Bearing remaining service life prediction method based on data reproduction multi-source information fusion
By using a multi-source information fusion method based on data reproduction, the problems of insufficient data and model adaptability in bearing remaining service life prediction were solved, achieving efficient prediction without target domain data and improving the model's adaptability and performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-16
- Publication Date
- 2026-03-13
AI Technical Summary
Existing bearing remaining service life prediction models lack labeled samples in real industrial scenarios and rely on target domain data for training, making it difficult to adapt to high-load and rapidly updated industrial environments and unable to efficiently utilize new data.
A multi-source information fusion method based on data reproduction is adopted. By using sliding window preprocessing, bidirectional gating mechanism and Mix-up strategy, combined with data reproduction mechanism, a multi-source domain prediction model is constructed to achieve general capture and information fusion of bearing degradation information.
Without requiring target domain data, it enables rapid prediction of bearing remaining service life, improves model adaptability and performance, and has good scalability and practical value.
Smart Images
Figure CN121658802A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of rolling bearing remaining service life prediction technology, specifically to a bearing remaining service life prediction method based on multi-source information fusion through data reproduction. Background Technology
[0002] Bearings, as critical components of rotating machinery, are prone to wear, fatigue spalling, or lubrication failure. These issues can lead to sudden malfunctions, causing production interruptions, safety accidents, and significantly increased maintenance costs. To improve equipment reliability and operational efficiency, bearing remaining useful life (RUL) prediction technology based on condition monitoring has been widely researched and applied. However, RUL prediction models typically rely on large amounts of high-quality target domain data to ensure prediction accuracy, which is often difficult to meet in real-world industrial scenarios.
[0003] In real-world industrial scenarios, bearings often operate under harsh conditions such as high loads and high speeds, making it difficult to collect stable operational data. Furthermore, RUL (Real User Label) tags are not naturally present, resulting in significant annotation costs. This often leads to a lack of labeled samples for target bearings. To address these issues, transfer learning (TL) is widely used to transfer knowledge from the source domain (bearings) to the target domain (bearings), reducing the need for labeled samples. However, TL still relies on target domain data for training, making it difficult to implement in scenarios where available data is scarce, such as during the initial stages of operation or with entirely new equipment. Therefore, researchers have further proposed domain generalization strategies that do not require the target domain to participate in training.
[0004] Domain generalization, as a solution for bearing RUL prediction under unknown data, has made some progress in the field of bearing RUL prediction. However, existing work still has the following limitations. First, current methods rely on building strong feature extractors to mine temporal features, ignoring the diversity and non-steady-state characteristics of the bearing degradation process, thus limiting the adaptability of the model. Second, most existing models are static structures, relying on training from scratch to fuse new source domain information, which cannot efficiently utilize newly added source domains or incremental data for continuous optimization, limiting their application in industrial contexts where equipment is rapidly updated and data is gradually accumulated.
[0005] To address the aforementioned issues, this invention proposes a bearing remaining service life prediction method based on multi-source information fusion through data reproduction. First, a sliding window approach is used to extract time and frequency domain features, followed by standardization to preprocess the source domain bearing vibration signals. After preprocessing, a bidirectional gating mechanism is used to mine the temporal features in the bearing's time and frequency domains, and a mix-up approach is employed to fuse information between different bearings, capturing the consistency of bearing degradation information. Subsequently, by gradually fusing newly added bearing information through data reproduction, a multi-bearing general degradation information fusion is achieved, enabling model construction. Summary of the Invention
[0006] To address bearing remaining service life (RUL) prediction when target data is unknown, this invention proposes a bearing RUL prediction method based on Multi-source Domain Mix-up Transfer (MDMT) using data reproduction, achieving bearing RUL prediction even with unknown data. First, a Mix-up-based dual-source domain bearing life prediction model is developed to capture general degradation information. Then, combined with data reproduction, general degradation information from multiple bearings is fused to enhance the model's accuracy in predicting bearing life under new scenarios with unknown data.
[0007] The technical solution adopted by this invention to solve its technical problem is:
[0008] A method for predicting the remaining service life of bearings based on multi-source information fusion through data reproduction, the method comprising the following steps:
[0009] Step 1) Obtain the bearing remaining service life prediction dataset
[0010] The data used in this invention comes from a publicly available dataset—the Xi'an Jiaotong University Bearing Accelerated Life Test Dataset (XJTU-SY). This dataset records vibration signals, temperature data, and life termination information of multiple industrial bearings through accelerated aging experiments, and is divided into training and test sets according to different operating conditions.
[0011] Step 2) Bearing data preprocessing
[0012] Preprocessing of bearing time-series vibration data based on time-domain and frequency-domain feature extraction, standardization, and sliding window strategies generates an input signal matrix to reduce input dimensionality, decrease vibration data noise, and improve the time-series characteristics of the data.
[0013] Step 3) Construction of the dual-source domain prediction model
[0014] Based on the bidirectional gating mechanism, the time-series features in the bearing time-frequency domain are mined, and the information fusion between different bearings is completed by combining Mix-up. The mixing ratio and model parameters are optimized by loss iteration to construct a dual-source domain prediction model.
[0015] Step 4) MDMT Model Construction
[0016] Multi-source information fusion is achieved by combining a data reproduction mechanism. Selected anchor samples are continuously input, and new source domain samples are introduced for further training. By iteratively updating the mixing ratio between training samples, multi-source information fusion is achieved, and an MDMT model is constructed.
[0017] Step 5) MDMT Performance Evaluation
[0018] The test set is input into MDMT to obtain the prediction results and compare them with the actual remaining lifetime values. The performance of the model is quantitatively evaluated using metrics such as mean absolute error (MAE), root mean square error (RMSE), and root mean squared error (R²).
[0019] Furthermore, the process of step 3) is as follows:
[0020] Step 3.1: Construct a feature extractor based on a bidirectional long short-term memory network, and mine temporal features through a three-layer stacked structure. Select... Construct a dual-source domain prediction model and randomly generate the initial mixing ratio. Constructing a hybrid domain As shown in the formula: ;
[0021] in, They represent in time The signal input matrix of the bearing, Indicates in time The bearing's RUL label, Indicates the mixing ratio. They represent Features extracted by the feature extractor Indicates the characteristics after mixing. Indicates the mixed label;
[0022] Step 3.2: Iteratively adjust the mixing ratio: Measured using MK-MMD distance. , and , Distance in Hilbert space , Then calculate the weighting factors. Finally, combined with weighting factors and the mixing rate of round e-1 Iteration is used to obtain the mixing ratio by introducing a uniform distribution and random perturbations. As training progresses, in the feature space Gradually from Towards The general information is captured upon proximity, and the specific calculation formula is as follows: ;
[0023] in, It's a temperature parameter, guiding... Towards Proximity, weighting factor Applying it in the form of a moving average , Indicates uniform distribution. Indicates the range of the disturbance. It is the total number of iterations. It is the current iteration batch;
[0024] Step 3.3: Design a regressor by combining a linear layer, ReLU activation function, and Dropout mechanism to output the RUL prediction result of the bearing. A multi-task joint loss function is also introduced; the specific learning objective function is shown below: ;
[0025] in, Indicates feature extractor, Indicates the regressor. Represents the MSE loss function. This represents the features extracted from the input signal matrix of the training set. This represents the features extracted from the mixed-domain signal matrix. For mixed part loss, Refers to the regression loss of the training set. They represent The regression loss, For the total loss, It is a trade-off parameter.
[0026] Furthermore, the process of step 4) is as follows:
[0027] Input training set and dual-source domain prediction models; As anchor samples are repeatedly reproduced, The dual-source domain bearing life prediction model is continuously trained by sequentially inputting data; based on the mixing ratio optimization, the model is then... Degradation information is migrated to anchor samples The general degradation information enables the fusion of multi-source information; when input By maintaining anchor point samples The training objective function is defined as follows: (The function remains unchanged and continues to learn multi-source general degradation information.) ;
[0028] in, These are the joint losses of anchor sample bearing S1 and newly added sample bearings S2, S3, and S4; the model sequentially applies the above loss functions. Minimize optimization to achieve multi-source information fusion.
[0029] The technical concept of this invention is as follows:
[0030] To address the lack of modeling data in the early stages of equipment operation, this invention proposes a bearing remaining service life prediction method based on multi-source information fusion through data reproduction, belonging to the field of bearing remaining service life prediction without target data. This invention develops a multi-source information fusion method based on a mix-up strategy. First, the raw data is preprocessed to improve the temporal characteristics of the bearing data. After preprocessing, general degradation information is captured using the mix-up strategy. Modeling of multi-source general degradation information is achieved based on data reproduction. Finally, the vibration data of the target bearing, after preprocessing, is input. This invention constructs a model in the new scenario of unknown data in the early stages of operation, realizing the prediction of the remaining service life of industrial bearings.
[0031] The beneficial effects of this invention are mainly reflected in the following aspects: it can quickly predict the remaining service life of bearings without the participation of target samples, thereby improving the adaptability of the model in actual deployment; it can continuously integrate degradation information from newly added source domains to improve the performance of the model, and has good scalability and practical value. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the method of the present invention;
[0033] Figure 2 The graph shows the bearing life prediction results. It contains four sub-graphs, which correspond to the remaining service life prediction results of test sets bearing2_2, bearing2_3, bearing2_4, and bearing2_5, respectively.
[0034] Figure 3 Box plot of absolute error for bearing life prediction. Detailed Implementation
[0035] The present invention will now be further described with reference to the accompanying drawings.
[0036] Reference Figures 1-3 A method for predicting the remaining service life of bearings based on the fusion of multi-source information from data reproduction, the method comprising the following steps:
[0037] (1) Obtain the bearing remaining service life prediction dataset
[0038] The data used in this invention comes from a publicly available dataset—the Xi'an Jiaotong University Bearing Accelerated Life Test Dataset (XJTU-SY). This dataset records vibration signals, temperature data, and life termination information of multiple industrial bearings through accelerated aging experiments to obtain the original dataset. Two typical operating conditions were selected for the experiment. In operating condition 1 (speed 2100 rpm, radial force 12 kN), five bearings were used as the training set. In operating condition 2 (speed 2250 rpm, radial force 11 kN), four bearings were used as the test set (bearing 2_1 was excluded because it exhibited abrupt degradation, with its vibration signal first decreasing and then suddenly diverging and failing, making its prediction results unrepresentative). ;
[0039] (2) Bearing data preprocessing
[0040] For the original vibration signal of the bearing, 23 statistical features were extracted using time-domain and frequency-domain analysis methods, including key indicators such as mean, standard deviation, skewness, kurtosis, maximum value, and spectral energy. Standardization was then used to reduce data dimensionality, decrease data noise, and improve time-series characteristics. Subsequently, a sliding window strategy was applied with a window size of 20 to generate the input signal matrix. B represents the batch size, T represents the time scale, and C represents the number of features. The corresponding label y represents the RUL of the last time, and the bearing training samples are constructed. ,in Indicates the time of bearing sample. Different bearing symbols are used to represent them. This indicates the number of samples within a single bearing. Indicates the number of bearings.
[0041] (3) Construction of dual-source domain prediction model
[0042] Step 3.1: Select A feature extractor based on a bidirectional long short-term memory network is constructed, and temporal features are mined through a three-layer stacked structure. An initial mixing ratio is randomly generated. Constructing a hybrid domain As shown in the formula: ;
[0043] in, They represent in time The signal input matrix of the bearing, Indicates in time The bearing's RUL label, Indicates the mixing ratio. They represent Features extracted by the feature extractor Indicates the characteristics after mixing. Indicates the mixed label;
[0044] Step 3.2: Iteratively adjust the mixing ratio: Measured using MK-MMD distance. , and , Distance in Hilbert space , Then calculate the weighting factors. Finally, combined with weighting factors and the mixing rate of round e-1 Iteration is used to obtain the mixing ratio by introducing a uniform distribution and random perturbations. As training progresses, in the feature space Gradually from Towards The general information is captured upon proximity, and the specific calculation formula is as follows: ;
[0045] in, It's a temperature parameter, guiding... Towards Proximity, weighting factor Applying it in the form of a moving average , Indicates uniform distribution. Indicates the range of the disturbance. It is the total number of iterations. This is the current iteration batch.
[0046] Step 3.3: Design a regressor by combining a linear layer, ReLU activation function, and Dropout mechanism to output the RUL prediction result of the bearing. A multi-task joint loss function is also introduced; the specific learning objective function is shown below: ;
[0047] in, Indicates feature extractor, Indicates the regressor. Represents the MSE loss function. This represents the features extracted from the input signal matrix of the training set. This represents the features extracted from the mixed-domain signal matrix. For mixed part loss, Refers to the regression loss of the training set. They represent The regression loss, For the total loss, It is a trade-off parameter.
[0048] (4) Construction of bearing prediction model based on multi-source information fusion
[0049] Select anchor point sample Then enter in sequence A dual-source domain bearing remaining service life prediction model is continuously trained to provide multi-source degradation information. Based on mixing ratio optimization, Bearing degradation information is migrated to anchor point samples. The model learns general degradation information from multiple sources to achieve multi-source information fusion. When a new sample is input, the model continuously learns general degradation information from multiple sources while keeping the anchor sample unchanged, thus completing the construction of the MDMT model.
[0050] The specific training process is as follows: ;
[0051] in, These are the joint losses for anchor sample bearing S1 and newly added sample bearings S2, S3, and S4, respectively. The model sequentially applies the above loss functions. Minimize optimization to achieve multi-source information fusion.
[0052] (5) MDMT performance evaluation
[0053] The test set signal matrix is input into a multi-source information fusion lifetime prediction model to predict the remaining service life of the target bearing. The model's predictive performance is then systematically evaluated using known real RUL labels. For accurate evaluation, tests are conducted on four different target bearings, and the model's performance is quantitatively assessed using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R².
[0054] As shown in Table 1 and Figure 2 As shown, compared to other methods, the MDMT method shows improvements in all three metrics. Although CCIRL performs best overall among existing methods, MDMT still achieves error reductions of 10.74% and 11.59% and accuracy improvements of 10.86% in MAE, RMSE, and R², respectively, indicating that it outperforms existing methods in all aspects. Figure 3 MDMT exhibits the smallest bin size and the lowest median error, indicating higher prediction accuracy at most time points. Although a certain number of outliers exist, this is attributed to MDMT's smaller overall prediction error, more concentrated distribution, and lower outlier identification threshold, while other methods have larger overall errors, making it difficult for them to reach the outlier identification limit due to high bias.
[0055] Table 1 Comparison of Prediction Results ;
[0056] This invention achieves feature fusion of multi-source bearing time-series data based on a mix-up strategy; simultaneously, it constructs a multi-source information fusion prediction model by combining a data reproduction mechanism. The synergistic effect of these mechanisms enables the proposed model to predict bearing life in new scenarios where data is unknown.
[0057] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for predicting the remaining service life of bearings based on multi-source information fusion through data reproduction, characterized in that, Includes the following steps: Step 1) Obtain the bearing remaining service life prediction dataset: Using a publicly available dataset, the vibration signals, temperature data, and end-of-life information of multiple industrial bearings were recorded through accelerated aging experiments to form the original dataset. The original dataset was then divided into a training set and a test set, where the training set data included... The test set data includes ; Step 2) Bearing data preprocessing: The original data is preprocessed based on time-domain and frequency-domain feature extraction, standardization and sliding window strategies to generate an input signal matrix, thereby reducing the input dimension, reducing vibration data noise and improving the temporal characteristics of the data. Step 3) Construction of the dual-source domain prediction model: Based on the bidirectional gating mechanism, the time-series features in the time and frequency domain features of bearings are mined, and the information fusion between different bearings is completed by combining Mix-up. The mixing ratio and model parameters are optimized by loss iteration to construct a dual-source domain prediction model. Step 4) MDMT Model Construction: Multi-source information fusion is achieved based on the data reproduction mechanism. Anchor point samples are continuously input, and new source domain samples are introduced for further training. By iteratively updating the mixing ratio between training samples, multi-source information fusion is achieved, and the MDMT model is constructed. Step 5) Bearing remaining service prediction performance assessment: The test set is input into MDMT to obtain the prediction results and compare them with the actual remaining lifetime values. The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used to quantitatively evaluate the model performance.
2. The bearing remaining service life prediction method based on multi-source information fusion through data reproduction as described in claim 1, characterized in that: In step 3), a mix-up strategy is used to extract general degradation information, and the process is as follows: 3.1) Select Construct a dual-source domain prediction model; randomly generate initial mixing ratios. Constructing a hybrid domain The specific formula is as follows: ； in, They represent in time The signal input matrix of the bearing, Indicates in time The bearing's RUL label, Indicates the mixing ratio. They represent Features extracted by the feature extractor Indicates the characteristics after mixing. Indicates the mixed label; 3.2) Iterative adjustment of mixing ratio: measured using MK-MMD distance. , and , Distance in Hilbert space , Then calculate the weighting factors. Finally, combined with weighting factors and the mixing rate of round e-1 Iteration is used to obtain the mixing ratio by introducing a uniform distribution and random perturbations. As training progresses, in the feature space Gradually from Towards The general information is captured upon proximity, and the specific calculation formula is as follows: ； in, It's a temperature parameter, guiding... Towards Proximity, weighting factor Applying it in the form of a moving average , Indicates uniform distribution. Indicates the range of the disturbance. It is the total number of iterations. It is the current iteration batch; The specific learning objective function of the model is shown below: ； in, Indicates feature extractor, Indicates the regressor. Represents the MSE loss function. This represents the features extracted from the input signal matrix of the training set. This represents the features extracted from the mixed-domain signal matrix. For mixed part loss, Refers to the regression loss of the training set. They represent The regression loss, For the total loss, It is a trade-off parameter.
3. The construction of the multi-source information fusion prediction model according to claim 1, characterized in that: In step 4), the fusion of multi-source information is achieved based on the data reproduction mechanism, and the process is as follows: 1) Input training set and dual-source domain prediction models; As anchor samples are repeatedly reproduced, The dual-source domain bearing life prediction model is continuously trained by sequentially inputting data; based on the mixing ratio optimization, the model is then... Degradation information is migrated to anchor samples The general degradation information enables the fusion of multi-source information; when input By maintaining anchor point samples The training objective function is defined as follows: (The function remains unchanged and continues to learn multi-source general degradation information.) ； in, These are the joint losses of anchor sample bearing S1 and newly added sample bearings S2, S3, and S4; the model sequentially applies the above loss functions. Minimize optimization to achieve multi-source information fusion.