Method for predicting and quantifying residual service life of bearing based on dynamic graph convolution distillation network

Abstract

The invention discloses a method for predicting and quantifying the residual service life of a bearing based on a dynamic graph convolution distillation network, and belongs to the field of prediction of the residual service life of the bearing, and the method comprises the steps: collecting a full-life vibration signal, and converting the full-life vibration signal into feature channel data; constructing a dynamic graph, and learning deep interaction between nodes by using a self-adaptive adjacent matrix; the teacher model extracts space-time joint features, and the space-time joint features are migrated to the student model through multi-level knowledge distillation; a student model outputs RUL through lightweight graph convolution, and a confidence interval is quantified by Bayesian bicorrelation uncertainty. According to the method, high precision is kept while the parameters are compressed, cross-working-condition robust prediction is achieved, and the industrial maintenance decision credibility is improved.

Description

Technical Field

[0001] This invention belongs to the field of bearing remaining service prediction, and particularly relates to a method for predicting and quantifying bearing remaining service using a dynamic graph convolutional distillation network. Background Technology

[0002] Modern industrial machinery typically withstands high loads and complex operating conditions. Bearings, as a crucial component of these machines, play a vital role in supporting and transmitting loads, significantly impacting equipment performance and reliability. Therefore, bearings are widely used in various industrial sectors, including machinery manufacturing, aerospace, energy generation and transmission, and rail transportation. Accurately assessing bearing health and predicting their remaining life is essential for predictive maintenance and minimizing equipment failure risks to maximize reliability. Thus, predicting the remaining life of bearings is of significant research importance.

[0003] Existing methods for predicting bearing remaining service based on deep learning suffer from the problem that large-scale training parameters reduce the computational efficiency of deep learning models, and the insufficient reliability of bearing RUL prediction results limits their effective application in industrial scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for predicting and quantifying the remaining service life of bearings using dynamic graph convolutional distillation networks, comprising:

[0005] Vibration signals are collected throughout the entire life cycle of the bearing, and the vibration signals are converted into characteristic channel data.

[0006] A dynamic graph is constructed based on the feature channel data. The dynamic graph includes nodes and edges. The nodes represent features in the feature channel data, and the edges represent the relationships between features.

[0007] The dynamic graph is updated using an adaptive adjacency matrix, which is used to learn deep interaction relationships between nodes.

[0008] Based on the dynamic graph, spatiotemporal joint features are extracted through the teacher model, and the spatiotemporal joint features are transferred to the student model through a multi-level knowledge distillation strategy.

[0009] The remaining service life of the bearing is predicted using the student model, and the uncertainty of the prediction results is quantified based on the Bayesian method to obtain the confidence interval of the bearing's RUL.

[0010] Optionally, converting the vibration signal into characteristic channel data includes:

[0011] The vibration signal is subjected to time-domain and frequency-domain feature extraction to obtain a multi-dimensional feature vector;

[0012] The multidimensional feature vectors are arranged in chronological order to construct a feature channel data matrix.

[0013] Optionally, the construction of the dynamic graph includes:

[0014] Sliding window sampling is performed on the feature channel data to obtain window data;

[0015] Perform a linear transformation on the window data and project it into a new feature space;

[0016] Calculate the similarity matrix between the transformed features, and construct an initial adjacency matrix based on the similarity matrix;

[0017] The initial adjacency matrix is ​​normalized to obtain the time dependency matrix;

[0018] A gating mechanism is introduced to weight the time dependency matrix to obtain a dynamic adjacency matrix;

[0019] A dynamic graph is constructed based on the dynamic adjacency matrix.

[0020] Optionally, the update of the adaptive adjacency matrix includes:

[0021] The similarity matrix is ​​normalized and nonlinearly mapped to obtain a normalized adjacency matrix;

[0022] The normalized adjacency matrix is ​​applied to the dynamic graph to update the connection weights between nodes.

[0023] Optionally, the teacher model extracts spatiotemporal joint features including:

[0024] The dynamic graph is input into a graph convolutional network to extract spatial features;

[0025] Multiple graph convolutional layers are stacked along the time dimension to extract temporal features;

[0026] By integrating the aforementioned spatial and temporal features, a spatiotemporal joint feature is obtained.

[0027] Optionally, the multi-level knowledge distillation strategy includes:

[0028] Align the intermediate layer features of the teacher model and the student model, and calculate the structural consistency loss;

[0029] Sparsity constraints are imposed on the adjacency matrices of the teacher and student models, and the sparsity loss is calculated.

[0030] The local structures of the teacher model and the student model are compared, and the local structure loss is calculated.

[0031] All losses are combined to construct a total distillation loss function, which is used to guide student model training.

[0032] Optionally, the total distillation loss function includes:

[0033] The difference loss between the outputs of the teacher model and the student model;

[0034] The weighted sum of structural consistency loss, sparsity loss, and local structural loss.

[0035] Optionally, the uncertainty quantification includes:

[0036] Statistical analysis was performed on multiple prediction results to obtain the prediction mean and cognitive uncertainty.

[0037] Based on the random uncertainty of predictive residual calculation;

[0038] By integrating the aforementioned cognitive uncertainty and accidental uncertainty, a dual-correlated uncertainty is obtained, and a confidence interval is constructed.

[0039] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0040] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] This invention employs a dynamic graph structure to uniformly model the temporal and spatial correlations of feature channel data, enabling the model to adaptively capture deep interactive features during bearing degradation, significantly improving prediction accuracy and robustness. The introduction of an adaptive adjacency matrix allows the graph structure to dynamically update with data evolution during training, enhancing the model's adaptability to different operating conditions and degradation modes, effectively avoiding the limitations of fixed structures in traditional graph networks. A multi-level spatiotemporal joint knowledge distillation strategy efficiently transfers knowledge from a large-capacity teacher model to a lightweight student model, maintaining high prediction performance while compressing model parameter size, meeting the dual requirements of real-time performance and deployment efficiency in industrial scenarios. Utilizing a Bayesian inference-based dual-correlation uncertainty quantification mechanism, cognitive uncertainty and accidental uncertainty are systematically separated and estimated, generating statistically significant confidence intervals, significantly improving the interpretability and credibility of prediction results. Attached Figure Description

[0043] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0044] Figure 1 This is a schematic flowchart illustrating the steps of a method for predicting and quantifying the remaining service life of bearings using a dynamic graph convolutional distillation network, as described in this invention.

[0045] Figure 2 This is a schematic diagram of GCN provided in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the gate control unit structure provided in an embodiment of the present invention.

[0047] Figure 4 This is a schematic diagram of the construction time correlation and spatial correlation data in an embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram of the multi-level spatiotemporal joint knowledge distillation strategy in an embodiment of the present invention.

[0049] Figure 6 This is a schematic diagram illustrating the quantification of dual-correlation uncertainty in an embodiment of the present invention.

[0050] Figure 7 This is a partial bearing RUL prediction result from the PHM2012 bearing dataset in this embodiment of the invention.

[0051] Figure 8 This is a schematic diagram comparing the prediction effects of various models in the PHM2012 bearing dataset in an embodiment of the present invention.

[0052] Figure 9 This is a partial uncertainty quantification result of the PHM2012 bearing dataset in this embodiment of the invention.

[0053] Figure 10 This is a partial bearing prediction result from the XJTU-SY dataset in this embodiment of the invention.

[0054] Figure 11 This is a comparison chart of the prediction performance of various models on the XJTU-SY dataset in this embodiment of the invention.

[0055] Figure 12 This is the partial uncertainty quantization result of the XJTU-SY bearing dataset in the embodiments of the present invention.

[0056] Figure 13 This is a visualization diagram of the performance quantification of each model in the ablation experiment in this embodiment of the invention.

[0057] Figure 14This is a schematic diagram illustrating the effect of multi-level spatiotemporal joint knowledge distillation in an embodiment of the present invention.

[0058] Figure 15 This is a schematic diagram illustrating the impact of partial channel data from the PHM2012 bearing dataset on the prediction performance of this model in an embodiment of the present invention.

[0059] Figure 16 This is a schematic diagram illustrating the impact of partial channel data from the XJTU-SY bearing dataset on the prediction performance of this model in an embodiment of the present invention. Detailed Implementation

[0060] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0061] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0062] Example 1

[0063] This embodiment provides a method for predicting and quantifying the remaining service life of bearings using a dynamic graph convolutional distillation network, including:

[0064] Vibration signals are collected throughout the entire life cycle of the bearing, and the vibration signals are converted into characteristic channel data.

[0065] A dynamic graph is constructed based on the feature channel data. The dynamic graph includes nodes and edges. The nodes represent features in the feature channel data, and the edges represent the relationships between features.

[0066] The dynamic graph is updated using an adaptive adjacency matrix, which is used to learn deep interaction relationships between nodes.

[0067] Based on the dynamic graph, spatiotemporal joint features are extracted through the teacher model, and the spatiotemporal joint features are transferred to the student model through a multi-level knowledge distillation strategy.

[0068] The remaining service life of the bearing is predicted using the student model, and the uncertainty of the prediction results is quantified based on the Bayesian method to obtain the confidence interval of the bearing's RUL.

[0069] The specific process includes:

[0070] Step 1: Synchronously collect bearing life cycle monitoring data and capture vibration signals from the healthy to the damaged bearing through a high-frequency vibration accelerometer;

[0071] Step 2: Convert the vibration data collected during the bearing operation to the feature channel. The vibration data includes normal bearing condition data and abnormal bearing condition data.

[0072] Step 3: Sample the transformed feature channel data through a window and divide it into rows and columns to obtain graph structure data of the temporal and spatial correlation of the teacher model;

[0073] Step 4: Use an adaptive adjacency matrix to automatically learn the relationships between nodes from the data, reflecting the deep interaction relationships in the bearing degradation process;

[0074] Step 5: Use a multi-level spatiotemporal joint knowledge distillation strategy to transfer the experiential knowledge of the teacher model to the student model, so that the student model has a similar predictive ability to the teacher model.

[0075] Step 6: The student model obtains the final RUL prediction through GCN, and uses a Bayesian-based method to quantify the confidence interval of the bearing RUL.

[0076] The execution process of step 3 includes the following steps:

[0077] Features are extracted from the original vibration data of the bearing, and then the extracted features are sampled using a moving window with a step size of 1 and a length of 10.

[0078] The sampled data window data is subjected to a linear transformation to project the input features onto a new feature space, as shown in the following expression:

[0079] ;

[0080] in, It is the input feature matrix. and These are learnable weights and biases.

[0081] according to Calculate the similarity matrix between the transformed features:

[0082] ;

[0083] Where S is the similarity matrix.

[0084] according to Application of similarity matrix The function was normalized to obtain the time dependency matrix:

[0085] ;

[0086] in, The time dependency matrix, For nodes and nodes Similarity between them For nodes Calculate the similarity with all nodes. This step is to normalize all similarities to between 0 and 1.

[0087] The importance of adjusting connections between nodes is determined by constructing an adaptive adjacency matrix, expressed as follows:

[0088] ;

[0089] in, yes Activation function These are learnable weights. It is a learnable bias.

[0090] Get a dimension as The gated vector is applied to the adjacency matrix, and then extended into matrix form through a broadcast mechanism, as shown in the following expression:

[0091] ;

[0092] in, It is a gating matrix.

[0093] use The time dependency matrix is ​​calculated and expressed as follows:

[0094] ;

[0095] in, This indicates element-wise multiplication.

[0096] The dynamic matrix of message passing is calculated as follows:

[0097] ;

[0098] in, It is a learnable weight matrix used to transform features to capture more complex dependencies.

[0099] The execution process of step 4 includes the following steps:

[0100] Calculated from step 3 Applying normalization and a nonlinear mapping to it, the expression is as follows:

[0101] ;

[0102] in, The adjacency matrix is ​​calculated along the row direction to make the sum of the connection weights of each node to all other nodes equal to 1, thus obtaining the adaptive adjacency matrix.

[0103] The execution process of step 5 includes the following steps:

[0104] The sampled data is projected into the latent space through a linear transformation, as shown in the following expression:

[0105]

[0106] in, Z represents the projection weights, and Z is the latent space.

[0107] Construct the time dependency matrix using inner product operations.

[0108] The structural consistency loss is calculated using the following expression:

[0109] ;

[0110] in, It is the input data. For structural consistency loss, To ensure structural consistency, the output is a desired reference matrix with a structure consistent with the input data. By calculating the sum of squared differences between the two, the model is forced to preserve the structural features of the input data during processing.

[0111] The sparsity loss of the adjacency matrix is ​​calculated as follows:

[0112] ;

[0113] The loss is due to the sparsity of the adjacency matrix. Let be the dimension of the adjacency matrix. For nodes and nodes The connection weights between them.

[0114] The local structural loss is calculated using the following expression:

[0115] ;

[0116] Among them, hyperparameters Control the strength of the constraints.

[0117] Based on the output of the teacher model and the output of the student model The total distillation loss is calculated using the following expression:

[0118] ;

[0119] in, This is the total distillation loss.

[0120] The execution process of step 6 includes the following steps:

[0121] For feature channel data, the posterior distribution of the parameters is calculated as follows:

[0122] ;

[0123] For feature channel data, the predicted distribution of the calculated parameters is expressed as follows:

[0124] ;

[0125] in, For model parameters, For input data.

[0126] Variational inference is used, with approximate distributions. The method approximates the posterior distribution, and the parameters Applying the Bernoulli variational distribution, the expression is as follows:

[0127] ;

[0128] Where L is the number of network layers. Let L be the Dropout mask matrix for the Lth layer. This represents the Dropout probability.

[0129] The predicted distribution is approximated by T Monte Carlo samplings, as shown in the following expression:

[0130] ;

[0131] in, These are the Dropout model parameters at the t-th sampling.

[0132] Since the Gaussian distribution is suitable for regression tasks such as bearing RUL prediction, we assume that the likelihood function is a Gaussian distribution, expressed as follows:

[0133] ;

[0134] in, It is the output of the neural network. Let Variance be the variance.

[0135] The mean of the predicted distribution is obtained as follows:

[0136] ;

[0137] The cognitive uncertainty estimate is obtained as follows:

[0138] ;

[0139] Using maximum likelihood estimation, during the prediction phase, all Dropouts in the model are ensured to be activated and predictions are made. These predictions are then used to evaluate the stochastic uncertainty in the model, as expressed below:

[0140] ;

[0141] Combining cognitive uncertainty and accidental uncertainty, the quantitative expression for dual-correlation uncertainty is as follows:

[0142] ;

[0143] Example 2

[0144] This embodiment provides a method for predicting and quantifying the remaining service life of bearings using a dynamic graph convolutional distillation network, including:

[0145] The following annotations provide details for the corresponding English abbreviations and terms, to facilitate subsequent descriptions and explanations:

[0146] RUL: Remaining service life;

[0147] GCN: Graph Convolutional Neural Network;

[0148] Dropout: Units dropped in a neural network;

[0149] The specific process of this embodiment includes:

[0150] S1: Synchronously collect bearing life cycle monitoring data and capture vibration signals from the healthy to the damaged bearing through a high-frequency vibration accelerometer;

[0151] S2: Convert the vibration data collected during the operation of the bearing to the feature channel. The vibration data includes normal bearing data and abnormal bearing data.

[0152] S3: The transformed feature channel data is sampled through a window and divided into rows and columns to obtain the graph structure data of the teacher model’s temporal and spatial correlations.

[0153] S4: Use an adaptive adjacency matrix to automatically learn the correlation between nodes from the data, reflecting the deep interaction relationships in the bearing degradation process;

[0154] S5: Use a multi-level spatiotemporal joint knowledge distillation strategy to transfer the experiential knowledge of the teacher model to the student model, so that the student model has a similar predictive ability to the teacher model.

[0155] S6: The student model obtains the final RUL prediction through GCN, and uses a Bayesian-based method to quantify the confidence interval of the bearing RUL.

[0156] For a detailed flowchart, please refer to Figure 3 In step S3, the present invention proposes a process for sampling the transformed feature channel data through a window and dividing it into rows and columns to obtain graph structure data of the teacher model's temporal and spatial correlations, including the following steps:

[0157] S0301: Extract the features from the original vibration data of the bearing, and then sample the extracted features using a moving window with a step size of 1 and a length of 10.

[0158] S0302: Project the input features onto a new feature space by performing a linear transformation on the sampled data window data, as shown in the following expression:

[0159] ;

[0160] in, It is the input feature matrix. and These are learnable weights and biases.

[0161] S0303: According to Calculate the similarity matrix between the transformed features:

[0162] ;

[0163] S0304: According to Application of similarity matrix The function was normalized to obtain the time dependency matrix:

[0164] ;

[0165] S0305: Constructing an adaptive adjacency matrix to adjust the importance of connections between nodes, expressed as follows:

[0166] ;

[0167] in, yes Activation function These are learnable weights. It is a learnable bias.

[0168] S0206: Obtain a dimension of The gated vector is applied to the adjacency matrix, and then extended into matrix form through a broadcast mechanism, as shown in the following expression:

[0169] ;

[0170] in, It is a gating matrix.

[0171] S0307: Use The time dependency matrix is ​​calculated and expressed as follows:

[0172] ;

[0173] in, This indicates element-wise multiplication.

[0174] S0308: Calculate the dynamic matrix of message passing, with the following expression:

[0175] ;

[0176] in, It is a learnable weight matrix used to transform features to capture more complex dependencies.

[0177] In step S4, an adaptive adjacency matrix is ​​used to automatically learn the correlations between nodes from the data, reflecting the deep interaction relationships in the bearing degradation process. This includes the following sub-steps:

[0178] S0401: Calculated from S0303 Applying normalization and a nonlinear mapping to it, the expression is as follows:

[0179] ;

[0180] in, The adjacency matrix is ​​calculated along the row direction to make the sum of the connection weights of each node to all other nodes equal to 1, thus obtaining the adaptive adjacency matrix.

[0181] In step S5, a multi-level spatiotemporal joint knowledge distillation strategy is used to transfer the experiential knowledge of the teacher model to the student model, enabling the student model to have predictive capabilities similar to the teacher model. This specifically includes the following sub-steps:

[0182] S0501: Project the sampled data into the latent space through a linear transformation, as shown in the following expression:

[0183] ;

[0184] in, It is the projected weight.

[0185] S0502: Repeat steps 2.4 to 2.7 to construct the time dependency matrix through inner product operations.

[0186] S0503: Calculate the structural consistency loss, the expression is as follows:

[0187] ;

[0188] in, This is the input data.

[0189] S0504: Calculate the sparsity loss of the adjacency matrix, as shown in the following expression:

[0190] ;

[0191] S0505: Calculate the local structural loss, the expression is as follows:

[0192] ;

[0193] Among them, hyperparameters Control the strength of the constraints.

[0194] S0506: Based on the output of the teacher model and the output of the student model The total distillation loss is calculated using the following expression:

[0195] ;

[0196] in, This is the total distillation loss.

[0197] In step S6, the student model obtains the final RUL prediction through GCN, and uses a Bayesian-based method to quantify the confidence interval of the bearing RUL, specifically including the following sub-steps:

[0198] S0601: For feature channel data, the posterior distribution of the parameters is calculated as follows:

[0199] ;

[0200] S0602: For feature channel data, calculate the predicted distribution of the parameters, as shown in the following expression:

[0201] ;

[0202] in, For model parameters, For input data.

[0203] S0603: Employing variational inference with approximate distributions The method approximates the posterior distribution, and the parameters Applying the Bernoulli variational distribution, the expression is as follows:

[0204] ;

[0205] Where L is the number of network layers. Let L be the Dropout mask matrix for the Lth layer. This represents the Dropout probability.

[0206] S0604: The predicted distribution is approximated by T Monte Carlo samplings, as shown in the following expression:

[0207] ;

[0208] in, These are the Dropout model parameters at the t-th sampling.

[0209] S0605: Since the Gaussian distribution is suitable for regression tasks such as bearing RUL prediction, it is assumed that the likelihood function is a Gaussian distribution, expressed as follows:

[0210] ;

[0211] in, It is the output of the neural network. Let Variance be the variance.

[0212] S0606: Obtain the mean of the predicted distribution, expressed as follows:

[0213] ;

[0214] S0607: Obtain the estimate of cognitive uncertainty, expressed as follows:

[0215] ;

[0216] S0608: Using maximum likelihood estimation, ensure that all Dropouts in the model are activated and make predictions during the prediction phase. Evaluate the stochastic uncertainty in the model using these prediction results. The expression is as follows:

[0217] ;

[0218] S0609: Combining cognitive uncertainty and accidental uncertainty, the quantitative expression for dual-correlation uncertainty is as follows:

[0219] ;

[0220] Furthermore, Figure 1 This is a flowchart of the prediction quantization process of the dynamic graph convolutional distillation network of the present invention. The network includes a teacher network and a student network. Knowledge is distilled from the teacher network and fed to the student network for prediction. The prediction is quantized using bicorrelation uncertainty to output a confidence interval.

[0221] Figure 2 This is a schematic diagram of the GCN provided by the present invention. It shows how to convert two-dimensional data extracted from feature channels using windows into a graph structure, and extract data features through convolution operations. First, the rows and columns of the window data are extracted to form node vectors in the graph. Then, the edges in the graph are constructed to connect them, resulting in a graph. Finally, convolution is used to obtain higher-dimensional features.

[0222] Figure 3 This is a schematic diagram of the gated unit structure provided by the present invention. To avoid the oversmoothing problem caused by excessive stacking of graph convolutional layers, a gating mechanism is used to selectively update and forget data, thereby enabling the graph network to autonomously retain the most relevant information and enhance the accuracy of model predictions.

[0223] The following is a further explanation with reference to specific embodiments:

[0224] Taking the PHM2012 bearing dataset and the XJTU-SY dataset as examples, eight time-domain features and six frequency-domain features were extracted from the horizontal and vertical directions of the original vibration signals, respectively. In the PHM2012 dataset, 14 bearings were selected under two operating conditions: Condition 1 (1800rpm / 4000N), bearing1-1 to bearing1-7; Condition 2 (1650rpm / 4200N), bearing2-1 to bearing2-7. In the XJTU-SY dataset, 15 bearings were selected under two operating conditions: Condition 1 (2100rpm / 12000N), bearing1-1 to bearing1-5; Condition 2 (2250rpm / 11000N), bearing2-1 to bearing2-5. Operating condition 3 (2400 rpm / 10000 N), bearing 3-1 to bearing 3-5. This example utilizes 13 features in the time-frequency domain, including kurtosis, entropy, fractal dimension, peak factor, impulse factor, crest factor, energy ratio, spectral flatness, mean, variance, skewness, peak vibration, and root mean square vibration. These features have proven effective in predicting the RUL of bearings. Two types of graphical data with different adjacency matrices are then constructed.

[0225] Figure 4 This is a schematic diagram of the A-GG gated unit structure provided by this invention. To avoid the oversmoothing problem caused by excessive stacking of graph convolutional layers, a gating mechanism based on two types of graph convolutional operations is proposed to selectively update and forget data, enabling the graph network to autonomously retain the most relevant information and enhance the accuracy of model predictions. In this example, the number of A-GG layers is set to 3 to obtain the optimal overall performance.

[0226] Figure 5This is a schematic diagram of the multi-level spatiotemporal joint knowledge distillation strategy provided by this invention. By utilizing the gradient in the backpropagation algorithm, the impact of each training step was analyzed, and the training processes of the two models were adjusted accordingly. The knowledge extraction process achieves adaptive learning and self-distillation of the multi-level spatiotemporal structure by constructing a dynamic structure. The teacher model generates a dynamic adjacency matrix in the hierarchical transmission and forms an internal distillation mechanism for structural knowledge by constraining behavior through local structural loss.

[0227] Figure 6 This is a schematic diagram illustrating the dual-correlation uncertainty quantification method of this invention. In statistics, these uncertainties can be divided into two categories: one is cognitive uncertainty, stemming from incomplete understanding of model structure, parameter estimation, etc.; the other is random uncertainty, mainly caused by inherent noise and random errors in the observed data. To assess the impact of these two types of uncertainty on the bearing RUL prediction process, this invention establishes a dual-correlation uncertainty quantification method based on bearing data and model parameters, encompassing both cognitive and random uncertainties.

[0228] Figure 7 This invention presents partial bearing RUL prediction results from the PHM2012 bearing dataset in this embodiment. Overall, the proposed method accurately tracks the actual RUL of the bearings. For bearings Bearing 1-5, Bearing 2-4, and Bearing 2-6, the predicted RUL curves closely match the actual RUL. Looking at different operating conditions, in condition 1, the overall prediction trend of the model basically matches the actual RUL, but there are prediction deviations at the end, such as in Bearing 1-6 and Bearing 1-7, and the average RMSE under condition 1 is higher than under condition 2. In condition 2, the model accuracy is significantly improved, especially for Bearing 2-6, whose prediction curve almost perfectly matches the actual value throughout, with a narrow and stable error band, making it the globally optimal case. However, the serious underestimation of life in the later stages of Bearing 2-5 in condition 2 exposes the model's sensitivity to sudden failure modes. For other bearings, such as Bearing 2-3 and Bearing 2-5, although the predicted curves have some deviations and fluctuations, they still generally conform to the bearing degradation life curve. In summary, dynamic graph structure modeling effectively improves the fitting ability to long-term degradation trends, while knowledge distillation reduces model complexity while maintaining a low average error. The errors arise because of interference in the bearing RUL prediction process. Specifically, bearing degradation is often affected by multiple factors, and the alternating effects of these factors cause the bearing degradation rate to constantly change, making it difficult to accurately capture degradation characteristics. Secondly, the training data used to train the model may contain a certain degree of noise and abrupt changes, which can negatively impact the learning of the prediction model, leading to misjudgments in bearing RUL prediction.

[0229] Figure 8 This invention presents a comparative diagram of the prediction performance of various models on the PHM2012 bearing dataset in this embodiment. It can be seen that in most cases, this method significantly outperforms the other models, achieving a minimum RMSE of 0.070 and an average RMSE of 0.094. The minimum RMSEs of the other models are as follows: TCN-GAU: 0.069, GTCN-GAU: 0.103, BGT: 0.095, and GNN: 0.174. This demonstrates that the proposed method has higher model prediction accuracy. From the perspective of different operating conditions, this method exhibits high generalization ability. Operating conditions 1 and 2 do not show a sharp increase in MAE and RMSE due to poor generalization ability, while the other models show varying degrees of fluctuation in evaluation indicators due to changes in operating conditions. For example, BGT's RMSE values ​​are relatively stable in operating condition 1, but change drastically in operating condition 2. A similar example is GTCN-GAU, which also shows RMSE fluctuations due to different operating conditions. From the perspective of model robustness, this method has higher model robustness. Whether it is working condition 1 or working condition 2, the changes of the various MAE and RMSE values ​​of the model are mostly kept within 20%, while the changes of the various MAE and RMSE values ​​of other models such as GNN and BGT obviously exceed this range.

[0230] Figure 9 This is a partial uncertainty quantification result of the PHM2012 bearing dataset in this embodiment of the invention. The confidence intervals of the prediction results for Bearing 1-5 and Bearing 1-6 are visualized. The red shaded area represents the 95% confidence interval. It can be seen that the interval range gradually shrinks over time and tightly encloses the prediction curve, effectively reflecting the uncertainty distribution characteristics during the degradation process.

[0231] Figure 10 This is a partial bearing prediction result from the XJTU-SY dataset in this embodiment of the invention. The model exhibits high overall prediction accuracy and stability. Among individual samples, Bearing 1-3 shows the best prediction accuracy, with an MSE of only 0.004, the lowest RMSE of 0.069, and a MAE of 0.056. Furthermore, R... 2 The RMSE of 0.910 indicates that the predicted curve fits the actual degradation curve very well. In contrast, Bearing 3-4 has the largest error, with an RMSE of 0.137, suggesting that its degradation process may be more complex, increasing the difficulty of prediction.

[0232] Figure 11This is a comparison chart of the prediction performance of various models on the XJTU-SY dataset in this embodiment of the invention. It can be seen that in most cases, the proposed method is significantly better than the other models, achieving the lowest RMSE and MAE. Its lowest RMSE reaches 0.069, and the average RMSE is 0.099, which proves the prediction performance of the proposed method.

[0233] Figure 12 This is the partial uncertainty quantification result of the XJTU-SY bearing dataset in this embodiment of the invention. The predicted bearing RUL curve maintains a high degree of consistency with the actual bearing RUL curve for most of the time period, and the confidence interval is relatively uniformly distributed without overexpansion. In Bearing 2-3, the predicted interval can effectively enclose the actual curve in the middle and later stages, while in Bearing 3-4, although the degradation curve fluctuates greatly in the middle stage, the model can still maintain a high coverage and a relatively stable interval width.

[0234] Figure 13 This is a visualization diagram illustrating the quantitative performance of each model in the ablation experiment of this invention. As can be seen from the diagram, after removing the adaptive adjacency matrix mechanism, due to the lack of node capture of bearing degradation features, the overall prediction performance of Model 2 decreases, with the average RMSE dropping to 0.174, while the average RMSE and MAE of the proposed method are 0.096 and 0.078, respectively. Secondly, Model 3, due to the loss of its dynamic graph structure, cannot capture deep correlation features, resulting in significant fluctuations in its prediction accuracy; the average RMSE and MAE are only 0.178 and 0.150, respectively.

[0235] Figure 14This is a schematic diagram illustrating the multi-level spatiotemporal joint knowledge distillation effect in this embodiment of the invention. Figure (a) shows the dynamic performance of the student model and the teacher model in terms of feature alignment, knowledge transfer, and performance gap during the training process. It can be seen that the feature alignment index is low in the early stages of training but rises steadily as training progresses, eventually approaching 1. This indicates that the student model gradually learns to imitate the intermediate feature representations of the teacher model, reflecting the effective guidance of the distillation mechanism on the deep information of the bearing data. Simultaneously, the knowledge transfer index rises rapidly in the first few epochs and fluctuates stably at a high level, indicating that the teacher model successfully transferred rich knowledge to the student model in the early stages of training. The performance gap generally shows a downward trend during training. Although there are slight fluctuations in the middle stages, the overall performance of the student model gradually approaches that of the teacher model, indicating that the student model possesses excellent generalization ability while maintaining model simplicity during the distillation process. Figure (b) further quantitatively demonstrates the comparison of parameter quantities and parameter similarity between the teacher and student models. The teacher model has approximately 120,000 parameters, while the student model has only about 4% of that. However, their performance similarity is as high as 0.96, indicating that the student model can maintain a knowledge structure that is highly similar to the teacher model while significantly compressing the parameters.

[0236] Figure 15 This diagram illustrates the impact of partial channel data from the PHM2012 bearing dataset on the model's prediction performance in this embodiment of the invention. The heatmaps for Bearing 1-6 show that the model initially focuses more on traditional time-domain features such as mean, variance, and peak vibration, while gradually shifting towards frequency-domain or mixed-domain features such as spectral flatness and energy ratio in later stages. This indicates that the model relies on basic statistics in the early degradation phase, but depends more on spectral features for discrimination during the accelerated degradation phase. In Bearing 2-4 samples, features such as spectral flatness, energy ratio, and impulse factor show significant contributions at multiple time steps, indicating that the spectral expression of fault symptoms is more prominent in this sample.

[0237] Figure 16 This is a schematic diagram illustrating the impact of partial channel data from the XJTU-SY bearing dataset on the prediction performance of this model in an embodiment of the invention. Bearing 1-4 shows strong responses to indicators such as kurtosis and fractal dimension at multiple time steps, suggesting that there may be slight surface defects in the early stages. Bearing 2-5, on the other hand, shows significantly enhanced sensitivity to features such as peak vibration and variance in the later stages, indicating a clear trend of accelerated failure.

[0238] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0239] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0240] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting and quantifying the remaining service life of bearings using a dynamic graph convolutional distillation network, characterized in that, include: Vibration signals are collected throughout the entire life cycle of the bearing, and the vibration signals are converted into characteristic channel data. A dynamic graph is constructed based on the feature channel data. The dynamic graph includes nodes and edges. The nodes represent features in the feature channel data, and the edges represent the relationships between features. The dynamic graph is updated using an adaptive adjacency matrix, which is used to learn deep interaction relationships between nodes. Based on the dynamic graph, spatiotemporal joint features are extracted through the teacher model, and the spatiotemporal joint features are transferred to the student model through a multi-level knowledge distillation strategy. The remaining service life of the bearing is predicted using the student model, and the uncertainty of the prediction results is quantified based on the Bayesian method to obtain the confidence interval of the bearing's RUL.

2. The method according to claim 1, characterized in that, Converting the vibration signal into characteristic channel data includes: The vibration signal is subjected to time-domain and frequency-domain feature extraction to obtain a multi-dimensional feature vector; The multidimensional feature vectors are arranged in chronological order to construct a feature channel data matrix.

3. The method according to claim 1, characterized in that, The construction of the dynamic graph includes: Sliding window sampling is performed on the feature channel data to obtain window data; Perform a linear transformation on the window data and project it into a new feature space; Calculate the similarity matrix between the transformed features, and construct an initial adjacency matrix based on the similarity matrix; The initial adjacency matrix is ​​normalized to obtain the time dependency matrix; A gating mechanism is introduced to weight the time dependency matrix to obtain a dynamic adjacency matrix; A dynamic graph is constructed based on the dynamic adjacency matrix.

4. The method according to claim 3, characterized in that, The update of the adaptive adjacency matrix includes: The similarity matrix is ​​normalized and nonlinearly mapped to obtain a normalized adjacency matrix; The normalized adjacency matrix is ​​applied to the dynamic graph to update the connection weights between nodes.

5. The method according to claim 1, characterized in that, The teacher model extracts spatiotemporal joint features including: The dynamic graph is input into a graph convolutional network to extract spatial features; Multiple graph convolutional layers are stacked along the time dimension to extract temporal features; By integrating the aforementioned spatial and temporal features, a spatiotemporal joint feature is obtained.

6. The method according to claim 5, characterized in that, The multi-level knowledge distillation strategy includes: Align the intermediate layer features of the teacher model and the student model, and calculate the structural consistency loss; Sparsity constraints are imposed on the adjacency matrices of the teacher and student models, and the sparsity loss is calculated. The local structures of the teacher model and the student model are compared, and the local structure loss is calculated. All losses are combined to construct a total distillation loss function, which is used to guide student model training.

7. The method according to claim 6, characterized in that, The total distillation loss function includes: The difference loss between the outputs of the teacher model and the student model; The weighted sum of structural consistency loss, sparsity loss, and local structural loss.

8. The method according to claim 1, characterized in that, The uncertainty quantification includes: Statistical analysis was performed on multiple prediction results to obtain the prediction mean and cognitive uncertainty. Based on the random uncertainty of predictive residual calculation; By integrating the aforementioned cognitive uncertainty and accidental uncertainty, a dual-correlated uncertainty is obtained, and a confidence interval is constructed.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images