Fan gearbox bearing residual life prediction method based on physical degradation mapping and multi-feature fusion
By constructing a physical degradation simulation model and using a multi-feature fusion method, the accuracy and stability issues of wind turbine gearbox bearing life prediction were solved, and reliable prediction of wind turbine gearbox bearings was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEAST DIANLI UNIVERSITY
- Filing Date
- 2026-01-08
- Publication Date
- 2026-04-17
AI Technical Summary
Existing technologies cannot effectively reflect the true physical wear patterns of wind turbine gearbox bearings, and are affected by factors such as wind speed, power, and temperature, resulting in inaccurate and unstable life predictions.
By constructing a physical degradation simulation model, extracting the simulated root mean square value and kurtosis, establishing a degradation mapping function, converting real vibration characteristics into a health index with physical meaning, and combining it with a long short-term memory network for multi-feature fusion prediction.
It achieves physical alignment of bearing wear, improves the accuracy and stability of prediction, is applicable to various types of wind turbine gearbox bearings, and has good generalization ability and engineering application value.
Smart Images
Figure CN121880818A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent operation and maintenance technology of wind power equipment, specifically involving a method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion. Background Technology
[0002] As a critical transmission component connecting the main shaft and the generator, the gearbox of a wind turbine generator bears complex loads over long periods, including cyclic alternating loads, torque impacts, fluctuations in lubrication conditions, and changes in ambient temperature. Fatigue spalling, pitting, or wear propagation in the gearbox bearings directly leads to turbine shutdown and maintenance, resulting in significant power generation losses and operation and maintenance costs. Therefore, health monitoring and remaining life prediction (RUL) of gearbox bearings have become an important technological direction for intelligent operation and maintenance in the wind power industry.
[0003] The following types of lifetime prediction technologies are currently mainly used in the wind power sector: I. Methods based on single-feature trends; Traditional methods typically use the root mean square (RMS), peak value, kurtosis, and root square amplitude of vibration signals to describe the degradation trend of equipment, and predict the remaining life through linear regression, polynomial fitting, or exponential models. However, these methods have significant limitations: (1) the operating conditions of wind turbines fluctuate greatly (wind speed, power, and load change randomly), resulting in non-monotonic, abrupt, and fluctuating characteristic curves; (2) a single feature cannot simultaneously characterize the wear energy changes and local defect impact processes of the bearing; (3) the degradation trend is significantly affected by sensor noise, structural resonance, and lubrication conditions, making it difficult to form a stable degradation pattern; (4) in the early degradation stage, the characteristic changes are not obvious, which can easily lead to misjudgment or premature prediction. Therefore, trend methods based on single features are difficult to meet the high reliability life prediction requirements of wind turbine gearbox bearings.
[0004] II. Fault diagnosis methods based on frequency domain and envelope spectrum; Frequency domain features (such as BPFO / BPFI / BSF) and envelope spectrum techniques are widely used in bearing fault diagnosis. However, they also face the following problems in life prediction: (1) they are sensitive to noise, as the background noise of wind turbine gearboxes often covers the fault frequency; (2) the energy characteristics will show "saturation" or "decline" after the fault intensifies, and the degradation trend is discontinuous; (3) the envelope spectrum is more suitable for fault identification and not suitable for long-term trend prediction. Therefore, it is difficult to accurately model the remaining life time by relying solely on frequency domain analysis.
[0005] III. Data-driven deep learning-based lifetime prediction; In recent years, neural networks such as LSTM, CNN, and GAN have been increasingly used for life prediction. However, their application in actual wind farms still has the following shortcomings: (1) Deep learning models rely on a large amount of full-cycle data that has degraded to failure, while actual wind turbines rarely have complete degradation data; (2) Without physical constraints, the model may learn operating noise rather than the essence of degradation, leading to unstable predictions; (3) The input features (such as RMS and Kurtosis) themselves are noisy and non-monotonic, making it difficult for the model to converge; (4) Different wind turbine models and different bearing operating conditions vary greatly, making it difficult for the model to transfer and resulting in weak generalization ability; (5) The prediction results lack interpretability and are difficult to meet the standards of actual industrial applications. Therefore, pure data-driven models still have significant limitations in the task of predicting the life of wind turbine gearbox bearings.
[0006] Fourth, existing technologies lack a degenerate expression method that achieves "physical consistency"; A key problem currently exists in the technology: the degradation process of the real vibration signal cannot directly reflect the physical wear law of the bearing. This is mainly manifested in the following ways: (1) the vibration characteristics fluctuate over time, while bearing wear is actually a monotonic cumulative process; (2) the real signal is affected by strong disturbances such as wind speed, power, and temperature changes; (3) the "real degradation degree" cannot be directly extracted from the real signal, resulting in a large deviation in life prediction; (4) the feature space has no physical meaning, and the model has poor interpretability.
[0007] Therefore, the industry still lacks a health index that can reflect the actual physical degradation process of bearings, suppress operating noise and environmental changes, and be directly used for life prediction. This is a major technical challenge in the field of intelligent operation and maintenance of wind turbines. Summary of the Invention
[0008] To address the aforementioned shortcomings, this invention aims to solve the core problem in existing technologies where the degradation process of actual equipment vibration signals cannot directly reflect the physical wear patterns of bearings. Specifically, the purpose of this invention is to overcome the deficiencies of existing models based on single vibration characteristics or purely data-driven models, and to achieve reliable, stable, and physically consistent prediction of the remaining life of wind turbine gearbox bearings.
[0009] To achieve the above objectives, the present invention provides the following solution: a method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion, comprising the following steps: S1. Obtain the vibration signal sequence of the wind turbine gearbox bearing, and extract the true root mean square value representing energy characteristics and the true kurtosis representing impact characteristics from the vibration signal; S2. Construct a degradation simulation model based on bearing structural parameters and operating conditions, and obtain simulated degradation signals; and extract simulated root mean square value and simulated kurtosis based on the simulated degradation signals; S3. Construct a physical degradation model, and based on the simulated root mean square value, the simulated kurtosis, and the physical degradation model, obtain the degradation mapping function; S4. Input the true root mean square value and the true kurtosis into the degradation mapping function to obtain the true degradation value. Perform monotonic processing on the true degradation value to obtain the health index. S5. Based on the health index, the true root mean square value, and the true kurtosis, a multidimensional degradation feature sequence is obtained, and the remaining life of the wind turbine gearbox bearing is obtained based on the multidimensional degradation feature sequence.
[0010] More preferably, the root mean square value is calculated as follows: ; In the formula, N Indicates the total number of vibration data sequences; The kurtosis is calculated as follows: ; In the formula, Mean .
[0011] More preferably, the physical degradation model includes: ; In the formula, d ( k ) represents the simulated degradation signal sequence.
[0012] More preferably, the health index includes: ; In the formula, This represents the actual degradation value.
[0013] More preferably, S5 includes the following steps: S51. Based on the health index, the true root mean square value, and the true kurtosis, a multidimensional degradation feature sequence is obtained: ; S52, using a length of L The time window constructs a multi-step time series from the multidimensional degenerative feature sequence: ; S53. Input the multi-step time series data into a Long Short-Term Memory (LSTM) network to obtain the remaining bearing life: ; In the formula, k This indicates the sample number at the current moment.
[0014] More preferably, the loss of the Long Short-Term Memory network includes: ; In the formula, M This represents the total number of training samples; Indicates the predicted value; y i Represents the actual value.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Physical alignment of the real degradation process is achieved, improving the accuracy of degradation characterization.
[0016] This invention constructs a physical degradation template by simulating degradation data and uses a degradation mapping function to convert real vibration characteristics into a degradation degree expression with physical meaning. This allows the real signal to characterize the wear and defect propagation trend of the bearing in a unified physical coordinate system. This method effectively overcomes the non-monotonicity problem caused by noise and operating conditions affecting the real characteristics, making the degradation curve more stable and accurate.
[0017] (2) A monotonic, interpretable health index was constructed to improve the stability of prediction.
[0018] This invention constructs the health index HIreal through monotonication, ensuring it increases strictly monotonically over time, consistent with the irreversible physical wear of bearings. This health index possesses a clear physical meaning and can be directly used to assess equipment health, avoiding the drawbacks of traditional features that exhibit large fluctuations and unclear trends, leading to unstable model predictions. The resulting prediction curve is smoother and more reliable.
[0019] (3) Multi-feature fusion enhances the degradation modeling capability and improves the robustness of the model.
[0020] This invention comprehensively utilizes four types of features—HIreal (health index), Dsim (physical degradation template), normalized RMS (normalized RMS), and normalized Kurtosis (normalized Kurtosis)—to fuse and express bearing wear, impact, energy changes, and prior physical information. The synergistic representation of the degradation process by these multi-source features improves the model's robustness to complex operating conditions and noise interference, making the prediction results more stable and applicable.
[0021] (4) Improve lifetime prediction accuracy by using LSTM-based time series modeling.
[0022] This invention employs a Long Short-Term Memory (LSTM) network to model multidimensional time series features, effectively capturing the time dependencies and long-term trends in bearing degradation. Compared to traditional regression methods, LSTM can more accurately describe the dynamic changes in degradation, making the remaining lifetime prediction results closer to the actual degradation trajectory and improving the accuracy of lifetime prediction.
[0023] (5) It is applicable to various types of wind turbine gearbox bearings and has good engineering application value.
[0024] This invention does not rely on a specific model or operating condition. It achieves data alignment across different operating conditions through a physical degradation template, demonstrating excellent generalization ability. This method is applicable to predicting the remaining life of various types of bearings, such as main bearings in wind turbine gearboxes and high-speed shaft bearings. It can be quickly applied to actual wind farm operation and maintenance, providing a reliable basis for fault early warning, maintenance decisions, and spare parts management. Attached Figure Description
[0025] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 This is a flowchart illustrating the method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion provided in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram comparing the simulated and actual RMS values in Example 1; Figure 3 This is a schematic diagram comparing the simulated and actual Kurtosis values in Example 1; Figure 4 This is a schematic diagram illustrating the mapping relationship between the equipment degradation index and the health index in Example 1; Figure 5 This is a schematic diagram of the multidimensional feature vector curve in Example 1; Figure 6 This is a schematic diagram comparing the theoretical remaining lifetime value and the actual predicted value in Example 1. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0029] Example 1: like Figure 1As shown, this embodiment provides a method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion. The method steps provided by the present invention are described in detail with a typical example of a large wind turbine gearbox as the monitoring object.
[0030] S1. Obtain the vibration signal sequence of the wind turbine gearbox bearing, and extract the true root mean square value representing the energy characteristics and the true kurtosis representing the impact characteristics from the vibration signal.
[0031] In this embodiment, a vibration acceleration sensor is placed near the bearing of the wind turbine gearbox to continuously collect vibration signals during the bearing's operation at a fixed sampling frequency. For each sequence of collected vibration data... After preprocessing, its true root mean square value is calculated. The root mean square (RMS) value is used to characterize the vibration energy level, and its calculation method is as follows: ; In the formula, N This represents the total number of vibration data sequences.
[0032] Simultaneously calculate the true kurtosis ( This is used to characterize the impact components and waveform sharpness, introducing a mean value. Using variance as an intermediate quantity, the kurtosis is calculated as follows: .
[0033] S2. Construct a degradation simulation model based on bearing structural parameters and operating conditions, and obtain simulated degradation signals; then extract simulated root mean square value and simulated kurtosis based on the simulated degradation signals.
[0034] To introduce physically meaningful degradation priors, this embodiment constructs a bearing degradation simulation model based on the structural parameters, operating loads, and rotational speed of the wind turbine gearbox bearing. This model simulates the complete degradation process of the bearing from its initial healthy state to its final failure state, resulting in a set of simulated degradation signals. The simulated root mean square value is calculated on the simulated degradation signals using the same method as in S1. With simulated kurtosis The distributions of the root mean square value and kurtosis in the real and simulated data are compared as follows: Figure 2 and Figure 3 As shown.
[0035] S3. Construct a physical degradation model and obtain the degradation mapping function based on the simulated root mean square value, simulated kurtosis and physical degradation model.
[0036] After smoothing and normalizing the simulated root mean square value and simulated kurtosis, a sequence of simulated degradation signals that increases with time is constructed. d ( kBased on this, a physical degradation model is constructed using the cumulative maximum form: .
[0037] Subsequently Linear normalization is performed over the entire range to make it fall within the [0,1] interval, which is used to characterize the relative degree of degradation of the bearing from healthy to failure during the simulated degradation process.
[0038] To ensure that real features correspond to the simulated degradation template, this embodiment constructs a degradation mapping function using simulated root mean square values and simulated kurtosis. f By using simulated features as input and simulated degradation templates as output, a multinomial regression model is employed to establish a mapping relationship between the feature space and the degradation space: ; Taking third-order polynomial regression as an example, the mapping form is: ; In the formula, and These represent the root mean square value and kurtosis, respectively. This represents the coefficients obtained from the fitting. This mapping function allows any true feature to be projected onto the physically degenerate coordinate system.
[0039] S4. Input the true root mean square value and the true kurtosis into the degradation mapping function to obtain the true degradation value. Perform monotonic processing on the true degradation value to obtain the health index.
[0040] The actual degradation values are as follows: .
[0041] Since real-world degradation sequences are typically significantly affected by noise and operating condition disturbances, this invention constructs a health index through monotonicity processing: .
[0042] and through normalization Mapping to the [0,1] interval gives the health index both physical meaning and conforms to the objective law of irreversible degeneration. The relationship between the simulated degeneration curve, the real degeneration mapping, and the monotonic health index is as follows: Figure 4 As shown.
[0043] S5. Based on the health index, the true root mean square value, and the true kurtosis, a multidimensional degradation feature sequence is obtained, and the remaining life of the wind turbine gearbox bearing is obtained based on the multidimensional degradation feature sequence.
[0044] To comprehensively characterize the degradation process of gearbox bearings, this embodiment uses a health index. Physical degradation template Together with the normalized true root mean square value and the true kurtosis, they form a multidimensional degenerate feature sequence, such as Figure 5 As shown: .
[0045] And using a length of L The time window constructs the feature vector sequence into a multi-step time series input: ; in L A value of 40 is used to capture long-term dependencies in bearing degradation.
[0046] The multidimensional degradation feature sequence is input into a Long Short-Term Memory (LSTM) network, which utilizes its internal memory gating structure to capture the degradation trend. The remaining bearing life is: ; In the formula, k This indicates the sample number at the current moment.
[0047] Then, normalize it to obtain the supervision signal: .
[0048] By minimizing the predicted value Compared with the true value Model training is performed using the mean squared error (MSE) between the two sides. ; In the formula, M This represents the total number of training samples.
[0049] Once training is complete, the latest sequence can be input into the LSTM model to obtain the predicted remaining lifetime. The theoretical remaining life value and the actual predicted value are as follows: Figure 6 As shown.
[0050] After the model infers from the real-time input multidimensional feature sequence, it can obtain the predicted remaining life of the wind turbine gearbox bearing. The prediction result shows a monotonically decreasing trend over time and can be used for equipment condition assessment, maintenance scheduling, and spare parts management. When the temperature drops below the set threshold, maintenance or replacement operations can be performed in advance as needed to reduce the risk of wind turbine shutdown and ensure the stability of equipment operation.
[0051] Through the closed-loop operation of the above-described embodiments, the present invention achieves deep coupling between equipment health prediction and inventory management, ultimately minimizing inventory costs while ensuring high equipment availability.
[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A fan gearbox bearing residual life prediction method based on physical degradation mapping and multi-feature fusion, characterized in that, Includes the following steps: S1. Obtain the vibration signal sequence of the wind turbine gearbox bearing, and extract the true root mean square value representing energy characteristics and the true kurtosis representing impact characteristics from the vibration signal; S2. Construct a degradation simulation model based on bearing structural parameters and operating conditions, and obtain simulated degradation signals; and extract simulated root mean square value and simulated kurtosis based on the simulated degradation signals; S3. Construct a physical degradation model, and based on the simulated root mean square value, the simulated kurtosis, and the physical degradation model, obtain the degradation mapping function; S4. Input the true root mean square value and the true kurtosis into the degradation mapping function to obtain the true degradation value. Perform monotonic processing on the true degradation value to obtain the health index. S5. Based on the health index, the true root mean square value, and the true kurtosis, a multidimensional degradation feature sequence is obtained, and the remaining life of the wind turbine gearbox bearing is obtained based on the multidimensional degradation feature sequence.
2. The method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion according to claim 1, characterized in that, The root mean square value is calculated as follows: ； In the formula, N Indicates the total number of vibration data sequences; The kurtosis is calculated as follows: ； In the formula, Mean .
3. The method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion according to claim 1, characterized in that, The physical degradation model includes: ； In the formula, d ( k ) represents a sequence of analog degraded signals.
4. The method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion according to claim 1, characterized in that, The health index includes: ； In the formula, This represents the actual degradation value.
5. The method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion according to claim 1, characterized in that, S5 includes the following steps: S51. Based on the health index, the true root mean square value, and the true kurtosis, a multidimensional degradation feature sequence is obtained: ; S52, using a length of L The time window constructs a multi-step time series from the multidimensional degenerative feature sequence: ; S53. Input the multi-step time series data into a Long Short-Term Memory (LSTM) network to obtain the remaining bearing life: ； In the formula, k This represents the sample number at the current moment.
6. The method for predicting the remaining life of wind turbine gearbox bearings based on physical degradation mapping and multi-feature fusion according to claim 5, characterized in that, The losses of the Long Short-Term Memory network include: ； In the formula, M This represents the total number of training samples; Indicates the predicted value; y i Represents the actual value.