Bearing residual service life prediction method based on adaptive clustering graph convolution

Abstract

The invention relates to the technical field of bearing residual service life prediction, in particular to a bearing residual service life prediction method based on adaptive clustering graph convolution, and the method specifically comprises the steps: capturing vibration signal sample data from healthy to damaged of a bearing through a high-frequency vibration accelerometer; performing transformation processing on the collected vibration signal sample data to obtain time-frequency domain representation of sub-sample data; inputting the time-frequency domain representation of the sub-sample data into a bidirectional long-short-term memory network to obtain a time dependence feature; calculating two types of adjacent matrixes based on the time domain feature data, obtaining a final adaptive graph through control factor fusion, and introducing a node representation updating network to update node features in the adaptive graph so as to obtain spatial correlation information in the adaptive graph; and obtaining a final residual service life prediction result by using a clustering graph pooling method based on the current node features and the future node features. According to the invention, local information can be effectively aggregated, so that the prediction precision of the model is improved.

Description

Technical Field

[0001] This invention relates to the field of bearing remaining service life prediction technology, and specifically to a bearing remaining service life prediction method based on adaptive clustering graph convolution. Background Technology

[0002] With the continuous development of artificial intelligence, intelligent industrial production has become an inevitable trend. Against this backdrop, predictive and health management of mechanical equipment, as a key means to achieve intelligent manufacturing, has received increasing attention in recent years. Bearings, as one of the most common components in mechanical equipment, have important functions such as load transmission, friction reduction, and vibration reduction, and are widely used in various fields, such as rail transportation, wind power generation, and aerospace. However, in complex and variable operating environments, bearings are subjected to alternating stress, impact loads, changes in lubrication conditions, and fluctuations in ambient temperature over a long period, leading to gradual performance degradation and even sudden failures. Bearing failure not only causes direct economic losses but can also pose a serious threat to the stability and reliability of the entire production system. Therefore, accurately predicting the remaining service life of bearings is of great significance for developing maintenance plans in advance, avoiding sudden failures, and improving equipment reliability.

[0003] Existing methods for predicting the remaining service life of bearings based on graph neural networks often use predefined graph structures to extract spatial features and directly stack node features to learn the final representation, thus failing to fully extract and utilize feature information. Summary of the Invention

[0004] The purpose of this invention is to provide a bearing remaining service life prediction method based on adaptive clustering graph convolution. This invention achieves rapid reproduction and data acquisition of the bearing failure process through multi-condition simulation. Then, it uses the acquired data to predict the bearing's remaining service life. Specifically, it uses short-time Fourier transform to process the original vibration signal of the bearing to obtain time-frequency domain features, and uses a bidirectional long short-term memory network to extract the time dependence in the condition monitoring data. Based on these time features, an adaptive graph structure is constructed to dynamically track the bearing's degradation characteristics. Finally, a node-level clustering graph pooling module is used to extract relevant features while capturing comprehensive local information, resulting in the final bearing remaining service life prediction. This invention can effectively aggregate local information, thereby improving the model's prediction accuracy. It can comprehensively and effectively extract feature correlations from condition monitoring data, and its excellent nonlinear learning capability provides a better final bearing remaining service life prediction result.

[0005] To achieve the above objectives, this invention provides a method for predicting the remaining service life of bearings based on adaptive clustering graph convolution, comprising the following steps:

[0006] Step 1: Based on the independently developed bearing accelerated life degradation test bench, synchronously collect bearing full life cycle monitoring data, and capture vibration signals from the healthy to the damaged bearing through a high-frequency vibration accelerometer;

[0007] Step 2: First, using a non-overlapping sampling method, the collected sample data is sampled through a window of length h. Then, the collected vibration signal sample data is processed using a short-time Fourier transform to obtain the time-frequency domain representation of the sub-sample data.

[0008] Step 3: Input the obtained features into a bidirectional long short-term memory network to capture and extract long-term time dependencies in the features, and obtain time-dependent features;

[0009] Step 4: Calculate two types of adjacency matrices based on time-domain feature data, and obtain the final adaptive graph by fusing them through control factors. The two types of adjacency matrices include a stable matrix and an adaptive matrix.

[0010] Step 5: Introduce a node representation update network to update the node features in the adaptive graph. The node representation update network updates the node features by considering the node's domain information in order to obtain the spatial correlation information in the adaptive graph.

[0011] Step 6: Use a clustering graph pooling method based on current node features and future node features. By considering the association between current and future nodes simultaneously, more comprehensive correlation information is captured, and the pooled node features after clustering are obtained. Finally, the pooled node features are input into the fully connected layer to obtain the final remaining lifetime prediction result.

[0012] Optionally, in step 2, processing the collected vibration signal sample data using short-time Fourier transform includes the following steps:

[0013] Step 2.1: Perform sliding non-overlapping sampling on the acquired vibration signal sample data using a window of length h, where each window's data is represented as U. i =(u1,u2,...,u h )∈R h ,i=1,2,...,L, where L represents the total number of sampling windows;

[0014] Step 2.2: Perform a short-time Fourier transform on the vibration signal sample data collected in each window to obtain the time-frequency domain representation of the vibration signal sub-samples. The calculation method for performing the short-time Fourier transform on the vibration signal data collected in each window satisfies the following:

[0015]

[0016] Where ABS[·] represents taking the absolute value, P(t) k ,f n ) indicates at frequency f n and time t k The amplitude at χ * dΔD represents the conjugate of the window function, dΔD represents the time index, and rΔR represents the frequency index.

[0017] Optionally, in step 3, the obtained subsample time-frequency domain representation is input into a bidirectional long short-term memory network to capture and extract long-term time dependencies in the time-frequency domain representation to obtain time-dependent features, including the following steps:

[0018] Step 3.1: Introduce forget gate, input gate, and output gate control mechanisms to regulate the flow of subsample time-frequency domain representation data;

[0019] Step 3.2: Combine the current time-frequency domain representation of the subsample data with the cell state c from the previous time step. t-1 The hidden layer state h at the previous moment t-1 A bidirectional long short-term memory network is input, and through gating mechanisms such as the forget gate, input gate, and output gate, the information flow and state update of time-frequency domain features are regulated in the time dimension. The bidirectional long short-term memory network is represented by the following expression:

[0020]

[0021] Where σ(·) represents the sigmoid activation function, tanh(·) represents the tanh activation function, and c t-1 p represents the cellular memory state of the previous moment. t It is the input data at the current moment; c t and h t These represent the cell state and hidden layer state at the current processing moment after processing by the bidirectional long short-term memory network, respectively; f t Indicates the output of the forget gate at time t; i t The input gate output represents time t; o t The output gate at time t controls the output of the current cell state, W. c W f W i W o Let b represent the trainable weight parameters. f b i b o b c This represents the trainable bias term. Indicates multiplication calculation;

[0022] Optionally, in step 4, two types of adjacency matrices are calculated based on time-dependent feature data, and the final adaptive graph is obtained by fusing them through control factors. This specifically includes the following steps:

[0023] Step 4.1: Construct two types of adaptive adjacency matrices, one stable matrix and one adaptive matrix. The stable matrix is ​​obtained by calculating the cosine similarity of time-dependent features, and the adaptive matrix is ​​obtained by performing linear transformation and nonlinear activation on the time-dependent features twice. The purpose is to suppress the overfitting phenomenon that may be caused by the adaptive adjacency matrix, and at the same time, to obtain the correlation between features based on a fixed similarity. The cosine similarity calculation satisfies the following:

[0024] c s,s′ =(p s ·p s′ ) / (||p s ||·||p s′ ||), s=1,2,...,L;

[0025] Among them, c s,s′ p represents the element in the s-th row and s'-th column of the cosine similarity matrix. s p represents the s-th row of the window data s′ s' represents the s'th row of the window data, and L represents the total number of rows of the window data;

[0026] Step 4.2: Apply the function Y(·) from this step to the data c of the cosine similarity matrix. s,s′ Calculate the stability matrix A s ∈R L×L Data a in s,s′ The value of a s,s′ Describing the stability matrix A s The value of row s and column s in the middle is calculated in a way that satisfies:

[0027]

[0028] Where λ is a manually set threshold. When a s,s′ When a = 0, it means there is no edge connecting the two time nodes, i.e., they are unrelated; when a s,s′ When = 1, it means that the two time nodes are connected by an edge in the graph, that is, they are related;

[0029] Step 4.3: Apply the fitness matrix A a ∈R L×LThe system dynamically tracks the feature changes of long-term time dependencies in the time-frequency domain representation using a trainable parameter matrix, forming an optimal graph topology. The adaptive matrix, used for dynamic tracking, is expressed as follows: This matrix dynamically tracks the changes in input data using a trainable parameter matrix, forming an optimal graph topology. Compared to a stable matrix, its value can be any value from 0 to 1, preventing problems such as vanishing and exploding gradients during automatic graph generation. It can be expressed as follows:

[0030]

[0031] in, and Let represent a trainable matrix, Tanh represent the hyperbolic tangent activation function, and P[n] represent the nth graph in the time series data;

[0032] Step 4.4: Control factors are introduced to adjust the proportions of the stable matrix and the adaptive matrix, and then fused to obtain the final adaptive graph. The adaptive graph satisfies the following expression: The adaptive adjacency matrix is ​​composed of the two stable matrices and the adaptive matrix mentioned above. Simultaneously, two control factors are introduced to adjust the proportions of the two types of matrices. The final expression for the adaptive adjacency matrix is ​​as follows:

[0033] A sa =γ1·A a +γ2·A s

[0034] Where γ1 and γ2 represent control factors;

[0035] Optionally, in step 5, updating the node representation network by considering the node's domain information to update the node features in order to obtain spatial correlation information in the adaptive graph includes the following steps:

[0036] Specifically, after obtaining the adaptive adjacency matrix, a node representation update operation is performed on each adaptive graph to obtain the future node features:

[0037]

[0038] Where N represents the total number of subgraphs constructed in this sample, Z represents the number of convolutional layers, and D... Z [n] represents the updated node features at layer Z, and σ represents the non-linear activation function. and This represents the adjacency matrix and node features of the nth subgraph after convolution at layer Z. When Z = 1,

[0039] W represents a matrix whose parameters are trainable;

[0040] Optionally, in step 6, a clustering graph pooling method based on current node features and future node features is used to obtain the clustered node features and the final prediction result. The execution process involves introducing a clustering graph pooling method based on current node features and future node features. By considering both current and future information, more comprehensive relevance features are captured, and finally, these features are input into a fully connected layer to obtain the final prediction result. Specifically, the process includes the following steps:

[0041] Step 6.1: Generate a clustering matrix C based on node features and node correlations. Z [n], by introducing a trainable parameter matrix This mechanism allows the model to dynamically adjust the weights of features and correlations: where the generated clustering matrix satisfies:

[0042]

[0043] Where SoftMax(·) represents the softmax activation function, Let D represent the adjacency matrix of the (Z-1)th layer. Z+1 [n] represents the node features updated by the (Z+1)th convolutional layer, || represents the concatenation operation, and σ represents the non-linear activation function;

[0044] Step 6.2: Transform the clustering matrix to capture local association information of the graph. By transforming the clustering matrix, local association information of the graph can be captured, thereby obtaining the pooling node features and the pooling adjacency matrix, specifically satisfying the following formula:

[0045]

[0046] Where [n] represents the nth graph of the time series, C Z [n] represents the clustering matrix of the Z-th layer. and This represents the pooling node characteristics and pooling adjacency matrix of layer Z. When Z=1,

[0047] Step 6.3: Update the correlation of the captured node features to obtain the updated pooled node features. Flatten the updated pooled node feature data and input it into the fully connected layer to obtain the final remaining lifetime prediction result. The prediction result is expressed by the following formula:

[0048] y t =σ(W t Flatten(g) t ));

[0049] Among them, y t The result represents the prediction, σ represents the ReLU activation function, and W represents the prediction result.t Let g represent the trainable weight matrix. t This represents the updated feature data, and Flatten indicates flattened calculation.

[0050] In summary, due to the adoption of the above technical solutions, the present invention has the following beneficial technical effects:

[0051] (1) The bearing remaining service life prediction method of the present invention realizes rapid reproduction and data acquisition of bearing failure process through multi-condition simulation. Then, the acquired data is used to predict the bearing remaining service life. Specifically, the original vibration signal of the bearing is processed by short-time Fourier transform to obtain time-frequency domain features, and the time dependence in the state monitoring data is extracted by bidirectional long short-term memory network.

[0052] (2) An adaptive graph structure is constructed based on this time feature, which aims to dynamically track the degradation characteristics of the bearing. The adaptive graph combines the advantages of the two matrices and effectively tracks the degradation characteristics of the bearing.

[0053] (3) Using node-level clustering graph pooling convolution to extract relevant features while capturing comprehensive local information can effectively aggregate local information, thereby improving the prediction accuracy of the model. It can comprehensively and effectively extract feature correlations from the condition monitoring data. Its excellent nonlinear learning ability provides better and final bearing remaining service life prediction results. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is the prediction process based on adaptive clustering graph convolution proposed in this invention;

[0056] Figure 2 This is the model structure of the bidirectional long short-term memory network of the present invention;

[0057] Figure 3 This is a schematic diagram of the adaptive adjacency matrix construction process of the present invention;

[0058] Figure 4 This is a schematic diagram of the clustering graph pooling method of the present invention;

[0059] Figure 5 This is a schematic diagram of the stable matrix, the adaptive matrix, and the resulting adaptive adjacency matrix of the present invention;

[0060] Figure 6 This is a comparative schematic diagram of the predicted remaining service life of bearings in an embodiment of the present invention;

[0061] Figure 7 This is a comparison chart of the average absolute error of bearing remaining service life prediction in the embodiments of the present invention;

[0062] Figure 8 This is a comparison chart of the application scores for predicting the remaining service life of bearings in this embodiment of the invention;

[0063] Figure 9 This is a schematic diagram illustrating the variation of prediction performance with prediction factors under various operating conditions in embodiments of the present invention;

[0064] Figure 10 This is a schematic diagram illustrating how the model prediction performance changes with the number of clustering and pooling layers in an embodiment of the present invention. Detailed Implementation

[0065] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0066] According to one aspect of the present invention, a method for predicting the remaining service life of bearings based on adaptive clustering graph convolution is provided, comprising the following steps:

[0067] Step 1: Based on the independently developed bearing accelerated life degradation test bench, vibration signal sample data of the bearing throughout its entire life cycle from healthy to damaged is captured using a high-frequency vibration accelerometer. During the bearing accelerated life test of this invention, bearing vibration signals were collected under three different operating conditions. These vibration signals contain data information in both the horizontal and vertical directions. The sampling frequency was set to 25.6 kHz, the sampling duration was 1.28 s, and data was collected every minute. When the bearing's acceleration monitoring value exceeded 20g, it was considered to have reached a failure state, at which point data collection was stopped.

[0068] Step 2: Process the collected vibration signal sample data using short-time Fourier transform to obtain the time-frequency domain representation of the sub-sample data; see the overall flowchart of this invention. Figure 1 The short-time Fourier transform processing of the collected vibration signal sample data includes the following steps:

[0069] Step 2.1: Using a non-overlapping sampling method, slide sampling is performed on the acquired vibration signal sample data through a window of length h. The data of each window can be represented as U. i =(u1,u2,...,uh )∈R h ,i=1,2,...,L, where L represents the total number of sampling windows;

[0070] Step 2.2: Perform a short-time Fourier transform on the vibration signal sample data collected in each window to obtain the time-frequency domain representation of the vibration signal sub-samples. The calculation method for performing the short-time Fourier transform on the vibration signal data collected in each window satisfies the following:

[0071]

[0072] Where ABS[·] represents taking the absolute value, that is, obtaining the real number amplitude value in the frequency domain representation after processing, P(t k ,f n ) indicates that the vibration signal is at frequency f n and time t k The amplitude at χ * dΔD represents the conjugate of the window function, dΔD represents the time index, and rΔR represents the frequency index;

[0073] Step 3: Input the time-frequency domain representation of the subsample data into a bidirectional long short-term memory network (BiLSTM) to capture and extract long-term temporal dependencies in the time-frequency domain representation, thereby obtaining temporal dependency features. In this embodiment of the invention, capturing and extracting long-term temporal dependencies in the time-frequency domain representation is achieved by introducing a bidirectional long short-term memory network. The execution process includes the following steps:

[0074] Step 3.1: By introducing three control mechanisms—forget gate, input gate, and output gate—into the bidirectional long short-term memory network, the flow of time-frequency domain representation data of subsamples is regulated, thereby effectively maintaining and updating cell states when processing long-term sequence data, ensuring long-term information transmission and effective processing;

[0075] Step 3.2: Input the input data into the bidirectional long short-term memory network, and regulate the information flow and state update of the time-frequency domain features in the time dimension through the gating mechanism of forget gate, input gate and output gate;

[0076] Figure 2 This describes the internal structure of the bidirectional long short-term memory network of the present invention. The bidirectional long short-term memory network contains three gate structures that use the σ (sigmoid activation function): the forget gate, the input gate, and the output gate. The input data of the bidirectional long short-term memory network includes the cell state c from the previous time step. t-1 The hidden layer state h at the previous moment t-1 and the input data p at the current time.t (The input data is the time-frequency domain representation of the subsamples), and its output is the cell state c at the current moment. t and hidden layer state h t The bidirectional long short-term memory network is represented by the following expression:

[0077]

[0078] Where σ(·) represents the sigmoid activation function, tanh(·) represents the tanh activation function, and c t-1 p represents the cellular memory state of the previous moment. t This is the input data at the current moment. t and h t These represent the cell state and hidden layer state at the current processing moment after processing by the bidirectional long short-term memory network, respectively. Furthermore, f t The output of the forget gate at time t determines the amount of past information to discard; i t The input gate output at time t determines the degree of new information input; t The output gate at time t controls the output of the current cell state, W. f W i W c Let b represent the trainable weight parameters. f b i b o b c This represents the trainable bias term. This represents data multiplication computation; the bidirectional long short-term memory network (LSTM) captures both short-term and long-term features simultaneously through a linear transmission path of cell states, thereby obtaining long-term temporal dependencies. Specifically, this long-term temporal dependency information is achieved through the continuous transmission of cell states: the current cell state c... t Directly inherited from the previous moment c t-1 Furthermore, by dynamically adjusting the fusion ratio of new and old information through forget gate and input gate, key information in the time-frequency domain representation can be retained across a longer time step rather than being covered by short-term noise.

[0079] Step 4: Calculate two types of adjacency matrices based on time-domain feature data, and fuse them using control factors to obtain the final adaptive graph. The two types of adjacency matrices include a stable matrix and an adaptive matrix. In this embodiment of the invention, calculating the stable matrix and the adaptive matrix and fusing them using control factors specifically includes the following steps:

[0080] Step 4.1: Construct an adaptive adjacency matrix containing a stable matrix and an adaptation matrix. The stable matrix is ​​obtained by calculating the cosine similarity of the time-dependent features. The adaptation matrix is ​​obtained by performing at least two linear transformations and nonlinear activations on the time-dependent features with time information. First, the linear transformations and nonlinear activations are calculated sequentially, and then the output is processed again through linear transformations and nonlinear activations to obtain the adaptation matrix. The adaptive adjacency matrix constructed in this invention consists of two parts, referred to as the stable matrix and the adaptation matrix, respectively. Figure 3 This is the construction process of the adaptive adjacency matrix of the present invention. First, the original vibration data of the bearing is processed by short-time Fourier transform, and then further features are extracted through a bidirectional long short-term memory network to construct an adaptive graph, finally obtaining a dynamic graph structure with temporal order relationships. This adaptive mechanism allows the graph to dynamically track the degradation state of the bearing and automatically adjust the graph topology during training, thereby enhancing the accuracy and robustness of model predictions. The stabilization matrix is ​​obtained by calculating the cosine similarity of spatiotemporal features. Its purpose is to suppress the overfitting phenomenon that may be caused by the adaptive adjacency matrix, and at the same time, to obtain the correlation between features based on a fixed similarity. The cosine similarity calculation satisfies the following:

[0081] c s,s′ =(p s ·p s′ ) / (||p s ||·||p s′ ||), s=1,2,...,L;

[0082] Among them, c s,s′ This represents the cosine similarity matrix at row s and s-th position. ’ The data in column p s The s-th row of the window data is represented by L, where L represents the total number of rows in the window.

[0083] Step 4.2: Apply the judgment function Y(·) to the elements c of the cosine similarity matrix. s,s′ The stability matrix A can be obtained. s ∈R L×L element a in s,s′ The value of a, where R represents the real number field, L×L represents the data dimension, and a s,s′ Describing the stability matrix A s The value in row s and column s' of the middle:

[0084]

[0085] Where λ is a manually set threshold, when a s,s′ When a = 0, it indicates that there is no edge connecting the two time nodes, meaning they are uncorrelated (the node feature values ​​of the two nodes are uncorrelated); when a s,s′When = 1, it means that the two time nodes are connected by an edge in the graph, that is, they are related;

[0086] Step 4.3: Apply the fitness matrix A a ∈R L×L The system dynamically tracks the feature changes of long-term time dependencies in the time-frequency domain representation using a trainable parameter matrix, and forms an optimal graph topology; the fitness matrix A a ∈R L×L The fitness matrix is ​​obtained by performing two linear transformations and nonlinear activations on the temporal feature data of the input time information. This matrix dynamically tracks the changes in the input data through a trainable parameter matrix, thereby forming the optimal graph topology. Compared to the stable matrix, the fitness matrix can have any value from 0 to 1, preventing problems such as vanishing and exploding gradients during automatic graph generation. The dynamic tracking of the fitness matrix is ​​expressed using the following mathematical formula:

[0087]

[0088] in, and Let represent a trainable matrix, Tanh represent the hyperbolic tangent activation function, and P[n] represent the nth graph in the time series data;

[0089] Step 4.4: Control factors are introduced to adjust the proportions of the stable matrix and the adaptive matrix before fusion to obtain the final adaptive graph. The adaptive adjacency matrix is ​​composed of the two matrices mentioned above, namely, the stable matrix and the adaptive matrix. Two control factors are introduced to adjust the proportions of the two types of matrices. The expression for the final adaptive adjacency matrix is ​​as follows:

[0090] A sa =γ1·A a +γ2·A s

[0091] Where γ1 and γ2 represent control factors.

[0092] S5: Introduce a node representation update network to update the node features in the adaptive graph. The node representation update network updates the node features by considering the node's neighborhood information to obtain the spatial correlation information in the adaptive graph. After obtaining the adaptive adjacency matrix, the adaptive graph (which includes the adjacency matrix and graph nodes) is constructed. Perform the following node representation update operation on each adaptive graph to obtain the future node features.

[0093]

[0094] Where N represents the total number of subgraphs constructed in this sample, Z represents the number of convolutional layers, and D... Z [n] represents the updated node features at layer Z, and σ represents the non-linear activation function. and This represents the adjacency matrix and node features of the nth subgraph after convolution at layer Z. When Z = 1, W represents a matrix whose parameters are trainable;

[0095] S6: After updating the node features, a clustering graph pooling method based on current node features and future node features is used. By considering the current and future node associations simultaneously, more comprehensive correlation information is captured, and clustered pooled node features are obtained. Finally, the pooled node features are input into the fully connected layer to obtain the final remaining lifetime prediction result. To this end, this invention updates the node features again on another level based on the node feature update. The current node features are the node features updated in step 5, and the future node features are the features after a pre-update of the node features updated in step 5.

[0096] like Figure 4 As shown, Figure 4 This is a schematic diagram of the clustering graph pooling method of the present invention. This method is a clustering graph pooling method based on current node features and future node features. By simultaneously considering current and future information, it captures more comprehensive relevance features to obtain the final prediction result, including the following steps:

[0097] Step 6.1: Based on the current node's feature D Z+1 [n] and node correlation (i.e., adjacency matrix) Calculate and generate clustering matrix C Z [n], and then introduce a trainable parameter matrix. The weights of node features and node correlations are dynamically adjusted, wherein the generated clustering matrix satisfies the following:

[0098]

[0099] Where SoftMax(·) represents the softmax activation function. D represents the adjacency matrix of the (Z-1)th layer (Z represents the number of layers in the adaptive clustering graph convolution), and D Z+1 [n] represents the node features updated by the (Z+1)th convolutional layer, || represents the concatenation operation, and σ represents the non-linear activation function;

[0100] Step 6.2: Transform the clustering matrix to capture local association information of the graph. By transforming the clustering matrix, local association information of the graph can be captured, thereby obtaining the features of pooling nodes. Pooling adjacency matrix Specifically, it satisfies the following formula:

[0101]

[0102] Where [n] represents the nth graph of the time series, C Z [n] represents the clustering matrix of the Z-th layer. and This represents the pooling node characteristics and pooling adjacency matrix of layer Z. When Z=1,

[0103] Step 6.3: Update the correlation of captured node features, obtain the updated node features, flatten the updated node feature data and input it into the fully connected layer; after capturing the correlation through the above steps, flatten the updated feature data and input it into the fully connected layer to obtain the final remaining lifetime prediction result, which is expressed by the following formula:

[0104] y t =σ(W t Flatten(g) t ));

[0105] Among them, y t The result represents the prediction, σ represents the ReLU activation function, and W represents the prediction result. t Let g represent the trainable weight matrix. t This represents the updated feature data. Flatten indicates flattening computation, which flattens the two-dimensional vector of the flattened layer into one-dimensional vector data.

[0106] The following is a further explanation with reference to specific embodiments:

[0107] Vibration signals throughout the entire lifecycle of bearings were collected using a constructed bearing test bench. This invention selected six bearings under three operating conditions: Condition 1 (2100 rpm / 12000 N), bearings 1-1, 1-2, and 1-4; Condition 2 (2250 rpm / 11000 N), bearings 2-2 and 2-4; and Condition 3 (2400 rpm / 10000 N), bearing 3. The acquired data was then used to predict the remaining service life of the bearings. Specifically, the original vibration signals of the bearings were processed using short-time Fourier transform to obtain time-frequency domain features, and a bidirectional long short-term memory network was used to extract the time dependence in the state monitoring data, obtaining time-dependent features. Then, an adaptive graph structure was constructed based on these time-dependent features to dynamically track the degradation characteristics of the bearings. Finally, a node-level clustering graph pooling module was used to extract correlation features while capturing comprehensive local information, resulting in the final predicted remaining service life of the bearings.

[0108] Figure 5This is a schematic diagram of the stable matrix, the adaptive matrix, and the resulting adaptive adjacency matrix provided by this invention. Figure 5 In the diagram, (a) represents the stable matrix, (b) represents the fitness matrix, and (c) represents the fused adaptive adjacency matrix. It can be seen that compared to the stable matrix, the fitness matrix exhibits greater sparsity, indicating that it can autonomously optimize its parameters to focus on more important features. Furthermore, the fitness matrix does not exhibit symmetry, which facilitates unbalanced aggregation operations between graph nodes, thereby capturing deeper features. Conversely, these characteristics of the fitness matrix make it more prone to overfitting, resulting in low prediction performance on the test set. The adaptive adjacency matrix effectively combines the advantages of both types of matrices, allowing the model to dynamically adjust the graph structure while minimizing overfitting, enhancing the model's generalization performance, and improving prediction accuracy. It is worth noting that the graphs shown are not mathematically related; they were selected simply for ease of observation.

[0109] In this invention, such as Figure 6 As shown, Figure 6 This diagram illustrates a comparison of bearing remaining service life prediction results provided by this invention. It includes bearings 1_1, 1_2, 1_4, 2_2, 2_4, and 3_3. The horizontal axis represents the bearing's operating time, and the vertical axis represents its remaining service life. The legend in the upper right corner illustrates other advanced prediction methods compared, including GCN, ChebGCN-LSTM, T-GCN, DC-DGCN, and SAGCN-SA. The results obtained using the proposed adaptive clustering graph convolutional network for bearing remaining service life prediction show that the proposed model achieves the best prediction performance, outperforming other graph neural network-based methods. The mean square error (RMSE) of the adaptive clustering graph convolutional network is 0.100, while the RMSE of other comparative models ranges from 0.114 to 0.165. Compared to the best-performing prediction model DC-DGCN, its RMSE is reduced by 12.3%. Compared to the SAGCN-SA model with an attention mechanism, its RMSE is reduced by 16%. Therefore, graph convolution based on clustering graph pooling operations has a more comprehensive feature extraction capability and can capture more refined degradation features. Meanwhile, the addition of an adaptive mechanism makes the proposed model more flexible during training, allowing it to dynamically adjust the graph structure to adapt to optimal parameters and achieve better predictive performance. Furthermore, it can be observed that models using only graph convolutional networks for feature extraction do not achieve satisfactory predictive results. This is because extracting only relevance features misses many time-dependent features, making the model more prone to overfitting and thus reducing its predictive performance on the test set.

[0110] Figure 7 and Figure 8This is a comparison chart of the mean absolute error (MAE) and application score for bearing remaining service life prediction provided by this invention. The horizontal axis represents different bearings, and the vertical axis represents the MAE value. The legend in the upper left corner shows the different advanced prediction models being compared, including GCN, ChebGCN-LSTM, T-GCN, DC-DGCN, and SAGCN-SA. It can be seen that the proposed prediction method achieves the optimal MAE value in most prediction tasks. For individual bearings, such as bearings 2_4 and 3_3, the performance is very close to the optimal prediction model, indicating that the adaptive clustering graph convolution has higher prediction accuracy and stronger applicability. Furthermore, in the application score index, the adaptive clustering graph convolution model shows a significant advantage, achieving the lowest application score in all prediction tasks. This indicates that the proposed model has higher prediction accuracy in the later stages of bearing degradation and is more suitable for application in real-world scenarios.

[0111] Figure 9 This is a schematic diagram illustrating the change in prediction performance with the predictor factor under various operating conditions provided by this invention. The horizontal axis represents the value of γ1, and the vertical axis represents the value of the root mean square error. The three different broken lines represent the prediction performance under different operating conditions. The magnitude of the control factor plays a crucial role in the allocation of matrix weights, thus affecting the model's prediction performance. To determine the optimal control factor value for the model's performance, the prediction performance of adaptive clustering graph convolution under various combinations of control factors was analyzed experimentally. This figure shows the prediction effect of the model's prediction performance with the control factor γ1 under three different operating conditions. Since γ1 + γ2 = 1, only the value of γ1 is shown in the figure. It can be seen that under all three operating conditions, the model's prediction performance increases with the increase of γ1. After reaching the optimal performance, further increasing γ1 will lead to a decrease in the model's prediction performance. This is because when γ1 is small, the model has to consider the predefined graph structure more, making it difficult to dynamically track degenerate features, resulting in poor prediction performance. Conversely, when γ1 is too large, the model will fall into a certain degree of overfitting due to the excessive proportion of the adaptive matrix, leading to a decrease in the prediction performance on the test set. Furthermore, it can be observed that condition 3 achieves the best predictive performance at a larger γ1 compared to conditions 1 and 2, indicating that the degradation characteristics of condition 3 are more complex and require greater utilization of the dynamic adjustment capability of the adaptive matrix.

[0112] Figure 10This diagram illustrates the variation of the model's prediction performance with the number of clustering pooling layers provided by this invention. The horizontal axis represents the number of clustering pooling layers, the left vertical axis represents training time, and the right vertical axis represents the values ​​of the mean square root error and mean absolute error. The bar chart represents the training time, and the line chart represents the prediction performance. When the number of clustering pooling layers is 3, the model achieves the best prediction accuracy. Although the training time is slightly longer than with 2 clustering pooling layers, the prediction performance is significantly improved. Therefore, this paper uses 3 clustering pooling layers. Furthermore, the trend in the diagram shows that when the number of clustering pooling layers is low, the model's prediction accuracy is poor. This is because fewer layers cannot extract deep, relevant features, resulting in poor prediction performance. When the number of clustering pooling layers continues to increase after 3, the model's prediction accuracy continuously decreases, almost losing its predictive ability at 6 layers. This is because excessive graph convolutional layers cause nodes to exhibit extremely similar characteristics after multiple aggregations of domain information, leading to prediction failure.

[0113] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for predicting the remaining service life of bearings based on adaptive clustering graph convolution, characterized in that, The prediction method includes the following steps: Step 1: Based on the self-developed bearing accelerated life degradation test bench, high-frequency vibration accelerometers are used to capture vibration signal sample data of the bearing from healthy to damaged. Step 2: Process the collected vibration signal sample data using short-time Fourier transform to obtain the time-frequency domain representation of the sub-sample data; Step 3: Input the time-frequency domain representation of the subsample data into the bidirectional long short-term memory network, and extract the long-term time dependencies in the time-frequency domain representation to obtain time-dependent features; Step 4: Calculate two types of adjacency matrices based on time-domain feature data, and obtain the final adaptive graph by fusing them through control factors. The two types of adjacency matrices include a stable matrix and an adaptive matrix. Step 5: Introduce a node representation update network to update the node features in the adaptive graph. The node representation update network updates the node features by considering the node's domain information in order to obtain the spatial correlation information in the adaptive graph. Step 6: Use a clustering graph pooling method based on current node features and future node features. By considering the association between current and future nodes simultaneously, more comprehensive correlation information is captured, and the pooled node features after clustering are obtained. Finally, the pooled node features are input into the fully connected layer to obtain the final remaining lifetime prediction result.

2. The bearing remaining service life prediction method based on adaptive clustering graph convolution as described in claim 1, characterized in that, Step 2 involves processing the collected vibration signal sample data using short-time Fourier transform, including the following steps: Step 2.1: Perform sliding non-overlapping sampling on the acquired vibration signal sample data using a window of length h, where each window's data is represented as U. i =(u1,u2,...,u h )∈R h ,i=1,2,...,L, where L represents the total number of sampling windows; Step 2.2: Perform a short-time Fourier transform on the vibration signal sample data collected in each window to obtain the time-frequency domain representation of the vibration signal sub-samples. The calculation method for performing the short-time Fourier transform on the vibration signal data collected in each window satisfies the following: Where ABS[·] represents taking the absolute value, P(t) k ,f n ) indicates at frequency f n and time t k The amplitude at χ * dΔD represents the conjugate of the window function, dΔD represents the time index, and rΔR represents the frequency index.

3. The bearing remaining service life prediction method based on adaptive clustering graph convolution as described in claim 1, characterized in that, In step 3, the obtained subsample time-frequency domain representation is input into a bidirectional long short-term memory network to capture and extract long-term time dependencies in the time-frequency domain representation, thereby obtaining time-dependent features. This includes the following steps: Step 3.1: Introduce forget gate, input gate, and output gate control mechanisms to regulate the flow of subsample time-frequency domain representation data; Step 3.2: Combine the current time-frequency domain representation of the subsample data with the cell state c from the previous time step. t-1 The hidden layer state h at the previous moment t-1 A bidirectional long short-term memory network is input, and through gating mechanisms such as the forget gate, input gate, and output gate, the information flow and state update of time-frequency domain features are regulated in the time dimension. The bidirectional long short-term memory network is represented by the following expression: Where σ(·) represents the sigmoid activation function, tanh(·) represents the tanh activation function, and c t-1 p represents the cellular memory state of the previous moment. t It is the input data at the current moment; c t and h t These represent the cell state and hidden layer state at the current processing moment after processing by the bidirectional long short-term memory network, respectively; f t Indicates the output of the forget gate at time t; i t The input gate output represents time t; o t The output gate at time t controls the output of the current cell state, W. c W f W i W o Let b represent the trainable weight parameters. f b i b o b c This represents the trainable bias term. This indicates multiplication calculation.

4. The bearing remaining service life prediction method based on adaptive clustering graph convolution as described in claim 1, characterized in that, In step 4, two types of adjacency matrices are calculated based on time-dependent feature data, and the final adaptive graph is obtained by fusing them through control factors. Specifically, the steps include: Step 4.1: Construct two types of adaptive adjacency matrices, one containing a stable matrix and the other a fitness matrix. The stable matrix is ​​obtained by calculating the cosine similarity of the time-dependent features, and the fitness matrix is ​​obtained by performing a linear transformation and a nonlinear activation on the time-dependent features twice. The cosine similarity calculation satisfies the following: c s,s′ =(p s ·p s′ ) / (||p s ||·||p s′ ||),s=1,2,...,L; Among them, c s,s′ p represents the element in the s-th row and s'-th column of the cosine similarity matrix. s p represents the s-th row of the window data s′ s' represents the s'th row of the window data, and L represents the total number of rows of the window data; Step 4.2: Use the judgment function Y(·) and the elements c of the cosine similarity matrix s,s′ Calculate the stability matrix A s ∈R L×L element a in s,s′ The value of is calculated in a manner that satisfies: Where λ is the set threshold, representing p s and p s′ The two time points are connected by an edge; Step 4.3: Apply the fitness matrix A a ∈R L×L The system dynamically tracks the feature changes of long-term time dependencies in the time-frequency domain representation using a trainable parameter matrix, and forms an optimal graph topology. The following mathematical expression is used for dynamic tracking of the adaptation matrix: in, and Let represent a trainable matrix, Tanh represent the hyperbolic tangent activation function, and P[n] represent the nth graph in the time series data; Step 4.4: A control factor is introduced to adjust the proportions of the stable matrix and the adaptive matrix, and then the results are fused to obtain the final adaptive graph. The adaptive graph satisfies the following expression: A sa =γ1·A a +γ2·A s ; γ1 and γ2 represent control factors.

5. The bearing remaining service life prediction method based on adaptive clustering graph convolution as described in claim 1, characterized in that, In step 5, the node representation update network updates node features by considering the node's domain information to obtain spatial relevance information in the adaptive graph, including: Perform a node representation update operation on each adaptive graph to obtain the features of future nodes: Where N represents the total number of subgraphs constructed in the samples, Z represents the number of convolutional layers, and D... Z [n] represents the updated node features at layer Z, and σ represents the non-linear activation function. and Let W represent the adjacency matrix and node features of the nth subgraph after convolution at layer Z; W represents the trainable matrix of parameters.

6. The bearing remaining service life prediction method based on adaptive clustering graph convolution as described in claim 1, characterized in that, In step 6, a clustering graph pooling method based on current node features and future node features is used to obtain the clustered node features and the final prediction result, including the following steps: Step 6.1: Based on the current node's feature D Z+1 [n] and node correlation Calculate and generate clustering matrix C Z [n], and then introduce a trainable parameter matrix. The weights of node features and node correlations are dynamically adjusted, and the generated clustering matrix satisfies: Where SoftMax(·) represents the softmax activation function. Let D represent the adjacency matrix of the (Z-1)th layer. Z+1 [n] represents the node features updated by the (Z+1)th convolutional layer, || represents the concatenation operation, and σ represents the non-linear activation function; Step 6.2: Transform the clustering matrix to capture local correlation information of the graph, thereby obtaining the pooling node features. Pooling adjacency matrix Specifically, it satisfies the following formula: Where [n] represents the nth graph of the time series, C Z [n] represents the clustering matrix of the Z-th layer. and This represents the pooling node characteristics and pooling adjacency matrix of the Z-th layer; Step 6.3: Update the correlation of the captured node features to obtain the updated pooled node features. Flatten the updated pooled node feature data and input it into the fully connected layer to obtain the final remaining lifetime prediction result. The prediction result is expressed by the following formula: y t =σ(W t ·Flatten(g t )); Among them, y t The result represents the prediction, σ represents the ReLU activation function, and W represents the prediction result. t Let g represent the trainable weight matrix. t This represents the updated feature data, and Flatten indicates flattened calculation.

Images