Bearing remaining service life prediction method based on adaptive gating chart convolutional network
Through the adaptive gated graph convolution network, the time and space correlation of bearing feature channels is solved, and the problem of edge information loss and excessive smoothing in the prior art is improved, and the accuracy and robustness of bearing residual service life prediction is improved. Suitable for predictive maintenance.
Patent Information
- Application Number
- CN202411062580.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-11-22
AI Technical Summary
There are problems in the existing bearing residual service life prediction method based on deep learning, which leads to low accuracy of model prediction.
Adaptive gating graph convolution network is adopted, and by building a multi-layer gating graph convolution module with an adaptive adjacency matrix, combining deep confidence network and bidirectional LSTM, the temporal correlation and spatial correlation of the fused feature channel are avoided from being too smooth. Problems and improve model prediction accuracy.
It significantly improves the accuracy of bearing residual service life prediction, enhances the generalization ability and robustness of the model, and can more accurately capture bearing degradation characteristics, suitable for predictive maintenance.
Smart Images

Figure CN119004970A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bearing remaining service life prediction, and in particular to a bearing remaining service life prediction method using an adaptive gated graph convolutional network. Background Art
[0002] Mechanical equipment in modern industry usually needs to withstand high loads and complex working conditions. Bearings, as an important component of mechanical components, undertake the key task of supporting and transmitting loads, and have an important impact on the performance and reliability of equipment. Therefore, bearings are widely used in various industrial fields, including machinery manufacturing, aerospace, energy generation and transportation, rail transit trains, etc. It is necessary to accurately infer the health status of bearings and predict the remaining life of bearings, and perform predictive maintenance based on this, in order to minimize the risk of failure of mechanical equipment and improve its reliability. Therefore, it is of great research significance to predict the remaining service life of bearings.
[0003] The existing methods for predicting the remaining service life of bearings based on deep learning have problems such as edge information loss and over-smoothing, which leads to low model prediction accuracy. Summary of the invention
[0004] The purpose of the present invention is to provide a bearing remaining service life prediction method based on an adaptive gated graph convolutional network, which takes into account the time correlation and spatial correlation of feature channel data, and integrates them with long-term and short-term time dependencies to obtain the best feature representation. By constructing a multi-layer gated graph convolution module with an adaptive adjacency matrix, the over-smoothing problem caused by stacking multi-layer graph convolutions is avoided, and the adaptive adjacency matrix can autonomously train parameters, thereby improving the model prediction accuracy.
[0005] To achieve the above object, the present invention provides a method for predicting the remaining service life of a bearing using an adaptive gated graph convolutional network, comprising the following steps:
[0006] Step 1: Collect vibration data during the operation of the bearing and convert it into a feature channel, wherein the vibration data includes normal state data and abnormal state data of the bearing;
[0007] Step 2: Sample the converted feature channel data through the window and divide it into rows and columns to obtain graph structure data of time correlation and spatial correlation;
[0008] Step 3: Use cosine similarity and covariance to obtain a predefined spatiotemporal correlation graph, and use the source matrix and target matrix to obtain two types of adaptive graphs;
[0009] Step 4: Input the obtained predefined graph and the graph with adaptive mechanism into a three-layer gated graph convolutional neural network to extract two types of correlations of bearing feature channels, and use a deep belief network to extract its low-dimensional features;
[0010] Step 5: The low-dimensional features extracted by DBN are fused with the long-term and short-term time dependencies extracted by the bidirectional LSTM and input into the final fully connected layer to obtain the final RUL prediction.
[0011] Optionally, the process of obtaining the graph structure data of time correlation and space correlation in step 2 includes the following steps:
[0012] Step 2.1: Extract the features of the original vibration data of the bearing, and then sample the extracted features using a moving window with a stride of 1 and a length of L;
[0013] Step 2.2: Use the columns of the window data as graph node vectors to build a spatial correlation graph, and use the rows of the window data as graph node vectors to build a temporal correlation graph.
[0014] Optionally, the execution process of step 3 includes the following steps:
[0015] Step 3.1: Use the covariance matrix V (i,n) ∈R S*S And the cosine similarity matrix C (i,n) ∈R T*T To calculate the correlation between the spatial graph and the time graph, the covariance matrix V (i,n) The expression is as follows:
[0016]
[0017] in is the matrix V (i,n) Elements in represents the Sth column of the nth sampling data of the ith bearing; E[·] is the expectation; S represents the number of extracted features.
[0018] Step 3.2: According to the matrix elements To obtain the adjacency matrix of the spatial graph Elements in
[0019] if If is positive, it means that the change trends of the two features are consistent. Set to 1, that is, in the spatial correlation graph G of the feature s There is an edge between s and s′; if If it is negative, it means that the change trends of the two features are different. Set to 0, that is, in the spatial correlation graph G of the feature sThere is no edge connection between s and s′;
[0020] Step 3.3: Preliminary definition of the temporal correlation graph G by the cosine similarity matrix t , cosine similarity C (i,n) Elements in The expression is as follows:
[0021]
[0022] in represents the t-th row feature of the n-th sampling data of the ith bearing, ||·|| is the quadratic norm of the vector; T represents the size of the time step.
[0023] Step 3.4: Apply the function o(·) to the cosine similarity matrix C (i,n) The elements in the adjacency matrix can be obtained Elements in δ is a threshold defined artificially, and the function o(·) is expressed as:
[0024]
[0025] Indicates that there is no edge connection between two time nodes in the time-related graph, which means that they are unrelated; Indicates that two time nodes are connected by an edge in the time-dependent graph, which means that they are related in time;
[0026] Step 3.5: Define two matrices and name them source matrices W r and the target matrix W g , firstly, W r and W g Multiply to get the adaptive feature space dependency, then introduce a nonlinear activation function Tanh to eliminate the residual secondary dependency, and finally get the adaptive adjacency matrix A through the softmax function adp , the expression is as follows:
[0027]
[0028] in is the learnable adaptive parameter matrix, It represents the adaptive spatial dependency after the hyperbolic activation function, and finally the 0-1 normalized adaptive adjacency matrix is generated by the softmax function;
[0029] Step 3.6: Calculate the adaptive adjacency matrix of time correlation in the same way as step 3.5 The expression is as follows:
[0030]
[0031] in is a learnable adaptive parameter matrix;
[0032] Step 3.7: Define the space-time Laplacian matrix, expressed as follows:
[0033]
[0034] Among them I s ,I t is the identity matrix, D s , D t is a diagonal matrix, and the values on its diagonal are the degrees of each node in the spatial graph and the temporal graph respectively;
[0035] Step 3.8: Orthogonal decomposition to obtain the eigenvectors and eigenvalues of the Laplacian matrix
[0036]
[0037] Among them, Λ=diag(λ0,λ1,λ2,...,λ X ) is the eigenvalue matrix, Q = [α0, α1, α2, ... α x ] is the eigenvector corresponding to the eigenvalue;
[0038] Step 3.9: Perform spectral convolution operation,
[0039]
[0040] Where s(·) represents the spectral graph convolution operation, U represents the graph data sampled with the window, and g s , g t Graphic filters representing spatial and temporal graphs, respectively;
[0041] Substituting into the equation, the spectral convolution equation in the form of a Laplace matrix is as follows:
[0042]
[0043] Step 3.10: Order This gives the scaled normalized Laplacian matrix
[0044]
[0045] Step 3.11: Introduce the first-order Chebyshev polynomial to transform the expression as follows:
[0046]
[0047] Where p is the Chebyshev order of the approximation, and p=1 is taken here; θ p(s)is the Chebyshev approximation vector, is the p-order Chebyshev approximation polynomial;
[0048] Step 3.12: Finally, the output of the adaptive Chebyshev graph convolution can be obtained by adding the weight matrix, which is expressed as follows:
[0049]
[0050] Optionally, the execution process of step 4 is to extract two types of data correlations through three layers of gated graph convolution layers and perform dimensionality reduction through DBN. Each layer of gated graph convolution layer is composed of two types of convolution and gating mechanisms. The first type of convolution is the self-attention graph convolution in which the predefined graph participates, and the second type of convolution is the spectral graph convolution in which the adaptive graph participates, which specifically includes the following steps:
[0051] Step 4.1: Input the two predefined graphs into the graph attention network with multi-head attention to further obtain more accurate edge weights, the output of which can be obtained by the following formula:
[0052]
[0053] in, A s and A t is the two types of correlation adjacency matrix adjusted by the attention mechanism, ⊙ is the element multiplication, M is the dynamic attention coefficient, so is a dynamic adjacency matrix with attention mechanism, Z L is the output of the multi-head attention mechanism at layer L. If L = 0, then set Z L =U (i,n) .
[0054] Step 4.2: In the graph data U (i,n) After the multi-head graph attention network and the adaptive Chebyshev graph convolution layer extract the correlation, a gate value β is used to control its update and forgetting data. The operation expression is as follows:
[0055]
[0056] Step 4.3: In the gated graph convolutional neural network module, β is the value of the adaptive graph convolutional network output obtained through a sigmoid function, and the corresponding update mechanism expression is as follows:
[0057]
[0058] Among them, 1-β represents the amount of information forgotten by the previous layer of gated graph convolutional neural network, and β represents the amount of information updated by the current layer of gated graph convolutional neural network; σ(·) is the ReLU activation function. After extracting the correlation through several layers of adaptive gated convolution, each node is updated to obtain a more reasonable high-dimensional representation;
[0059] Step 4.4: Output the correlation information extracted from the gated graph convolutional neural network module into a DBN network with two hidden layers to achieve the extraction of low-dimensional abstract features.
[0060] Optionally, the execution process of step 5 includes the following steps:
[0061] Step 5.1: Input the feature channel data into the bidirectional LSTM to extract the time dependency. The update mechanism of the bidirectional LSTM is as follows:
[0062]
[0063] Where σ(·) is the activation function, W and b are the weight matrices and bias parameters of each gate, i t 、f t , o t are the input gate, forget gate, and output gate at the tth position, respectively, c t is the cell state at the tth position, h t is the hidden layer information at the tth position;
[0064] Step 5.2: In the self-attention Bi-LSTM, the original data is first projected into a higher-dimensional space to reduce the overlap of information and create information in subspaces under different representations. The expression is as follows:
[0065] Q=U (i,n) W Q , K=U (i,n) W K , V=U (i,n) W V
[0066] Among them, Q, K, and V represent the time step data U (i,n) The query subspace, key subspace, and value subspace obtained, W Q , W K , W V are their corresponding weight matrices respectively;
[0067] Step 5.3: Scale the obtained multiple subspaces to obtain the attention output, which is expressed as follows:
[0068]
[0069] Among them, d kIndicates the dimension of the input data.
[0070] Step 5.4: Concatenate the attention outputs of multiple heads to get the final multi-head attention output, which is expressed as follows:
[0071]
[0072] Where H represents the number of attention heads and U represents the serial operation;
[0073] Step 5.5: In the final spatiotemporal fusion layer, the two types of features are flattened and input into the fully connected layer for final prediction, which is expressed as follows:
[0074]
[0075] Where W is the weight matrix in the FC layer, b is the bias vector in the FC layer, and σ(·) is the activation function.
[0076] The present invention provides a method for predicting the remaining service life of a bearing using an adaptive gated graph convolutional network, which specifically includes sampling the original vibration signal of the bearing and converting it into a feature channel; sampling the data of the feature channel through a window, and dividing it into rows and columns respectively to obtain graph structure data of time correlation and space correlation, and using cosine similarity and covariance to preliminarily define the adjacency matrix of the two types of graphs for input into the subsequent graph neural network; using the source matrix and the target matrix to obtain the adaptive two types of graphs and combine them with the predefined Figure 1 The gated graph convolutional network constructed with the same input; the correlation features extracted from the gated graph convolutional network are input into the deep belief network (DBN) to extract its low-dimensional abstract features; the long-term and short-term time dependencies in the feature channel data are extracted through the bidirectional LSTM and fused with the output of the DBN, and finally input into a fully connected layer to obtain the final remaining bearing service life prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0078] Figure 1 It is a schematic flow chart of the steps of a method for predicting the remaining useful life of a bearing using an adaptive gated graph convolutional network of the present invention.
[0079] Figure 2 It is a schematic diagram of the process of predefining two types of correlation graphs of the present invention.
[0080] Figure 3 It is a schematic diagram of the spectral graph convolution framework with an adaptive adjacency matrix of the present invention.
[0081] Figure 4 It is a schematic diagram of the structure of the A-GG gating unit provided by the present invention.
[0082] Figure 5 It is a schematic diagram of the LSTM unit structure provided by the present invention.
[0083] Figure 6 It is a schematic diagram of the basic structure of the DBN model provided by the present invention.
[0084] Figure 7 It is a schematic diagram comparing two different types of adjacency matrix graphs of the present invention.
[0085] Figure 8 It is the RUL prediction result of some bearings in the PHM2012 bearing data set in the embodiment of the present invention.
[0086] Fig. 9 It is a schematic diagram comparing the prediction effects of various models of the PHM2012 bearing data set in an embodiment of the present invention.
[0087] Fig.10 It is a partial bearing prediction diagram of the XJTU-SY data set in an embodiment of the present invention.
[0088] Fig.11 It is a comparison chart of the prediction effects of various models of the XJTU-SY data set in an embodiment of the present invention.
[0089] Fig.12 It is a schematic diagram of the quantification and visualization of the performance of each model in the ablation experiment in the embodiment of the present invention.
[0090] Fig.13 It is a schematic diagram of the model prediction effect under different cosine similarity threshold values in an embodiment of the present invention.
[0091] Fig.14 It is a schematic diagram of the training error and training time of the model under different A-GG layers in an embodiment of the present invention. DETAILED DESCRIPTION
[0092] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.
[0093] The following is an annotation of the corresponding English abbreviations for the convenience of subsequent description:
[0094] RUL: Remaining useful life;
[0095] DBN: Deep Belief Network;
[0096] A-GG: Adaptive Gated Graph Convolutional Neural Network;
[0097] The present invention provides a method for predicting the remaining service life of a bearing using an adaptive gated graph convolutional network, comprising the following steps:
[0098] S1: collecting vibration data during the operation of the bearing and converting the data into a feature channel, wherein the vibration data includes normal state data and abnormal state data of the bearing;
[0099] S2: The converted feature channel data is sampled through a window and divided into rows and columns to obtain graph structure data of time correlation and spatial correlation;
[0100] S3: Use cosine similarity and covariance to obtain the predefined spatiotemporal correlation graph, and use the source matrix and target matrix to obtain the adaptive two-type graph;
[0101] S4: The obtained predefined graph and the graph with adaptive mechanism are input into a three-layer gated graph convolutional neural network to extract two types of correlations of bearing feature channels, and a deep belief network is used to extract its low-dimensional features;
[0102] S5: The low-dimensional features extracted by DBN are fused with the long-term and short-term time dependencies extracted by the bidirectional LSTM and input into the final fully connected layer to obtain the final RUL prediction.
[0103] For a detailed flowchart, see Figure 1 In step S2, the present invention proposes a method for extracting two types of correlations, and the process of obtaining graph structure data of time correlation and spatial correlation includes the following steps:
[0104] S0201: Various features of the original vibration data of the bearing are extracted, and then the extracted features are sampled using a moving window with a stride of 1 and a length of L.
[0105] S0202: Using the columns of window data as graph node vectors, a spatial correlation graph is established; using the rows of window data as graph node vectors, a temporal correlation graph is established.
[0106] In step S3, two graph data with different topological structures are established, and the spectral graph convolution operation is used in the adaptive graph, which specifically includes the following sub-steps:
[0107] S0301: Use the covariance matrix V (i,n) ∈R S*S And the cosine similarity matrix C (i,n) ∈RT*T To calculate the correlation of its data respectively. Covariance matrix V (i,n) It can be obtained by formula (1), where is the matrix V (i,n) Elements in represents the Sth column of the nth sampling data of the ith bearing; E[·] is the expectation.
[0108]
[0109] S0302: According to the matrix elements To obtain the adjacency matrix of the spatial graph Elements in if If is positive, it means that the change trends of the two features are consistent. Set to 1, that is, in the spatial correlation graph G of the feature s There is an edge between s and s′. If If it is negative, it means that the change trends of the two features are different. Set to 0, that is, in the spatial correlation graph G of the feature s There is no edge connection between s and s′.
[0110] S0303: Preliminary definition of the temporal correlation graph G by the cosine similarity matrix t , cosine similarity C (i,n) Elements in It can be calculated by formula (2), where represents the t-th row feature of the n-th sampling data of the ith bearing, and ||·|| is the binary norm of the vector.
[0111]
[0112] S0304: Apply the function o(·) defined in equation (3) to the cosine similarity matrix C (i,n) The elements in the adjacency matrix can be obtained Elements in δ is an artificially defined threshold. Indicates that there is no edge connection between two time nodes in the time-related graph, which means that they are unrelated; Indicates that two time nodes are connected by an edge in the time-dependent graph, which means that they are related in time.
[0113]
[0114] S0305: Define two matrices and name them source matrices W r and the target matrix W g , firstly, W r and Wg Multiply to get the adaptive feature space dependency, then introduce a nonlinear activation function Tanh to eliminate the residual secondary dependency, and finally get the adaptive adjacency matrix A through the softmax function adp . Adaptive adjacency matrix for spatial correlation It can be calculated by the formula, where is the learnable adaptive parameter matrix, It represents the adaptive spatial dependency after the hyperbolic activation function, and finally the 0-1 normalized adaptive adjacency matrix is generated by the softmax function, as shown in formula (4).
[0115]
[0116] S0306: Similarly, the adaptive adjacency matrix of time correlation is calculated by formula It is expressed as formula (5), where is a learnable adaptive parameter matrix.
[0117]
[0118] S0307: Use spectral convolution operation to extract spatial correlation and temporal correlation in the feature channel. First, define the spatiotemporal Laplacian matrix, which can be expressed as formula (6), where I s ,I t is the identity matrix, D s , D t is a diagonal matrix whose diagonal values are the degrees of each node in the spatial graph and the temporal graph, respectively.
[0119]
[0120] S0308: Perform orthogonal decomposition on the above equation to obtain the eigenvectors and eigenvalues of the Laplace matrix, as shown in equation (7). Where Λ=diag(λ0,λ1,λ2,...,λ X ) is the eigenvalue matrix, Q = [α0, α1, α2, ... α x ] is the eigenvector corresponding to the eigenvalue.
[0121]
[0122] S0309: The spectral graph convolution operation can be mathematically expressed as equation (8), where s(·) represents the spectral graph convolution operation, U represents the graph data sampled with the window, and g s , g t They represent the graph filters of the spatial graph and the temporal graph respectively. Substituting them into the equation, we can get the spectral graph convolution equation (9) in the form of Laplacian matrix.
[0123]
[0124]
[0125] S0310: In order to make the calculation easier, the first-order Chebyshev polynomial is generally used to approximate the graph convolution operation, thereby reducing the computational cost. First, let The scaled normalized Laplace matrix is obtained, which is expressed as formula (10).
[0126]
[0127] S0311: After introducing the first-order Chebyshev polynomial, equation (10) can be expressed as equation (11). Where p is the Chebyshev order of the approximation, and p=1 is taken here; θ p(s) is the Chebyshev approximation vector, is the p-order Chebyshev approximation polynomial.
[0128]
[0129] S0312: Finally, by adding the weight matrix, the output of the adaptive Chebyshev graph convolution can be obtained, which is expressed as formula (12).
[0130]
[0131] In step S4, a gated graph convolution mechanism is designed, and its output is reduced in dimension through DBN, which includes the following sub-steps:
[0132] S0401: Input the two predefined graphs obtained into the graph attention network with multi-head attention to further obtain more accurate edge weights, thereby more efficiently aggregating domain information, and its output can be obtained by formula (13).
[0133]
[0134] in, A s and A t is the two types of correlation adjacency matrix adjusted by the attention mechanism, ⊙ is the element multiplication, M is the dynamic attention coefficient, so is a dynamic adjacency matrix with attention mechanism, Z L is the output of the multi-head attention mechanism at layer L. If L = 0, then set Z L =U (i,n) .
[0135] S0402: In the graph data (i,n)After the multi-head graph attention network and the adaptive Chebyshev graph convolution layer extract the correlation, a gate value β is used to control its update and forgetting data, which can be obtained by formula (14).
[0136] β=sigmoid(H i,n ) (14)
[0137] S0403: In the A-GG module, β is the value output by the adaptive graph convolutional network obtained through a sigmoid function. In this gating mechanism, 1-β represents the amount of information forgotten by the previous layer A-GG, and β represents the amount of information updated by the current layer A-GG. The above update mechanism can be expressed by formula (15).
[0138]
[0139] Among them, σ(·) is the ReLU activation function. After extracting the correlation through several layers of adaptive gated convolution, each node is updated to obtain a more reasonable high-dimensional representation.
[0140] S0404: Output the correlation information extracted from the A-GG module into a DBN network with two hidden layers to achieve the extraction of low-dimensional abstract features.
[0141] The execution process of step S5 includes the following steps:
[0142] S0501: Input the feature channel data into the bidirectional LSTM to extract the time dependency. The update mechanism of the bidirectional LSTM can be expressed by equation (16), where σ(·) is the activation function, W and b are the weight matrices and bias parameters of each gate, i t 、f t , o t are the input gate, forget gate, and output gate at the tth position, respectively, c t is the cell state at the tth position, h t is the hidden layer information at the tth location.
[0143]
[0144] S0502: In the self-attention Bi-LSTM, the original data is first projected into a higher dimensional space to reduce the overlap of information and create information of subspaces under different representations, as shown in Equation (17).
[0145] Q=U (i,n) W Q , K=U (i,n) W K , V=U (i,n) W V (17)
[0146] In the formula, Q, K, and V represent the time step data U (i,n) The query subspace, key subspace, and value subspace obtained, W Q , W K , W V are their corresponding weight matrices respectively.
[0147] S0503: Scale the obtained multiple subspaces to obtain the attention output, which can be expressed as formula (18).
[0148]
[0149] Among them, d k Indicates the dimension of the input data.
[0150] S0504: Connect the attention outputs of multiple heads in series to obtain the final multi-head attention output, which is expressed as equation (19), where H represents the number of attention heads and U represents the series operation.
[0151]
[0152] S0505: In the final spatiotemporal fusion layer, the two types of features are flattened and input into the fully connected layer for final prediction, which can be expressed by formula (20), where W is the weight matrix in the FC layer, b is the bias vector in the FC layer, and σ(·) is the activation function.
[0153]
[0154] Furthermore, Figure 2 The present invention provides a schematic diagram of the process of predefining two types of correlation graphs. It shows how to convert two-dimensional data extracted by using a window in a feature channel into a graph structure. First, the rows and columns of the window data are extracted respectively to form node vectors in the two types of correlation graphs, and then the edge connections in the graph are constructed by covariance calculation and cosine similarity calculation to obtain a predefined adjacency matrix of the graph, and finally the node vectors and the adjacency matrix are combined to obtain two types of correlation graphs.
[0155] Figure 3 Schematic diagram of the spectral graph convolution framework with adaptive adjacency matrix provided by the present invention. In this figure, the source matrix W r and the target matrix W g The adaptive adjacency matrix is first obtained by multiplication, and then combined with the node data extracted by the window to obtain two types of adaptive graphs. Finally, the two types of graphs are spectrally convolved to obtain the deep spatiotemporal correlation of the feature channel data.
[0156] The following is further described in conjunction with specific embodiments:
[0157] Taking the PHM2012 bearing dataset and the XJTU-SY dataset as examples, 8 time domain features and 6 frequency domain features were extracted from the horizontal and vertical directions of the original vibration signal. In the PHM2012 dataset, 17 bearings were selected under three working conditions, namely: working condition 1 (1800rpm / 4000N), bearing1-1 to bearing1-7. Working condition 2 (1650rpm / 4200N), bearing2-1 to bearing2-7. Working condition 3 (1500rpm / 5000N), bearing3-1 to bearing3-3. In the XJTU-SY dataset, 10 bearings were selected under two working conditions, namely: working condition 1 (2100rpm / 12000N), bearing1-1 to bearing1-5. Condition 2 (2250rpm / 11000N), bearing2-1 to bearing2-5. The eight features in the time domain of this example include mean, standard deviation, maximum and minimum values, root mean square, skewness, kurtosis, and peak-to-peak value. The frequency domain features are extracted by fast Fourier transforming the original signal, including median frequency, average frequency, maximum amplitude, frequency of maximum amplitude, band power, and maximum power spectral density. These features have been proven to be effective in predicting the RUL of bearings. Then two graph data with different types of adjacency matrices are constructed.
[0158] Figure 4 This is a schematic diagram of the A-GG gating unit structure provided by the present invention. In order to avoid the over-smoothing problem caused by too many stacking of graph convolution layers, a gating mechanism based on two types of graph convolution operations is proposed to selectively update and forget data, so that the graph network can autonomously retain the most relevant information and enhance the accuracy of model prediction. In this example, the number of A-GG layers is set to 3 to obtain the best overall performance.
[0159] Figure 5 It is a schematic diagram of the LSTM unit structure provided by the present invention. After obtaining the two types of correlations in the feature channel data, it is also hoped to extract the local and long-term time dependencies in the original data through LSTM to further improve the model prediction accuracy. Here, the present invention proposes a new bidirectional LSTM with a self-attention mechanism, which can solve the problems of slow response to dynamic changes and unstable gradients in the previous recurrent network. In this example, the forward and reverse information of the bidirectional LSTM is taken and fused, and the final fused feature dimension is 40.
[0160] Figure 6It is a schematic diagram of the basic structure of the DBN model provided by the present invention. Considering that the current graph reading often uses operations such as splicing, maximum pooling, and average to obtain the correlation representation in the graph. However, extracting a single data feature cannot represent all nodes, which may lead to information omission and fail to fully reflect the abstract correlation information. The present invention proposes a spatiotemporal fusion layer based on a deep belief network (DBN). In this example, the number of hidden layers is 2.
[0161] Figure 7 It is a schematic diagram of the comparison of two different adjacency matrix diagrams provided by the present invention. First, it can be seen from Figure (a) and Figure (c) that the adaptive adjacency matrix shows greater sparsity, which makes it possible to consider the optimal edge weight more accurately, especially in Figure (a), it can be clearly captured that the weights occupied by certain features are large, proving that these features play a more important role in the prediction of RUL at a specific moment. But at the same time, the increase in sparsity will cause the model to be more prone to overfitting, and the predefined adjacency matrix can well alleviate the risk of overfitting. It can be seen from Figure (b) and Figure (d) that the conventional predefined adjacency matrix can also characterize the correlation between features to a certain extent, and can play a mitigating role in the overfitting that may be caused by the adaptive adjacency matrix. Secondly, it can be seen from the comparison of the two types of adjacency matrices that the adaptive adjacency matrix has an asymmetry that the traditional adjacency matrix does not have, so that the node aggregation is freed from the inherent balance characteristics, thereby increasing the robustness of the model and further improving the prediction accuracy.
[0162] Figure 8 It is the RUL prediction result of some bearings in the PHM2012 bearing data set provided by the present invention. It can be seen that the model of the present invention can accurately track the actual RUL of the bearing. For bearings Bearing1_5, Bearing2_4 and Bearing2_6, the predicted RUL curve is very consistent with the actual RUL. In other bearings such as Bearing1_6 and Bearing1_7, the predicted curve of the bearing has certain deviations and fluctuations in the second half, but it still conforms to the life curve of the bearing degradation as a whole. The reason for the error is that there is a certain interference in the process of predicting the remaining service life of the bearing. Specifically, the degradation of the bearing is often affected by multiple stresses at the same time. The alternating effects of these stresses will cause the degradation rate of the bearing to continue to produce new changes, making it difficult to accurately capture the degradation characteristics. Secondly, there may be a certain degree of noise information and mutation information in the training data used when training the RUL model, and this information will have a negative benefit on the learning of the prediction model, thereby causing misjudgment of the life of the bearing.
[0163] Fig. 9It is a schematic diagram comparing the prediction effects of various models of the PHM2012 bearing data set provided by the present invention. It can be seen that in most cases, the proposed method is significantly better than several other typical time series prediction models, achieving the minimum RMSE and MAE, and its average RMSE is as low as 0.099, while the lowest average RMSE of the remaining models ranges from 0.136 to 0.221. The proposed method improves the accuracy of model prediction. From the overall perspective of the RMSE and MAE graphs, the differences of the proposed methods in most bearings under these two indicators are the minimum values compared with other models. Compared with the C-LSTM with better prediction performance, the average difference in working condition 1 is 0.0096, while that of C-LSTM is 0.0214, indicating that the performance of the model on the prediction curve is more stable. From the perspective of input data, the input of TCN and C-LSTM is also time step data, while the input of the proposed method is two types of association graphs composed of time step data, so it can be concluded that converting the signal into graph data is helpful to improve the prediction performance.
[0164] Fig.10 It is a partial bearing prediction graph of the XJTU-SY data set provided by the present invention. As can be seen from the figure, the difference between the predicted curve and the true curve is very small. Although there are certain fluctuations in the predicted value of the bearing in certain specific cases, the overall prediction curve can still track the true value well. It should be noted that this set of data sets is not comparable with the PHM2012 bearing data set, because the bearings tested by the XJTU-SY data set have more failure modes, and most of their failure locations are also different, which greatly increases the randomness of the prediction and makes the extraction of degradation features more difficult. Secondly, the bearing training data provided by the XJTU-SY data set is less than that of the PHM2012 bearing data set, which also increases the difficulty of model prediction.
[0165] Fig.11It is a comparison chart of the prediction effects of each model of the XJTU-SY data set provided by the present invention. From the overall perspective of the two evaluation indicators, DCAGGCN shows better prediction performance than other models in most cases, with the lowest average RMSE of 0.108 and the lowest average MAE of 0.093, while the lowest average RMSE of the remaining models is 0.170 to 0.315. From this data, it can be seen that in the XJTU-SY data set with greater prediction difficulty, the average RMSE of each model has decreased to varying degrees, but the prediction performance of the proposed method remains at a relatively high level. This is because its unique adaptability and extraction of deep correlations allow the model to accurately identify the degradation characteristics of bearings and use them efficiently. From the RMSE comparison chart, it can be seen that the performance of the TGRU and CNN models is relatively unstable, and the prediction performance they show on different bearings varies greatly. Although the proposed method does not achieve the best evaluation effect on some bearings, its overall stability is stronger, indicating that the model has higher robustness. Under two bearings with different operating conditions, the prediction performance of the proposed method does not fluctuate greatly, and the RUL of bearings under different working conditions can be well predicted. This further illustrates that the model has strong generalization and can effectively extract the degradation features in the model.
[0166] Fig.12 It is a schematic diagram of the quantitative visualization of the performance of each model in the ablation experiment provided by the present invention. The different modules and structural composition methods of the ablation experiment are as follows: Model I is the proposed method, Model II removes the adaptive adjacency matrix, Model III removes the gating mechanism, and reduces the graph convolution layer to one layer, and Model IV removes the multi-head self-attention LSTM for extracting time dependency. It can be seen that after removing the adaptive adjacency matrix mechanism, the overall prediction performance of Model II decreases due to the lack of dynamic capture of bearing degradation features. Secondly, Model III loses the gating mechanism and thus cannot capture deep correlation features, causing its prediction accuracy to fluctuate greatly. In addition, Model IV does not consider time dependency, which causes some information to be lost in the final feature representation, thereby reducing the prediction accuracy, but due to the existence of the adaptive mechanism, its prediction distribution is relatively stable.
[0167] Fig.13 It is a schematic diagram of the model prediction effect under different threshold values of cosine similarity provided by the present invention. It can be seen that the optimal threshold value is about 0.7. As the threshold increases, RMSE and MAE both decrease first and then increase. The reason for this phenomenon is that when the threshold is too small, its filtering effect will also be poor, resulting in the edge of the graph data not being able to well reflect the actual correlation between features. When the threshold is too large, the threshold filter will filter out a large part of the effective information, making the connection of the graph data too sparse, and the extraction of correlation is hindered.
[0168] Fig.14 It is a schematic diagram of the training error and training time of the model under different A-GG layers provided by the present invention. It can be seen that when the number of A-GG layers is 3-4, the model obtains a better prediction effect, and after continuing to increase the number of A-GG layers, the model prediction ability shows negative growth. This is because too many layers cause the model to be over-smoothed to a certain extent, but the existence of the gating mechanism well suppresses the occurrence of this phenomenon. Considering the comprehensive training time, it can be seen that when the number of A-GG layers of the model is three, the best overall prediction performance can be obtained.
[0169] In summary, the present invention has the following beneficial technical effects:
[0170] (1) The present invention extracts two types of correlation features in the bearing feature channel, and uses an adaptive gated graph convolutional network to extract deep features. Finally, it fuses them with the long-term and short-term time dependencies extracted by the bidirectional LSTM to obtain the final RUL prediction, which can be used for the RUL prediction of bearings, greatly improving the generalization ability of the model and the accuracy of prediction.
[0171] (2) In the present invention, a dual-correlation neural network is proposed to enhance the characterization and extraction capabilities of bearing features of different scales.
[0172] (3) In the present invention, an adaptive graph structure is proposed, which can adaptively adjust the topological structure of the graph during the model training process to find the optimal graph construction form, thereby improving the robustness and accuracy of the model.
[0173] (4) In the present invention, a gated graph convolutional layer is proposed, which can avoid the over-smoothing problem caused by excessive stacking of the graph convolutional layer, so that the graph network can autonomously retain the most relevant information, thereby improving the stability of the model.
[0174] (5) The dual-correlation adaptive gated graph convolutional network designed in the present invention has the significant advantages of high prediction accuracy, strong prediction robustness and good generalization performance. It can meet the practical application needs of bearing life prediction in predictive maintenance, and also has broad application potential in other fields.
[0175] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.
Claims
1. A bearing remaining service life prediction method based on adaptive gated graph convolutional network, It is characterized in that The following steps are involved: Step 1: Collect vibration data during the operation of the bearing and convert it into a feature channel, wherein the vibration data includes normal state data and abnormal state data of the bearing; Step 2: Sample the converted feature channel data through the window and divide it into rows and columns to obtain graph structure data of time correlation and spatial correlation; Step 3: Use cosine similarity and covariance to obtain a predefined spatiotemporal correlation graph, and use the source matrix and target matrix to obtain two types of adaptive graphs; Step 4: Input the obtained predefined graph and the graph with adaptive mechanism into a three-layer gated graph convolutional neural network to extract two types of correlations of bearing feature channels, and use a deep belief network to extract its low-dimensional features; Step 5: The low-dimensional features extracted by DBN are fused with the long-term and short-term time dependencies extracted by the bidirectional LSTM and input into the final fully connected layer to obtain the final RUL prediction.
2. The method for predicting the remaining useful life of a bearing using an adaptive gated graph convolutional network as claimed in claim 1, It is characterized in that The process of obtaining the graph structure data of time correlation and space correlation in step 2 includes the following steps: Step 2.1: Extract the features of the original vibration data of the bearing, and then sample the extracted features using a moving window with a stride of 1 and a length of L; Step 2.2: Use the columns of the window data as graph node vectors to build a spatial correlation graph, and use the rows of the window data as graph node vectors to build a temporal correlation graph.
3. The method for predicting the remaining useful life of a bearing using an adaptive gated graph convolutional network as claimed in claim 2, It is characterized in that The execution process of step 3 includes the following steps: Step 3.1: Use the covariance matrix V (i,n) ∈R S*S And the cosine similarity matrix C (i,n) ∈R T*T To calculate the correlation between the spatial graph and the time graph, the covariance matrix V (i,n) The expression is as follows: in is the matrix V (i,n) Elements in represents the Sth column of the nth sampling data of the i-th bearing; E[·] is the expectation, and S represents the number of features extracted; Step 3.2: According to the matrix elements To obtain the adjacency matrix of the spatial graph Elements in if If is positive, it means that the change trends of the two features are consistent. Set to 1, that is, in the spatial correlation graph G of the feature s There is an edge between s and s′; if If it is negative, it means that the change trends of the two features are different. Set to 0, that is, in the spatial correlation graph G of the feature s There is no edge connection between s and s′; Step 3.3: Preliminary definition of the temporal correlation graph G by the cosine similarity matrix t , cosine similarity C (i,n) Elements in The expression is as follows: in represents the t-th row feature of the n-th sampling data of the i-th bearing, ||·|| is the second norm of the vector, and T represents the size of the time step; Step 3.4: Apply the function o(·) to the cosine similarity matrix C (i,n) The elements in can get the adjacency matrix Elements in δ is a threshold defined artificially, and the function o(·) is expressed as: Indicates that there is no edge connection between two time nodes in the time-related graph, which means that they are unrelated; Indicates that two time nodes are connected by an edge in the time-dependent graph, which means that they are related in time; Step 3.5: Define two matrices and name them source matrices W r and the target matrix W g First, W r and W g Multiply to get the adaptive feature space dependency, then introduce a nonlinear activation function Tanh to eliminate the residual secondary dependency, and finally get the adaptive adjacency matrix A through the softmax function adp , the expression is as follows: in is the learnable adaptive parameter matrix, It represents the adaptive spatial dependency after the hyperbolic activation function, and finally the 0-1 normalized adaptive adjacency matrix is generated by the softmax function; Step 3.6: Calculate the adaptive adjacency matrix of time correlation in the same way as step 3.5 The expression is as follows: in is a learnable adaptive parameter matrix; Step 3.7: Define the space-time Laplacian matrix, expressed as follows: Among them I s ,I t is the identity matrix, D s , D t is a diagonal matrix, and the values on its diagonal are the degrees of each node in the spatial graph and the temporal graph respectively; Step 3.8: Orthogonal decomposition to obtain the eigenvectors and eigenvalues of the Laplacian matrix Where Λ=diag(λ 0 ,λ 1 ,λ 2 ,...,λ X ) is the eigenvalue matrix, Q = [α 0 ,α 1 ,α 2 ,...α x ] is the eigenvector corresponding to the eigenvalue; Step 3.9: Perform spectral convolution operation, Where s(·) represents the spectral graph convolution operation, U represents the graph data sampled with the window, and g θ(s) , g θ(t) Graphic filters representing spatial and temporal graphs, respectively; Substituting into the equation, the spectral convolution equation in the form of a Laplace matrix is as follows: Step 3.10: Order This gives the scaled normalized Laplacian matrix Step 3.11: Introduce the first-order Chebyshev polynomial to transform the expression as follows: Where p is the Chebyshev order of the approximation, and p=1 is taken here; θ p(s) is the Chebyshev approximation vector, is the p-order Chebyshev approximation polynomial; Step 3.12: Finally, by adding the weight matrix Θ, the output of the adaptive Chebyshev graph convolution can be obtained, and the expression is as follows:
4. The method for predicting the remaining useful life of a bearing using an adaptive gated graph convolutional network as claimed in claim 3, It is characterized in that The execution process of step 4 is to extract two types of data correlations through three layers of gated graph convolution layers and perform dimensionality reduction through DBN. Each layer of gated graph convolution layers is composed of two types of convolution and gating mechanisms. The first type of convolution is the self-attention graph convolution in which the predefined graph participates, and the second type of convolution is the spectral graph convolution in which the adaptive graph participates. Specifically, it includes the following steps: Step 4.1: Input the two predefined graphs into the graph attention network with multi-head attention to further obtain more accurate edge weights, the output of which can be obtained by the following formula: in, A s and A t is the two types of correlation adjacency matrix adjusted by the attention mechanism, ⊙ is the element multiplication, M is the dynamic attention coefficient, so is a dynamic adjacency matrix with attention mechanism, Z L is the output of the multi-head attention mechanism at layer L. If L = 0, then set Z L =U (i,n) ; Step 4.2: In the graph data U (i,n) After the multi-head graph attention network and the adaptive Chebyshev graph convolution layer extract the correlation, a gate value β is used to control its update and forgetting data. The operation expression is as follows: β=sigmoid(H i,n ); Step 4.3: In the gated graph convolutional neural network module, β is the value of the adaptive graph convolutional network output obtained through a sigmoid function, and the corresponding update mechanism expression is as follows: Among them, 1-β represents the amount of information forgotten by the previous layer of gated graph convolutional neural network, and β represents the amount of information updated by the current layer of gated graph convolutional neural network; σ(·) is the ReLU activation function. After extracting the correlation through several layers of adaptive gated convolution, each node is updated to obtain a more reasonable high-dimensional representation; Step 4.4: Output the correlation information extracted from the gated graph convolutional neural network module into a DBN network with two hidden layers to achieve the extraction of low-dimensional abstract features.
5. The method for predicting the remaining useful life of a bearing using an adaptive gated graph convolutional network as claimed in claim 4, It is characterized in that The execution process of step 5 includes the following steps: Step 5.1: Input the feature channel data into the bidirectional LSTM to extract the time dependency. The update mechanism of the bidirectional LSTM is as follows: Where σ(·) is the activation function, W and b are the weight matrices and bias parameters of each gate, i t 、f t , o t are the input gate, forget gate, and output gate at the tth position, respectively, c t is the cell state at the tth position, h t is the hidden layer information at the tth position; Step 5.2: In the self-attention Bi-LSTM, the original data is first projected into a higher-dimensional space to reduce the overlap of information and create information in subspaces under different representations. The expression is as follows: Q=U (i,n) W Q ,K=U (i,n) W K ,V=U (i,n) W V Among them, Q, K, and V represent the time step data U (i,n) The query subspace, key subspace, and value subspace obtained, W Q , W K , W V are their corresponding weight matrices respectively; Step 5.3: Scale the obtained multiple subspaces to obtain the attention output, which is expressed as follows: Among them, d k Indicates the dimension of the input data; Step 5.4: Concatenate the attention outputs of multiple heads to get the final multi-head attention output, which is expressed as follows: Where H represents the number of attention heads and U represents the serial operation; Step 5.5: In the final spatiotemporal fusion layer, the two types of features are flattened and input into the fully connected layer for final prediction, which is expressed as follows: Where W is the weight matrix in the FC layer, b is the bias vector in the FC layer, and σ(·) is the activation function.
Citation Information
Cited By
Sodium-ion battery health state estimation method based on multi-space convolution gating network
CN120009739A
Method for predicting and quantifying residual service life of bearing based on dynamic graph convolution distillation network
CN121301804A
A method for bearing residual useful life prediction quantification of a dynamic graph convolution distillation network
CN121301804B
Bearing device residual life prediction model and method based on TCN and decay diagram
CN121479217A
Bearing device residual life prediction model and method based on tcn and decay map
CN121479217B
