A bearing residual life prediction method and system of a double-scale space-time hybrid supergraph

Abstract

The application provides a bearing residual life prediction method and system of a double-scale space-time mixed hypergraph, and the method comprises the following steps: obtaining a node of a graph data structure, extracting time domain features and frequency domain features of the node; obtaining a node feature matrix by using a sliding window; calculating an adjacency matrix of a semi-connected ordinary graph and a correlation matrix of a semi-connected hypergraph, generating an adjacency matrix of a full-connected ordinary graph and a correlation matrix of a full-connected hypergraph; inputting the adjacency matrix of the full-connected ordinary graph and the semi-connected ordinary graph and the correlation matrix of the full-connected hypergraph and the semi-connected hypergraph into a graph convolutional neural network and a hypergraph convolutional neural network, obtaining two fused high-order-low-order spatial features through feature fusion, then obtaining double-scale fused high-order-low-order spatial features, obtaining a time sequence dependence relationship through a BiLSTM network, inputting the BiLSTM network into a KAN prediction layer, and obtaining a residual life prediction value of a rolling bearing. The application utilizes complex space-time correlation information in a vibration signal, and has obvious advantages.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing remaining service life prediction technology, and in particular to a bearing remaining service life prediction method and system based on a dual-scale spatiotemporal hybrid hypergraph. Background Technology

[0002] In modern industrial production, rolling bearings play a crucial role in supporting the normal operation of rotating components in increasingly automated and integrated machinery. Under complex external operating conditions, the performance and expected life of rolling bearings are prone to irreversible degradation, seriously threatening the safety and continuity of production operations. Remaining useful life (RUL) prediction, as an important technology in predictive health management (PHM), primarily aims to accurately predict the end of normal equipment operation by monitoring the performance degradation status of key components. Therefore, rolling bearing RUL prediction is essential for developing ad-hoc equipment maintenance plans, not only reducing downtime and maintenance costs but also comprehensively improving production efficiency.

[0003] In the field of rolling bearing RUL prediction based on deep learning, researchers have focused on temporal models such as Temporal Convolutional Networks (TCN), Long Short-Term Memory Networks (LSTM), and Transformers. Leveraging the model's ability to extract temporal information, these models are used as temporal encoders to identify features and trends in time series, thereby completing the RUL prediction task. The aforementioned temporal model-based RUL prediction methods typically use a sliding window to read adjacent sequences or pixels to extract features, which is a typical Euclidean data processing method. However, the vibration signals of rolling bearings contain non-Euclidean data containing performance degradation information. This non-Euclidean data has a clear spatial topological structure that changes with the mechanical degradation state, but temporal models, limited by the structure, cannot capture the spatial dependencies of this data. In single-sensor prediction tasks, temporal models are not very sensitive to features of non-adjacent sequences; in multi-sensor prediction tasks, the spatial correlations between sensors are also ignored.

[0004] To address the limitations of temporal models in processing non-Euclidean data, graph-structured data and graph neural networks (GNNs) offer a promising solution for simultaneously capturing temporal and spatial information. Nodes and edges in graph-structured data represent the data's own information and the relationships between different nodes, respectively. The initial design of GNNs was to learn the correlations between graph-structured data during modeling, naturally representing the spatial relationships between different nodes in the system. Current research, based on confirming the existence of spatial correlations in rolling bearing vibration signals, uses node proximity, sample similarity, and node correlation as entry points for mining performance degradation features. Similar to deep learning RUL prediction methods based on Euclidean data, GNN-type models occupy the niche previously held by CNN-type models in feature extraction. This involves first using GNN-type models to deconstruct graph data to learn spatial features, and then using temporal models to capture the temporal dependencies between these spatial features.

[0005] In existing GNN rolling bearing remaining service life prediction technology, the default node relationship in ordinary graph structure is only pairwise, that is, the edge in ordinary graph structure can only connect two nodes. It cannot effectively model the complex relationship between nodes in real-world scenarios. In other words, ordinary graph data structure is insufficient in learning more complex relationship features such as higher-order statistics and heterogeneous time dynamics between multiple time nodes. Summary of the Invention To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting the remaining life of bearings using a dual-scale spatiotemporal hybrid hypergraph, thereby resolving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for predicting the remaining life of bearings using a dual-scale spatiotemporal hybrid hypergraph includes the following steps: In the full-life vibration signal of a rolling bearing, a single sample is regarded as a node in a graph data structure, and the time-domain and frequency-domain features of the node in the horizontal and vertical directions are extracted respectively to obtain a feature matrix. The feature matrix of all nodes in the whole life vibration signal of the rolling bearing is obtained by using a sliding window to extract the feature matrix of all nodes; Based on the KNN algorithm, the adjacency matrix of the semi-connected ordinary graph and the incidence matrix of the semi-connected hypergraph are calculated using the node feature matrix. Then, all nodes are traversed to generate the adjacency matrix of the fully connected ordinary graph and the incidence matrix of the fully connected hypergraph. Construct a graph convolutional neural network and a hypergraph convolutional neural network. Input the adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively to update node features. Based on the attention weight feature fusion module, obtain the first fused high-order-low-order spatial features. The adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph are respectively input into the graph convolutional neural network and the hypergraph convolutional neural network for node feature update. Based on the attention weight feature fusion module, the second fused high-order-low-order spatial features are obtained. The attention-weighted feature fusion module integrates the first fused high- and low-level spatial features and the second fused high- and low-level spatial features to obtain dual-scale fused high- and low-level spatial features. After learning the temporal dependencies of the dual-scale fused high- and low-level spatial features through a BiLSTM network, the information is input into the KAN prediction layer to obtain the remaining life prediction value of the rolling bearing.

[0007] According to one aspect of the above technical solution, the time-domain features include maximum value, minimum value, variance, peak-to-peak value, median, 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 energy, median energy, maximum frequency, minimum frequency, average frequency, frequency center, standard deviation frequency, and root mean square frequency.

[0008] According to one aspect of the above technical solution, the step is based on the KNN algorithm, using the node feature matrix to calculate the adjacency matrix of the semi-connected general graph and the association matrix of the semi-connected hypergraph, wherein the adjacency matrix of the semi-connected general graph... The expression is: ; The association matrix of the semi-connected hypergraph The expression is: ; in, and Let n represent the node feature matrix, and n represent the number of nodes.

[0009] According to one aspect of the above technical solution, in the steps of inputting the adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature updating, and in the steps of inputting the adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature updating, the expression for node feature updating through the graph convolutional neural network is: ; The expression for updating node features using a hypergraph convolutional neural network is: ; in, x l+1This represents the result obtained after updating through a graph convolutional neural network. l The node features of layer +1, σ represents the activation function, D represents the node degree matrix calculated by the graph convolutional neural network, and W represents the learnable weight matrix for the node feature transformation process in the graph convolutional neural network. This represents the adjacency matrix with added self-loops. x’ l+1 This represents the result obtained after updating through a hypergraph convolutional neural network. l +1 layer node features, D v D represents the node degree matrix computed using a hypergraph convolutional neural network. e W represents the degree matrix of the hyperedge. e W represents the hyperedge weights learned in a hypergraph convolutional neural network. l H represents the learnable weight matrix for the feature transformation process of nodes in the l-th layer hypergraph convolutional neural network, and H represents the correlation matrix. According to one aspect of the above technical solution, the updated node features of the fully connected general graph are represented as X. 1 G The updated node features of the fully connected hypergraph are represented as X. 1 HG The step, based on the attention weight feature fusion module, yields the expression for the first fusion of high-order and low-order spatial features as follows: ; in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 1 m f represents the first fusion weight. 1 m This represents the first fusion of high-order and low-order spatial features.

[0010] According to one aspect of the above technical solution, the updated node features of the semi-connected general graph are represented as X. 2 G The updated node features of the semi-connected hypergraph are represented as X. 2 HG The step, based on the attention weight feature fusion module, yields the expression for the second fused high-order-low-order spatial features as follows: ; in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 2 m f represents the second fusion weight. 2 mThis represents the second fusion of high-order and low-order spatial features.

[0011] According to one aspect of the above technical solution, the attention-weight-based feature fusion module integrates the first fused high-order-low-order spatial features and the second fused high-order-low-order spatial features to obtain the expression for the dual-scale fused high-order-low-order spatial features as follows: ; in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 3 m f represents the third fusion weight. 3 m This represents the dual-scale fusion of high-order and low-order spatial features.

[0012] According to one aspect of the above technical solution, the BiLSTM network consists of a forward LSTM layer and a backward LSTM layer, both of which are connected to the KAN prediction layer. The input of the BiLSTM network is a first fused high-order-low-order spatial feature, a second fused high-order-low-order spatial feature, and a dual-scale fused high-order-low-order spatial feature.

[0013] According to one aspect of the above technical solution, the expression for the KAN prediction layer is: ; in, and Let these represent the external and internal functions of the KAN network, respectively. x This represents the output of the BiLSTM network, where γ represents the dimension. , , w represents a trainable parameter that controls the overall output size, and c i B represents several trainable parameters. i (x) represents a predefined spline basis function.

[0014] This invention also provides a bearing remaining life prediction system based on a dual-scale spatiotemporal hybrid hypergraph, comprising: Extraction module: used to treat a single sample as a node in a graph data structure in the whole life vibration signal of a rolling bearing, and extract the time domain features and frequency domain features of the node in the horizontal and vertical directions respectively to obtain a feature matrix; The extraction module is used to extract the feature matrices of all nodes in the full-life vibration signal of a rolling bearing using a sliding window, thus obtaining the node feature matrix. The generation module is used to calculate the adjacency matrix of the semi-connected ordinary graph and the incidence matrix of the semi-connected hypergraph based on the KNN algorithm and the node feature matrix. It then traverses all nodes to generate the adjacency matrix of the fully connected ordinary graph and the incidence matrix of the fully connected hypergraph. The first computation module is used to construct a graph convolutional neural network and a hypergraph convolutional neural network. The adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph are respectively input into the graph convolutional neural network and the hypergraph convolutional neural network for node feature update. The feature fusion module based on attention weights obtains the first fused high-order-low-order spatial features. The second calculation module is used to input the adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature update, and to obtain the second fused high-order-low-order spatial features based on the attention weight feature fusion module. The third calculation module is used for the attention-weighted feature fusion module. It integrates the first fused high-order-low-order spatial features and the second fused high-order-low-order spatial features to obtain dual-scale fused high-order-low-order spatial features. After learning the temporal dependency of the dual-scale fused high-order-low-order spatial features through a BiLSTM network, it is input into the KAN prediction layer to obtain the remaining life prediction value of the rolling bearing.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This paper proposes a method to extract spatial degradation features of ordinary graph data structures and hypergraph data structures at fully connected and semi-connected scales, respectively. An attention-weighted feature fusion module is introduced to achieve feature fusion across high-order and low-order feature dimensions, improving the utilization of graph data structures. A spatiotemporal feature extraction network is constructed, executing in a spatial-temporal sequence to capture the correlation of node information sequentially from both temporal and spatial dimensions. Furthermore, a KAN network is used to replace the traditional MLP prediction layer, achieving higher prediction accuracy with fewer parameters.

[0016] The dual-scale spatiotemporal hybrid hypergraph model proposed in this invention effectively utilizes the complex spatiotemporal correlation information in vibration signals, and has significant advantages over existing traditional graph neural network-based prediction methods. Attached Figure Description

[0017] Figure 1 This is a flowchart of the bearing remaining life prediction method using a dual-scale spatiotemporal hybrid hypergraph in the first embodiment of the present invention; Figure 2 This is a graph showing the prediction results for the test set Bearing1-3 in the first embodiment of the present invention; Figure 3This is a structural block diagram of the bearing remaining life prediction system based on a dual-scale spatiotemporal hybrid hypergraph in the second embodiment of the present invention. The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] 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.

[0019] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0020] 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.

[0021] Please see Figure 1 The image shows a bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to the first embodiment of the present invention, comprising the following steps: S10, In the whole life vibration signal of the rolling bearing, a single sample is regarded as a node of the graph data structure, and the time domain features and frequency domain features of the node in the horizontal and vertical directions are extracted respectively to obtain the feature matrix. S20, use a sliding window to extract the feature matrix of all nodes in the full-life vibration signal of the rolling bearing, and obtain the node feature matrix; S30, based on the KNN algorithm, uses the node feature matrix to calculate the adjacency matrix of the semi-connected ordinary graph and the incidence matrix of the semi-connected hypergraph, traverses all nodes, and generates the adjacency matrix of the fully connected ordinary graph and the incidence matrix of the fully connected hypergraph. S40, construct a graph convolutional neural network and a hypergraph convolutional neural network, input the adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature update, and obtain the first fused high-order-low-order spatial features based on the feature fusion module of attention weight; S50, the adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph are respectively input into the graph convolutional neural network and the hypergraph convolutional neural network for node feature update, and the second fused high-order-low-order spatial features are obtained based on the feature fusion module of attention weight. S60, the attention-weighted feature fusion module integrates the first fused high-order-low-order spatial features and the second fused high-order-low-order spatial features to obtain dual-scale fused high-order-low-order spatial features. After learning the temporal dependency of the dual-scale fused high-order-low-order spatial features through a BiLSTM network, the feature fusion module inputs the feature fusion module into the KAN prediction layer to obtain the remaining life prediction value of the rolling bearing.

[0022] Understandably, this invention improves the utilization of graph data structures by proposing a method to extract spatial degradation features of ordinary graph data structures and hypergraph data structures at fully connected and semi-connected scales, respectively, and introducing an attention-weighted feature fusion module to complete feature fusion across high-order and low-order feature dimensions. A spatiotemporal feature extraction network executing in a spatial-temporal sequence is constructed, capturing the correlation of node information sequentially from the temporal and spatial dimensions, and using a KAN network to replace the traditional MLP prediction layer, achieving higher prediction accuracy with fewer parameters.

[0023] The dual-scale spatiotemporal hybrid hypergraph model proposed in this invention effectively utilizes the complex spatiotemporal correlation information in vibration signals, and has significant advantages over existing traditional graph neural network-based prediction methods.

[0024] Furthermore, the time-domain features include maximum value, minimum value, variance, peak-to-peak value, median, 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 energy, median energy, maximum frequency, minimum frequency, average frequency, frequency center, standard deviation frequency, and root mean square frequency.

[0025] Furthermore, for step S20, to maximize the use of the training set data, the feature matrix of the rolling bearing's full-life vibration signal obtained in step S10 is used to extract continuous horizontal and vertical vibration signals through a sliding window with a step size of 1 and a length of 32. After completing the feature extraction in step S10, a training set of data with dimensions [32, 24×2] is obtained for subsequent prediction model training, meaning that the feature matrix in each node is 48-dimensional. After sampling, a node feature matrix of a graph data structure is obtained. Each node in the graph data structure corresponds to the time-domain and frequency-domain features of a sampled signal in the rolling bearing's full-life vibration signal, and is accompanied by a corresponding remaining service life label. The sliding window length is a custom hyperparameter.

[0026] Furthermore, in step S30, based on the KNN algorithm, the adjacency matrix of the semi-connected general graph and the association matrix of the semi-connected hypergraph are calculated using the node feature matrix. The expression is: ; The association matrix of the semi-connected hypergraph The expression is: ; in, and Let n represent the node feature matrix, and n represent the number of nodes.

[0027] Understandably, a regular graph is represented by a node feature matrix and an adjacency matrix, while a hypergraph is represented by a node feature matrix and an association matrix. Based on the window data obtained in step S20, a graph data structure is constructed, mapping independent vibration signal samples to independent nodes, and establishing edge connections (i.e., adjacency matrices) between nodes according to the distance relationships between their feature matrices, thus forming a regular graph data structure used to characterize bearing degradation. Furthermore, to reduce the adverse effects of insufficient node connections in the graph data structure on model training, the KNN algorithm is used to perform neighborhood search on nodes, establishing edge connections between each node and several of its nearest neighbors. In the defined semi-connected regular graph, a single graph data structure consists of 32 nodes, with each node establishing edge connections with half of the nodes in the graph. In the defined fully connected regular graph, the adjacency matrix is ​​a fixed square matrix of all 1s, where every node in the graph establishes edge connections with all other nodes, and the all-1 adjacency matrix completely represents the node association information across the entire graph. Similarly, the sample data processed by the sliding window are mapped one by one to the node feature matrices in the hypergraph, and hyperedges connecting multiple nodes are constructed based on the distance relationships between the node feature matrices, thus forming a hypergraph data structure representing the high-order associations of the graph data structure. Furthermore, taking all nodes as the central nodes, hyperedge connections at the fully connected and semi-connected scales are calculated respectively to construct fully connected hypergraphs and semi-connected hypergraphs.

[0028] Furthermore, for steps S40 and S50, the steps of inputting the adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature updating, and the steps of inputting the adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature updating, the expression for node feature updating through the graph convolutional neural network is: ; The expression for updating node features using a hypergraph convolutional neural network is: ; in, x l+1 This represents the result obtained after updating through a graph convolutional neural network. l The node features of layer +1, σ represents the activation function, D represents the node degree matrix calculated by the graph convolutional neural network, and W represents the learnable weight matrix for the node feature transformation process in the graph convolutional neural network. This represents the adjacency matrix with added self-loops. x’ l+1 This represents the result obtained after updating through a hypergraph convolutional neural network. l +1 layer node features, D vD represents the node degree matrix computed using a hypergraph convolutional neural network. e W represents the degree matrix of the hyperedge. e W represents the hyperedge weights learned in a hypergraph convolutional neural network. l H represents the learnable weight matrix for the feature transformation process of nodes in the l-th layer hypergraph convolutional neural network, and H represents the correlation matrix.

[0029] Understandably, if the adjacency matrix of a fully connected ordinary graph is input into a graph convolutional neural network, This represents the adjacency matrix of a fully connected ordinary graph. If the adjacency matrix of a semi-connected ordinary graph is input into a graph convolutional neural network, This represents the adjacency matrix of a semi-connected ordinary graph. The same applies to hypergraphs, which will not be elaborated here. The fully connected ordinary graph and fully connected hypergraph constructed in step S30 are input into the graph convolutional neural network and the hypergraph convolutional neural network, respectively, to update the node features in the graph structure and extract spatial association information at different connection scales. The dimensionality changes of the node feature matrix in the graph data structure remain consistent in both the graph convolutional network and the hypergraph convolutional network. The dimensionality changes performed in the two-layer network structure are: [32,48], [32,300], [32,200]. The node feature matrix updated by the graph convolutional network represents the low-order spatial features based on pairwise connections between nodes, while the node feature matrix updated by the hypergraph convolutional network reflects the high-order spatial features of multiple nodes iterating together. The spatial feature extraction methods for semi-connected ordinary graphs and semi-connected hypergraphs are consistent with the fully connected scale. Through the above steps, the updated node features of fully connected ordinary graphs, semi-connected ordinary graphs, fully connected hypergraphs, and semi-connected hypergraphs can be calculated.

[0030] Furthermore, for steps S40, S50, and S60, the updated node features of the fully connected general graph are represented as X. 1 G The updated node features of the fully connected hypergraph are represented as X. 1 HG The step, based on the attention weight feature fusion module, yields the expression for the first fusion of high-order and low-order spatial features as follows: ; in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 1 m f represents the first fusion weight. 1 m This represents the first fusion of high-order and low-order spatial features; The updated node features of the semi-connected general graph are represented as X.2 G The updated node features of the semi-connected hypergraph are represented as X. 2 HG The step, based on the attention weight feature fusion module, yields the expression for the second fused high-order-low-order spatial features as follows: ; in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 2 m f represents the second fusion weight. 2 m This represents the second fusion of higher-order and lower-order spatial features; The attention-weight-based feature fusion module integrates the first fused high-order-low-order spatial features and the second fused high-order-low-order spatial features to obtain the expression for the dual-scale fused high-order-low-order spatial features as follows: ; in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 3 m f represents the third fusion weight. 3 m This represents the dual-scale fusion of high-order and low-order spatial features.

[0031] Understandably, the updated node features output by the graph convolutional network and the node feature matrix output by the hypergraph convolutional network are input into the attention-weighted feature fusion module. This module adaptively calculates the corresponding attention weights based on the importance of different feature branches to the prediction task, and performs weighted fusion of low-order and high-order spatial features to obtain fused high-order-low-order spatial features at both the fully connected and semi-connected scales. Considering the potential redundant edge connections introduced by the fully connected graph (hypergraph) data structure, an attention mechanism is introduced to assign different weights to the updated node features for weighting, achieving adaptive high-order-low-order spatial feature fusion. Then, the feature fusion module further integrates the fused high-order-low-order spatial features at both the fully connected and semi-connected scales to obtain dual-scale fused high-order-low-order spatial features.

[0032] Furthermore, the BiLSTM network consists of a forward LSTM layer and a backward LSTM layer, both of which are connected to the KAN prediction layer. The inputs of the BiLSTM network are a first fused high-order-low-order spatial feature, a second fused high-order-low-order spatial feature, and a dual-scale fused high-order-low-order spatial feature.

[0033] Understandably, the BiLSTM network is a variant of the LSTM network, designed to address the gradient vanishing and bidirectional information extraction problems inherent in LSTM networks. It aims to enable the network to mine long-term correlations over time. The BiLSTM network consists of forward LSTM layers and backward LSTM layers, processing time-series information in two directions simultaneously. The forward LSTM layer trains the time series forward, while the backward LSTM layer trains the time series backward; both are directly connected to the KAN prediction layer. The input to the BiLSTM network consists of three fused high-order-low-order spatial features: a first fused high-order-low-order spatial feature, a second fused high-order-low-order spatial feature, and a dual-scale fused high-order-low-order spatial feature. These three features collectively extract the spatial dimension degradation features of the rolling bearing. The BiLSTM then utilizes its own characteristics to extract the temporal dimension degradation features of the rolling bearing, thus achieving spatiotemporal feature fusion. The dual-scale fusion of high-order and low-order spatial features is input into the BiLSTM network. By combining forward and backward temporal modeling, the evolution of spatial features in the time dimension and the correlation between time series are learned to extract the temporal degradation features of the rolling bearing operating state over time. In the single-layer BiLSTM network, the data change dimensions are: [32,200], [32,100].

[0034] Furthermore, the expression for the KAN prediction layer is: ; in, and Let these represent the external and internal functions of the KAN network, respectively. x This represents the output of the BiLSTM network, where γ represents the dimension. , , w represents a trainable parameter that controls the overall output size, and c i B represents several trainable parameters. i (x) represents a predefined spline basis function.

[0035] Understandably, after obtaining the temporally dependent enhanced feature representation through the BiLSTM network, it is input into the KAN prediction layer. In the two KAN networks, the data change dimensions are [32,50] and [50,1]. The KAN prediction layer fits complex nonlinear relationships through a combination of learnable univariate functions, realizing the mapping from spatiotemporal features containing rolling bearing degradation information to remaining service life. Finally, the output of the KAN prediction layer is transformed into a RUL prediction value in the interval [0,1] by the sigmoid function, which is the predicted value of the remaining service life of the rolling bearing.

[0036] To verify the predictive performance of the proposed model, experiments were conducted using the XJTU-SY dataset. The dataset collected complete lifetime vibration signals of rolling bearings in both horizontal and vertical directions. The sampling frequency was set to 25.6 kHz, with one sample collected per minute. Each sampled data file contained 32,768 data points. Taking operating condition one in the dataset as an example, Bearing1-1 and Bearing1-2 were used as the training set, and Bearing1-3, Bearing1-4, and Bearing1-5 were used as the test set.

[0037] The prediction experiment used the original vibration signal as input, extracted time-domain and frequency-domain features, and then constructed a graph data structure to import into the prediction model. Experiments determined that the loss function of the prediction model during training was set to mean squared error, the number of training iterations was 100, the learning rate was set to 0.001, the training batch size was 64, and the Adam optimizer was used to update hyperparameters during model iteration.

[0038] Using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) 2 Three evaluation metrics widely used in prediction tasks are used to assess model performance.

[0039] To verify the effectiveness of the prediction model, comparative experiments were conducted using Transformer, CNN-BiGRU, and GCN-LSTM models. Transformer and CNN-BiGRU are traditional time-series prediction models, while GCN-LSTM, with a framework similar to the proposed model, sequentially extracts spatial features and temporal dependencies during the performance degradation process of rolling bearings. The experimental parameters and data preprocessing for all comparative experiments remained consistent with the proposed method.

[0040] Figure 2 The figure compares the prediction results for Bearing1-3, with RUL Label representing the remaining service life. As can be seen from the figure, the proposed model's overall predictions are closest to the actual values, effectively capturing the degradation trend of the bearing in both the initial and late stages of its lifespan. Compared to the Transformer, CNN-BiGRU, and GCN-LSTM models, the proposed model more accurately describes the degradation trend of the remaining lifespan in the mid-to-late stages of prediction, avoiding the significant prediction lag and abnormal fluctuations found in the Transformer model, and demonstrating the best prediction accuracy and stability. The table below compares the prediction performance evaluation metrics of different models on the test set under operating condition 1.

[0041]

[0042] As shown in the table above, the proposed model achieved the best prediction performance in all test samples. Compared to the comparison models, the proposed model showed significantly lower RMSE and MAE evaluation metrics. Compared to GCN-LSTM, which is also a graph neural network, the average values ​​of the two evaluation metrics decreased by 44.37% and 51.06%, respectively. Meanwhile, the proposed model achieved the highest coefficient of determination (0.905), indicating that it can more accurately predict the performance degradation trend of rolling bearings. In Bearing1-5, the prediction performance of the comparison models all showed varying degrees of decline, but the proposed model still maintained high prediction accuracy, demonstrating its stronger robustness.

[0043] In summary, the bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph in the above embodiments of the present invention effectively utilizes the complex spatiotemporal correlation information in vibration signals, and has significant advantages compared with existing traditional prediction methods based on graph neural networks. Please refer to Figure 3 The figure shows a bearing remaining life prediction system based on a dual-scale spatiotemporal hybrid hypergraph in the second embodiment of the present invention, comprising: Extraction module 11: used to treat a single sample as a node in a graph data structure in the whole life vibration signal of a rolling bearing, and extract the time domain features and frequency domain features of the node in the horizontal and vertical directions respectively to obtain a feature matrix; Extraction module 12: Used to extract the feature matrix of all nodes in the full-life vibration signal of the rolling bearing using a sliding window, and obtain the node feature matrix; Generation Module 13: Based on the KNN algorithm, it calculates the adjacency matrix of the semi-connected ordinary graph and the incidence matrix of the semi-connected hypergraph using the node feature matrix, traverses all nodes, and generates the adjacency matrix of the fully connected ordinary graph and the incidence matrix of the fully connected hypergraph. First computation module 14: used to construct graph convolutional neural network and hypergraph convolutional neural network, input the adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature update, and obtain the first fused high-order-low-order spatial features based on the attention weight feature fusion module. The second calculation module 15 is used to input the adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature update, and obtain the second fused high-order-low-order spatial features based on the attention weight feature fusion module. The third calculation module 16 is used for the attention weight-based feature fusion module. It integrates the first fused high-order-low-order spatial features and the second fused high-order-low-order spatial features to obtain dual-scale fused high-order-low-order spatial features. After learning the temporal dependency of the dual-scale fused high-order-low-order spatial features through a BiLSTM network, it is input into the KAN prediction layer to obtain the remaining life prediction value of the rolling bearing.

[0044] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0045] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for predicting the remaining life of bearings using a dual-scale spatiotemporal hybrid hypergraph, characterized in that, Includes the following steps: In the full-life vibration signal of a rolling bearing, a single sample is regarded as a node in a graph data structure, and the time-domain and frequency-domain features of the node in the horizontal and vertical directions are extracted respectively to obtain a feature matrix. The feature matrix of all nodes in the whole life vibration signal of the rolling bearing is obtained by using a sliding window to extract the feature matrix of all nodes; Based on the KNN algorithm, the adjacency matrix of the semi-connected ordinary graph and the incidence matrix of the semi-connected hypergraph are calculated using the node feature matrix. Then, all nodes are traversed to generate the adjacency matrix of the fully connected ordinary graph and the incidence matrix of the fully connected hypergraph. Construct a graph convolutional neural network and a hypergraph convolutional neural network. Input the adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively to update node features. Based on the attention weight feature fusion module, obtain the first fused high-order-low-order spatial features. The adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph are respectively input into the graph convolutional neural network and the hypergraph convolutional neural network for node feature update. Based on the attention weight feature fusion module, the second fused high-order-low-order spatial features are obtained. The attention-weighted feature fusion module integrates the first fused high- and low-level spatial features and the second fused high- and low-level spatial features to obtain dual-scale fused high- and low-level spatial features. After learning the temporal dependencies of the dual-scale fused high- and low-level spatial features through a BiLSTM network, the information is input into the KAN prediction layer to obtain the remaining life prediction value of the rolling bearing.

2. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 1, characterized in that, The time-domain features include maximum value, minimum value, variance, peak-to-peak value, median, 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 energy, median energy, maximum frequency, minimum frequency, average frequency, frequency center, standard deviation frequency, and root mean square frequency.

3. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 1, characterized in that, The steps described are based on the KNN algorithm, using the node feature matrix to calculate the adjacency matrix of the semi-connected general graph and the association matrix of the semi-connected hypergraph. The adjacency matrix of the semi-connected general graph... The expression is: ； The association matrix of the semi-connected hypergraph The expression is: ； in, and Let n represent the node feature matrix, and n represent the number of nodes.

4. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 3, characterized in that, The steps of inputting the adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature updating, and the steps of inputting the adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature updating, are described by the expression for node feature updating via the graph convolutional neural network: ； The expression for updating node features using a hypergraph convolutional neural network is: ； in, x l+1 This represents the result obtained after updating through a graph convolutional neural network. l The node features of layer +1, σ represents the activation function, D represents the node degree matrix calculated by the graph convolutional neural network, and W represents the learnable weight matrix for the node feature transformation process in the graph convolutional neural network. This represents the adjacency matrix with added self-loops. x’ l+1 This represents the result obtained after updating through a hypergraph convolutional neural network. l +1 layer node features, D v D represents the node degree matrix computed using a hypergraph convolutional neural network. e W represents the degree matrix of the hyperedge. e W represents the hyperedge weights learned in a hypergraph convolutional neural network. l H represents the learnable weight matrix for the feature transformation process of nodes in the l-th layer hypergraph convolutional neural network, and H represents the correlation matrix.

5. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 4, characterized in that, The updated node features of the fully connected general graph are represented as X. 1 G The updated node features of the fully connected hypergraph are represented as X. 1 HG The step, based on the attention weight feature fusion module, yields the expression for the first fusion of high-order and low-order spatial features as follows: ； in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 1 m f represents the first fusion weight. 1 m This represents the first fusion of high-order and low-order spatial features.

6. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 5, characterized in that, The updated node features of the semi-connected general graph are represented as X. 2 G The updated node features of the semi-connected hypergraph are represented as X. 2 HG The step, based on the attention weight feature fusion module, yields the expression for the second fused high-order-low-order spatial features as follows: ； in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 2 m f represents the second fusion weight. 2 m This represents the second fusion of high-order and low-order spatial features.

7. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 6, characterized in that, The attention-weight-based feature fusion module integrates the first fused high-order-low-order spatial features and the second fused high-order-low-order spatial features to obtain the expression for the dual-scale fused high-order-low-order spatial features as follows: ； in, , Represents the attention weights obtained during training, ReLU represents the activation function, and W... 3 m f represents the third fusion weight. 3 m This represents the dual-scale fusion of high-order and low-order spatial features.

8. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 1, characterized in that, The BiLSTM network consists of a forward LSTM layer and a backward LSTM layer, both of which are connected to the KAN prediction layer. The inputs of the BiLSTM network are a first fused high-order-low-order spatial feature, a second fused high-order-low-order spatial feature, and a dual-scale fused high-order-low-order spatial feature.

9. The bearing remaining life prediction method based on a dual-scale spatiotemporal hybrid hypergraph according to claim 1, characterized in that, The expression for the KAN prediction layer is: ； in, and Let these represent the external and internal functions of the KAN network, respectively. x This represents the output of the BiLSTM network, where γ represents the dimension. , , w represents a trainable parameter that controls the overall output size, and c i B represents several trainable parameters. i (x) represents a predefined spline basis function.

10. A bearing remaining life prediction system based on a dual-scale spatiotemporal hybrid hypergraph, characterized in that, include: Extraction module: used to treat a single sample as a node in a graph data structure in the whole life vibration signal of a rolling bearing, and extract the time domain features and frequency domain features of the node in the horizontal and vertical directions respectively to obtain a feature matrix; The extraction module is used to extract the feature matrices of all nodes in the full-life vibration signal of a rolling bearing using a sliding window, thus obtaining the node feature matrix. The generation module is used to calculate the adjacency matrix of the semi-connected ordinary graph and the incidence matrix of the semi-connected hypergraph based on the KNN algorithm and the node feature matrix. It then traverses all nodes to generate the adjacency matrix of the fully connected ordinary graph and the incidence matrix of the fully connected hypergraph. The first computation module is used to construct a graph convolutional neural network and a hypergraph convolutional neural network. The adjacency matrix of the fully connected ordinary graph and the correlation matrix of the fully connected hypergraph are respectively input into the graph convolutional neural network and the hypergraph convolutional neural network for node feature update. The feature fusion module based on attention weights obtains the first fused high-order-low-order spatial features. The second calculation module is used to input the adjacency matrix of the semi-connected ordinary graph and the correlation matrix of the semi-connected hypergraph into the graph convolutional neural network and the hypergraph convolutional neural network respectively for node feature update, and to obtain the second fused high-order-low-order spatial features based on the attention weight feature fusion module. The third calculation module is used for the attention-weighted feature fusion module. It integrates the first fused high-order-low-order spatial features and the second fused high-order-low-order spatial features to obtain dual-scale fused high-order-low-order spatial features. After learning the temporal dependency of the dual-scale fused high-order-low-order spatial features through a BiLSTM network, it is input into the KAN prediction layer to obtain the remaining life prediction value of the rolling bearing.

Images