includes steps S1041 to S1042: S1041, in the temporal convolution branch, the constructed temporal convolutional network includes three layers of temporal convolutional residual blocks with different dilation coefficients. Each temporal convolutional residual block consists of a dilated causal convolutional layer, a normalization layer, a nonlinear activation layer, and a residual connection layer. S1042, by using a self-attention mechanism and linear transformation, the temporal features obtained from the temporal convolution branch are transformed into a query matrix, a key matrix, and a value matrix. By calculating the correlation weights between different prediction time steps, the temporal features of each prediction time step are weighted and fused to obtain the temporal enhancement features.

[0038] In practical implementation, the above-mentioned standardized continuous feature sequences will be used as the basis for further processing. The input is fed into the temporal convolutional branch to learn temporal features. The temporal convolutional network contains three TCN residual blocks, each consisting of a dilated causal convolutional layer, a normalization layer, a nonlinear activation layer, and a residual connection layer. The standardized feature sequence is input into the temporal convolutional branch to learn temporal features. The main difference between the three TCN residual blocks lies in the different dilation coefficients in the dilated causal convolutional layer. By defining different expansion coefficients in the expanded causal convolutional layers of the TCN residual block, the model progressively expands its receptive field to extract degradation information at different time scales. After learning layer by layer through the temporal convolutional branch, the temporal characteristics of the rolling bearing performance degradation process are obtained. The network structure and data dimension changes of the temporal convolutional branch are shown in Table 2. Table 2. Network structure and data dimension changes of the temporal convolutional branch.

[0039] The feature update process of the temporal convolution branch is as follows:

[0040] The self-attention mechanism uses the temporal features obtained via the TCN branch as input data and obtains the query matrix through a linear transformation layer. Key matrix and value matrix The self-attention mechanism calculates the correlation weights between different prediction time steps and uses these weights to perform weighted fusion of the temporal features at each prediction time step, thus obtaining temporal enhancement features based on the self-attention mechanism. The computational process of the self-attention mechanism is as follows:

[0041] in, This represents the scaling factor, used to suppress potential numerical fluctuations during the self-attention mechanism computation process.

[0042] S105, the spatial enhancement features and the temporal enhancement features are input into the spatiotemporal feature fusion model. The weights of the two types of features are dynamically adjusted through a gating mechanism to calculate the spatiotemporal fusion features. The spatiotemporal fusion features are then input into the fully connected prediction layer to obtain the prediction result of the remaining service life of the rolling bearing.

[0043] Furthermore, step S105 specifically includes steps S1051 to S1053: S1051, The spatial enhancement features and the temporal enhancement features are input together into the spatiotemporal feature fusion model, and the spatial enhancement features and temporal enhancement features are concatenated in the feature dimension to obtain the spatiotemporal joint features; S1052, The gating weights used to measure the correlation between the two types of features and the performance degradation state are calculated through the gating mechanism. The two types of features are then weighted and fused according to the gating weights to obtain the spatiotemporal fusion features. S1053, after inputting the spatiotemporal fusion features into the fully connected prediction layer, the features are mapped to the [0,1] interval through the fully connected layer and the Sigmoid activation function to obtain the prediction result of the remaining service life of the rolling bearing.

[0044] In practical implementation, the spatial enhancement features obtained above will be used... and temporal enhancement features The input is fed into a spatiotemporal feature fusion model (or time-series feature fusion model) to achieve a joint representation of the spatial and temporal correlations in the performance degradation process of rolling bearings. The spatiotemporal feature fusion model concatenates spatial and temporal enhancement features along the feature dimension to obtain joint spatiotemporal features. Subsequently, two fully connected layers are used to obtain gating weights that characterize the contribution of spatial augmentation features and temporal augmentation features. Gating weights The spatiotemporal fusion feature is obtained by weighted fusion of spatial enhancement features and temporal enhancement features. The calculation process of the spatiotemporal feature fusion model based on the gating mechanism is as follows:

[0045]

[0046]

[0047] In the formula, , , , This represents the weights and biases of the fully connected layer in the spatiotemporal feature fusion model. This represents the Sigmoid activation function.

[0048] Spatiotemporal fusion features are transformed into RUL prediction results in the [0,1] interval through a fully connected prediction layer and a Sigmoid activation function. The calculation process is as follows:

[0049] The introduction of gating mechanism in the spatiotemporal feature fusion model enables the prediction model to dynamically adjust and reduces the dependence on a single feature branch; the changes in data dimension of the spatiotemporal feature fusion model and prediction layer are shown in Table 3. Table 3. Data Dimension Changes in the Spatiotemporal Feature Fusion Model and Prediction Layer

[0050] The proposed method will be further explained below using RUL prediction experiments on the PHM2012 dataset: The sampling frequency in the PHM2012 dataset is set to 25.6 kHz, with 2560 sampling points recorded for each sample within 0.1 seconds. The dataset provides rolling bearing vibration signals throughout their entire life cycle under three operating conditions, with a vibration signal amplitude of 20g set as the complete failure threshold for the bearing. For a comparison of the prediction performance of some rolling bearing samples from the PHM2012 dataset under different models, please refer to [link to relevant documentation]. Figures 3 to 6 The detailed information of the PHM2012 dataset and the experimental setup for RUL prediction are shown in Table 4: Table 4. Prediction Experiment Setup Based on PHM2012 Dataset

[0051] Please see Figure 2The prediction model proposed in this embodiment is built on the PyTorch framework and implemented using Python 3.9. The sliding window length in the prediction model is set to 32, the sliding step size is fixed at 1, and the number of similarity edges connecting a single node in the heterogeneous graph data structure is set to 3. The model has 30 training iterations, a training batch size of 256, and uses the Adam optimizer for model parameter training with a learning rate of 0.001. The prediction model uses root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (COP). Performance evaluation was conducted, and three prediction models—CNN-BiLSTM, Transformer-GRU, and GCN-TCN—were built for comparative experiments. The network frameworks of CNN-BiLSTM and Transformer-GRU were executed sequentially, while the GCN-TCN model framework was similar to that in this embodiment, serving as a spatiotemporally parallel prediction model. Table 5 shows a comparison of the prediction performance metrics for the test set under different models. Table 5 Comparison of prediction performance evaluation metrics for the test set in PHM2012 under different models

[0052] As shown in Table 5, the prediction model based on heterogeneous graph relationship modeling and spatiotemporal feature fusion proposed in this invention exhibits good RUL prediction performance in multiple test bearing samples of the PHM2012 dataset. The average RMSE of the test samples is as low as 0.091, the average MAE is as low as 0.07, and the average coefficient of determination reaches 0.893. Compared with the best GCN-TCN model in the comparison model, it improves by an average of 21.9% in the three indicators, indicating that the predicted RUL curve has a high consistency with the linear degradation label. The spatiotemporal parallel prediction model fully learns the spatial and temporal characteristics of the degradation process. The analysis of the prediction results under different working conditions shows that the proposed model achieves the best evaluation index in most test samples, indicating that the proposed model has a certain adaptability and stability to the performance degradation process of rolling bearings under different speed and load conditions.

[0053] In summary, the bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion in the above embodiments of the present invention constructs a spatiotemporally parallel prediction model, extracts spatial and temporal features of the bearing performance degradation process on the heterogeneous graph convolutional branch and the temporal convolutional branch, respectively, and uses a gating mechanism to achieve spatiotemporal feature fusion, outputting the model's RUL prediction result. The proposed heterogeneous graph data structure improves the utilization of degradation information in vibration signals by defining multiple node types and edge connection types, enhancing the prediction model's ability to represent spatial correlation features. Compared with homogeneous graph data structures, heterogeneous graph data structures and corresponding heterogeneous graph convolutional networks avoid homogeneous processing of heterogeneous features by the prediction model, enhancing the targeting of feature extraction. At the same time, node attention and self-attention mechanisms are introduced to enhance the features of key nodes and key time steps in the prediction data, and the contribution weight of spatiotemporal features is adaptively adjusted through the gating mechanism, making the spatiotemporally fused features better reflect the performance degradation law of rolling bearings, improving the prediction accuracy and stability of the model.

[0054] Example 2 In another aspect, this invention proposes a bearing life prediction system based on the fusion of heterogeneous graphs and spatiotemporal features. Please refer to [link / reference]. Figure 7 The figure shows a bearing life prediction system based on heterogeneous graph and spatiotemporal feature fusion according to a second embodiment of the present invention. The system includes: The signal acquisition module 11 is used to acquire the vibration signals of the rolling bearing in the horizontal and vertical directions. It takes a single vibration signal sample as the smallest feature extraction unit, calculates its time domain features and frequency domain features respectively, and combines the features in the two directions to form a complete feature set. Data processing module 12 is used to slice data on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and to perform standardized data processing on the continuous feature sequence; The spatial feature enhancement module 13 is used to define two node types based on the time domain features and the frequency domain features, and to define two edge connection types based on the similarity of features and the matching relationship between the same feature in different directions. It transforms the continuous feature sequence into a heterogeneous graph data structure, inputs the heterogeneous graph data structure into a heterogeneous graph convolution branch, updates the node features based on similarity edges and matching edges to obtain spatial features, and enhances the spatial features through a node attention mechanism to obtain spatially enhanced features. The temporal feature enhancement module 14 is used to input the continuous feature sequence into the temporal convolution branch, learn the temporal features using multiple temporal convolution residual blocks with different dilation coefficients, and achieve temporal feature enhancement through a self-attention mechanism to obtain temporal enhanced features. The life prediction module 15 is used to input the spatial enhancement features and the temporal enhancement features into the spatiotemporal feature fusion model, dynamically adjust the weights of the two types of features through a gating mechanism, calculate the spatiotemporal fusion features, and input the spatiotemporal fusion features into the fully connected prediction layer to obtain the remaining life prediction result of the rolling bearing.

[0055] Furthermore, the data processing module 12 is specifically used for: The sliding window algorithm is used to slice the data on the complete feature set according to the preset sliding window length and the sliding step size is 1, so as to obtain the feature vector of each vibration signal acquisition sample in the continuous feature sequence corresponding to each time step in the continuous feature sequence in two acquisition directions. The feature dimension of the feature vector is equal to the sum of the types of extracted features. The continuous feature sequence increases the number of training samples while retaining the performance degradation information between adjacent sampling samples. The maximum and minimum values ​​of the extracted features from the training data are used as the scale for feature standardization. The data input to the prediction model are then subjected to min-max standardization according to the following expression.

[0056] Furthermore, the spatial feature enhancement module 13 is specifically used for: Construct the heterogeneous graph data structure ,in, Represents a set of nodes. Denotes the set of edges. Represents a set of node types. Represents a set of edge connection types. express Node type mapping, express Edge connection type mapping; The nodes obtained from time-domain features and frequency-domain features are defined as time-domain feature nodes and frequency-domain feature nodes. By calculating the absolute value of the Pearson correlation coefficient between different feature nodes, the most relevant nodes are selected to establish the similarity edge connection. The matching edge connection is established by searching for feature nodes with the same feature in the horizontal and vertical directions. In the heterogeneous graph convolution branch, the node update process of a single-layer heterogeneous graph convolutional network is as follows:

[0057] In the formula, This represents the input features of a single-layer heterogeneous graph convolutional network. The initial input of the first layer of the heterogeneous graph convolutional network is a heterogeneous graph. The node feature matrix, This represents the spatial features of the output of a single-layer heterogeneous graph convolutional network; By using a node attention mechanism and linear transformation, the spatial features obtained from the convolutional branch of the heterogeneous graph are transformed into a key matrix and a value matrix. The temporal form of the node feature matrix of the heterogeneous graph is linearly transformed into a query matrix. By traversing all feature nodes in the heterogeneous graph through the query matrix at each prediction time step, spatial enhancement features based on the node attention mechanism are obtained.

[0058] Furthermore, the time-series feature enhancement module 14 is specifically used for: In the temporal convolution branch, the constructed temporal convolutional network contains three layers of temporal convolutional residual blocks with different dilation coefficients. Each temporal convolutional residual block consists of a dilated causal convolutional layer, a normalization layer, a nonlinear activation layer, and a residual connection layer. The temporal features obtained from the temporal convolution branch are transformed into a query matrix, a key matrix, and a value matrix using a linear transformation through a self-attention mechanism. By calculating the correlation weights between different prediction time steps, the temporal features of each prediction time step are weighted and fused to obtain the temporal enhancement features.

[0059] Furthermore, the lifetime prediction module is specifically used for: The spatial enhancement features and the temporal enhancement features are input into the spatiotemporal feature fusion model, and the spatial enhancement features and temporal enhancement features are concatenated in the feature dimension to obtain the spatiotemporal joint features; The gating mechanism calculates the gating weights used to measure the correlation between the two types of features and the performance degradation state. The two types of features are then weighted and fused according to the gating weights to obtain the spatiotemporal fusion features. After the spatiotemporal fusion features are input into the fully connected prediction layer, they are mapped to the [0,1] interval through the fully connected layer and the Sigmoid activation function to obtain the prediction result of the remaining service life of the rolling bearing.

[0060] The functions or operation steps implemented by the above modules and units are largely the same as those in the above method embodiments, and will not be repeated here.

[0061] The bearing life prediction system based on heterogeneous graph and spatiotemporal feature fusion provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0062] Example 3 This invention also proposes a computer, please refer to [link / reference]. Figure 8The computer shown in the third embodiment of the present invention includes a memory 10, a processor 20, and a computer program 30 stored in the memory 10 and executable on the processor 20. When the processor 20 executes the computer program 30, it implements the above-described bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion.

[0063] The memory 10 includes at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10 can be an internal storage unit of a computer, such as the computer's hard disk. In other embodiments, the memory 10 can be an external storage device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Furthermore, the memory 10 can include both internal and external storage units of the computer. The memory 10 can be used not only to store application software and various types of data installed on the computer, but also to temporarily store data that has been output or will be output.

[0064] In some embodiments, the processor 20 may be an electronic control unit (ECU), a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip, used to run program code stored in the memory 10 or process data, such as executing access restriction programs.

[0065] It should be pointed out that, Figure 8 The structure shown does not constitute a limitation on the computer. In other embodiments, the computer may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0066] This invention also proposes a storage medium storing a computer program that, when executed by a processor, implements the bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion as described above.

[0067] Those skilled in the art will understand that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0068] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0069] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion, characterized in that, include: The vibration signals of the rolling bearing in the horizontal and vertical directions are acquired. Each vibration signal sample is used as the smallest feature extraction unit. Its time domain features and frequency domain features are calculated respectively, and the features in the two directions are combined to form a complete feature set. Data slices are performed on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and the continuous feature sequence is then subjected to standardized data processing. Two node types are defined based on the time-domain features and the frequency-domain features. Two edge connection types are defined based on the similarity of features and the matching relationship between the same feature in different directions. The continuous feature sequence is transformed into a heterogeneous graph data structure. The heterogeneous graph data structure is input into the heterogeneous graph convolution branch. Node features are updated based on similarity edges and matching edges to obtain spatial features. Spatial feature enhancement is achieved through a node attention mechanism to obtain spatially enhanced features. The continuous feature sequence is input into the temporal convolution branch, and temporal features are learned using multiple temporal convolution residual blocks with different dilation coefficients. Temporal feature enhancement is achieved through a self-attention mechanism to obtain temporally enhanced features. The spatial enhancement features and the temporal enhancement features are input into the spatiotemporal feature fusion model. The weights of the two types of features are dynamically adjusted through a gating mechanism to calculate the spatiotemporal fusion features. The spatiotemporal fusion features are then input into the fully connected prediction layer to obtain the prediction result of the remaining service life of the rolling bearing. The steps involved in transforming the continuous feature sequence into a heterogeneous graph data structure, inputting the heterogeneous graph data structure into a heterogeneous graph convolutional branch, updating node features based on similarity edges and matching edges to obtain spatial features, and enhancing spatial features through a node attention mechanism to obtain spatially enhanced features include: Construct the heterogeneous graph data structure ,in, Represents a set of nodes. Denotes the set of edges. Represents a set of node types. Represents a set of edge connection types. express Node type mapping, express Edge connection type mapping; The nodes obtained from time-domain features and frequency-domain features are defined as time-domain feature nodes and frequency-domain feature nodes. By calculating the absolute value of the Pearson correlation coefficient between different feature nodes, the most relevant nodes are selected to establish the similarity edge connection. The matching edge connection is established by searching for feature nodes with the same feature in the horizontal and vertical directions. In the heterogeneous graph convolution branch, the node update process of a single-layer heterogeneous graph convolutional network is as follows: In the formula, This represents the input features of a single-layer heterogeneous graph convolutional network. The initial input of the first layer of the heterogeneous graph convolutional network is a heterogeneous graph. The node feature matrix, This represents the spatial features of the output of a single-layer heterogeneous graph convolutional network; By using a node attention mechanism and linear transformation, the spatial features obtained from the convolutional branch of the heterogeneous graph are transformed into a key matrix and a value matrix. The temporal form of the node feature matrix of the heterogeneous graph is linearly transformed into a query matrix. By traversing all feature nodes in the heterogeneous graph through the query matrix at each prediction time step, spatial enhancement features based on the node attention mechanism are obtained.

2. The bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion according to claim 1, characterized in that, The time-domain features include maximum value, minimum value, average value, variance, median, range, peak value, absolute mean, root mean square, standard deviation, kurtosis, skewness, waveform factor, peak factor, impulse factor, and margin factor. The frequency-domain features include total spectral energy, median spectral amplitude, maximum spectral amplitude, minimum spectral amplitude, mean spectral amplitude, spectral frequency centroid, spectral frequency standard deviation, and spectral frequency root mean square.

3. The bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion according to claim 1, characterized in that, The steps of slicing data on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and then standardizing the continuous feature sequence include: The sliding window algorithm is used to slice the data on the complete feature set according to the preset sliding window length and the sliding step size is 1, so as to obtain the feature vector of each vibration signal acquisition sample in the continuous feature sequence corresponding to each time step in the continuous feature sequence in two acquisition directions. The feature dimension of the feature vector is equal to the sum of the types of extracted features. The continuous feature sequence increases the number of training samples while retaining the performance degradation information between adjacent sampling samples. The maximum and minimum values ​​of the extracted features from the training data are used as the scale for feature standardization. The data input to the prediction model are then subjected to min-max standardization according to the following expression.

4. The bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion according to claim 1, characterized in that, The steps of inputting the continuous feature sequence into the temporal convolution branch, learning temporal features using multiple temporal convolution residual blocks with different dilation coefficients, and enhancing temporal features through a self-attention mechanism to obtain enhanced temporal features include: In the temporal convolution branch, the constructed temporal convolutional network contains three layers of temporal convolutional residual blocks with different dilation coefficients. Each temporal convolutional residual block consists of a dilated causal convolutional layer, a normalization layer, a nonlinear activation layer, and a residual connection layer. The temporal features obtained from the temporal convolution branch are transformed into a query matrix, a key matrix, and a value matrix using a linear transformation through a self-attention mechanism. By calculating the correlation weights between different prediction time steps, the temporal features of each prediction time step are weighted and fused to obtain the temporal enhancement features.

5. The bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion according to claim 1, characterized in that, The steps of inputting the spatiotemporal fusion features into a fully connected prediction layer to obtain the remaining service life prediction result of the rolling bearing include: The spatial enhancement features and the temporal enhancement features are input into the spatiotemporal feature fusion model, and the spatial enhancement features and temporal enhancement features are concatenated in the feature dimension to obtain the spatiotemporal joint features; The gating mechanism calculates the gating weights used to measure the correlation between the two types of features and the performance degradation state. The two types of features are then weighted and fused according to the gating weights to obtain the spatiotemporal fusion features. After the spatiotemporal fusion features are input into the fully connected prediction layer, they are mapped to the [0,1] interval through the fully connected layer and the Sigmoid activation function to obtain the prediction result of the remaining service life of the rolling bearing.

6. A bearing life prediction system based on heterogeneous graph and spatiotemporal feature fusion, characterized in that, include: The signal acquisition module is used to acquire the vibration signals of the rolling bearing in the horizontal and vertical directions. It takes a single vibration signal sample as the smallest feature extraction unit, calculates its time domain features and frequency domain features respectively, and combines the features in the two directions to form a complete feature set. The data processing module is used to slice the data on the complete feature set according to a preset sliding window length and step size to obtain a continuous feature sequence, and to perform standardized data processing on the continuous feature sequence. The spatial feature enhancement module is used to define two node types based on the time domain features and the frequency domain features, and two edge connection types based on the similarity of features and the matching relationship between the same feature in different directions. It transforms the continuous feature sequence into a heterogeneous graph data structure, inputs the heterogeneous graph data structure into a heterogeneous graph convolution branch, updates the node features based on similarity edges and matching edges to obtain spatial features, and enhances the spatial features through a node attention mechanism to obtain spatially enhanced features. The temporal feature enhancement module is used to input the continuous feature sequence into the temporal convolution branch, learn the temporal features using multiple temporal convolution residual blocks with different dilation coefficients, and enhance the temporal features through a self-attention mechanism to obtain the temporal enhanced features. The life prediction module is used to input the spatial enhancement features and the temporal enhancement features into the spatiotemporal feature fusion model, dynamically adjust the weights of the two types of features through a gating mechanism, calculate the spatiotemporal fusion features, and input the spatiotemporal fusion features into the fully connected prediction layer to obtain the remaining life prediction result of the rolling bearing. Specifically, the spatial feature enhancement module is used for: Construct the heterogeneous graph data structure ,in, Represents a set of nodes. Denotes the set of edges. Represents a set of node types. Represents a set of edge connection types. express Node type mapping, express Edge connection type mapping; The nodes obtained from time-domain features and frequency-domain features are defined as time-domain feature nodes and frequency-domain feature nodes. By calculating the absolute value of the Pearson correlation coefficient between different feature nodes, the most relevant nodes are selected to establish the similarity edge connection. The matching edge connection is established by searching for feature nodes with the same feature in the horizontal and vertical directions. In the heterogeneous graph convolution branch, the node update process of a single-layer heterogeneous graph convolutional network is as follows: In the formula, This represents the input features of a single-layer heterogeneous graph convolutional network. The initial input of the first layer of the heterogeneous graph convolutional network is a heterogeneous graph. The node feature matrix, This represents the spatial features of the output of a single-layer heterogeneous graph convolutional network; By using a node attention mechanism and linear transformation, the spatial features obtained from the convolutional branch of the heterogeneous graph are transformed into a key matrix and a value matrix. The temporal form of the node feature matrix of the heterogeneous graph is linearly transformed into a query matrix. By traversing all feature nodes in the heterogeneous graph through the query matrix at each prediction time step, spatial enhancement features based on the node attention mechanism are obtained.

7. The bearing life prediction system based on heterogeneous graph and spatiotemporal feature fusion according to claim 6, characterized in that, The data processing module is specifically used for: The sliding window algorithm is used to slice the data on the complete feature set according to the preset sliding window length and the sliding step size is 1, so as to obtain the feature vector of each vibration signal acquisition sample in the continuous feature sequence corresponding to each time step in the continuous feature sequence in two acquisition directions. The feature dimension of the feature vector is equal to the sum of the types of extracted features. The continuous feature sequence increases the number of training samples while retaining the performance degradation information between adjacent sampling samples. The maximum and minimum values ​​of the extracted features from the training data are used as the scale for feature standardization. The data input to the prediction model are then subjected to min-max standardization according to the following expression.

8. A readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion as described in any one of claims 1 to 5.

9. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the bearing life prediction method based on heterogeneous graph and spatiotemporal feature fusion as described in any one of claims 1 to 5.

Images