Wind turbine gearbox residual life prediction method based on failure path consistency constraint

Abstract

The application discloses a wind turbine gearbox residual life prediction method based on failure path consistency constraint, comprising: dividing the wind turbine gearbox into structural risk units and constructing a structural risk atlas; collecting multi-source online data and preprocessing; calculating node risk value, edge coupling strength, cumulative structural damage and theoretical life prior; extracting initial degradation features, using the structural risk atlas to generate feature gating vectors to obtain gated degradation features; using a health index modeling network to compress the degradation features into health indicators, training the loss with consistency constraints; constructing a cross-condition migration loss; training the prediction model composed of a time series encoder and a regression head in two stages; outputting residual life prediction value, health state grade and the highest risk structural unit. The application runs through the whole process of feature construction, health index constraint, migration training and life regression with the structural failure path, improving the prediction accuracy and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment condition monitoring, reliability assessment and intelligent operation and maintenance technology, specifically involving a method for predicting the remaining service life of wind turbine gearboxes. Background Technology

[0002] The wind turbine gearbox is a key speed-increasing component in the wind turbine's drivetrain, typically comprising a planetary gear system, parallel gear stages, high-speed shafting, bearing support structures, and a lubrication and cooling system. Because wind turbines operate in complex environments with varying wind speeds, loads, impact loads, temperature fluctuations, and changes in lubrication conditions, gearbox degradation is usually not the result of a single component failing independently. Rather, it is the combined effect of gear meshing wear, bearing fatigue, abnormal load transmission, lubrication deterioration, and thermal coupling effects. Therefore, predicting the remaining service life of a wind turbine gearbox requires not only utilizing online monitoring data such as vibration, oil temperature, oil particle size, speed, power, and torque, but also considering the impact of the transmission path, support relationships, and failure propagation on service life degradation.

[0003] In existing technologies, Salameh et al., in their paper "Gearbox condition monitoring in windturbines: A review," pointed out that multi-source monitoring methods, including vibration analysis, oil analysis, acoustic emission analysis, temperature monitoring, and SCADA data analysis, play an important role in the condition assessment of wind turbine gearboxes. This study demonstrates that multi-source monitoring data can provide a basis for fault identification; however, it primarily focuses on condition monitoring and fault diagnosis, without further transforming transmission paths, bearing support relationships, and lubrication coupling relationships into structural constraints that can participate in RUL model training.

[0004] In their paper "Machinery health prognostics: A systematic review from data acquisition to RUL prediction," Lei et al. summarized the process of mechanical equipment health prediction, from data acquisition and health indicator construction to remaining service life prediction, pointing out that health indicator construction and RUL prediction are key links in predictive maintenance. This study shows that RUL prediction requires a continuous modeling process among degradation data, health indicators, and service life output. However, existing data-driven methods often focus on feature extraction and prediction accuracy, neglecting the correspondence between health indicators and service life output and specific structural failure paths.

[0005] In their paper "A review on physics-informed data-driven remaining useful life prediction: Challenges and opportunities," Li et al. pointed out that introducing physical mechanisms, domain knowledge, or degradation laws can help improve the interpretability and generalization ability of RUL predictions. However, for complex transmission components such as wind turbine gearboxes with clear transmission paths, support relationships, and lubrication coupling relationships, existing physics-informed data-driven methods mostly introduce constraints from the perspective of general degradation laws or damage equations, and rarely use structural failure paths as a unified prior to integrate into the degradation feature construction, health index constraints, cross-condition transfer training, and life regression processes.

[0006] Regarding patented technologies, US9360393B2, "Determining damage and remaining useful life of rotating machinery including drive trains, gearboxes, and generators," proposes estimating the remaining useful life of rotating machinery components based on cumulative damage and damage limits. WO2019043600A1, "Remaining useful life estimator," proposes estimating the remaining useful life of rotating equipment based on high-frequency vibration signals, through preprocessing, feature extraction, fault classification, and nonlinear regression. These patents embody the idea of ​​life estimation based on damage accumulation and vibration feature regression, but they do not establish a structural risk map for the internal transmission path, support relationships, and lubrication coupling relationships of wind turbine gearboxes, nor do they use structural risk relationships for degradation feature gating, health index constraints, cross-condition migration training, and high-risk structural unit location. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method for predicting the remaining life of wind turbine gearboxes based on failure path consistency constraints, so as to solve the problems in the prior art such as insufficient combination of multi-source monitoring data and structural failure mechanism, unclear physical meaning of health indicators, insufficient stability of cross-condition life prediction, and difficulty in locating specific risk locations in the prediction results.

[0008] The wind turbine gearbox remaining life prediction method based on failure path consistency constraints of this invention includes:

[0009] S1. Based on the transmission path, bearing support relationship, lubrication and cooling relationship and fault propagation relationship of the wind turbine gearbox, the wind turbine gearbox is divided into a finite number of structural risk units. With the structural risk units as nodes and the torque transmission, bearing support, lubrication and thermal coupling and fault propagation relationship as edges, a structural risk map of the wind turbine gearbox is constructed.

[0010] S2. Collect multi-source online data during the operation of the wind turbine gearbox, including real-time operating data; preprocess the multi-source online data and perform sliding window slicing to form the t-th window-level multi-source sample. ;

[0011] S3. Calculate the risk value of each node in the structural risk map under the t-th sampling window based on historical failure data and real-time operating condition data collected in step S2. and the edge coupling strength between each node And then from Combining the results yields the node risk vector. ,Depend on Combining them yields the edge coupling matrix Based on the node risk vector Calculate the cumulative structural damage and theoretical life prior ;in , , , They constitute the same failure path prior, and correspond one-to-one with the window-level samples in step S2 in time;

[0012] S4, from window-level multi-source samples Extracting the initial degenerate feature vector Construct a feature-structure risk unit mapping matrix B, and use the node risk vector obtained in step S3. and edge coupling matrix The feature gating vector is generated using the following formula:

[0013]

[0014] in, Let ξ be the feature gating vector; ξ be the gating gain coefficient; and ρ be the coupling risk impact coefficient. The feature gating vector is multiplied element-wise with the initial degraded feature vector to obtain the result. and Degradation characteristics after co-modulated structural risk gating :

[0015]

[0016] In the formula, ⊙ represents element-wise multiplication;

[0017] S5. Use a pre-trained health indicator modeling network to process the degradation features obtained in step S4. Compressed into one-dimensional health indicators The training loss function of the health indicator modeling network includes health indicators. The cumulative structural damage amount calculated in step S3 Consistency constraints between :

[0018]

[0019] Where κ is the damage mapping coefficient;

[0020] S6. Use full-life test data or historical fault samples with remaining life tags as source domain data, and online data collected from the target wind field as target domain data; based on the health indicators obtained in step S5... The health status of samples in the source and target domains is stratified, and structural risk map embedding vectors are constructed within the same health status. Cross-condition migration loss is constructed using the structural risk map embedding vector as a constraint. ;in The loss is the alignment loss for the distribution of degradation features within the same health state. Consistency constraints for failure paths within the same health state. To constrain weights;

[0021] S7. Train the remaining lifetime prediction model consisting of a time-series encoder and a regression head, using structural risk-gated degradation features across P consecutive windows during training. Health indicators HI, nodal risk vector R, cumulative structural damage D, and theoretical life prior. The output of the timing encoder is used as the input to the regression head, and the regression head outputs the remaining lifetime prediction value. ;

[0022] The training consists of two phases: the first phase uses the real remaining lifetime labels of the source domain for supervised training; the second phase, while maintaining the supervision constraints of the source domain, introduces health index constraints, theoretical lifetime flexibility constraints, and cross-condition transfer constraints constructed in step S6 for the target domain. The theoretical lifetime flexibility constraints refer to the difference between predicted lifetime and theoretical lifetime prior. The flexible consistency constraints between them do not force or replace data-driven prediction results;

[0023] S8. Output the predicted remaining useful life and health status level of the current wind turbine gearbox, wherein the predicted remaining useful life is output by the trained remaining useful life prediction model, and the health status level is based on the health indicators obtained in step S5. Determine, and based on node risk values , edge coupling matrix Coupling risks and health indicators after transmission And the degree of anomaly of the features associated with the m-th structural risk unit, and calculate the risk index of each structural risk unit. Identify the highest risk structural unit And output it.

[0024] Furthermore, the risk value mentioned in step S3 The calculation formula is:

[0025]

[0026] In the formula, Let m be the set of components belonging to the m-th structural risk unit. Based on the failure rate, This is a comprehensive working condition correction factor. The percentage of historical failures is represented by θ, where θ is the fusion coefficient. These are the normalization coefficients;

[0027] The edge coupling strength The calculation formula is:

[0028]

[0029] In the formula, The initial coupling coefficients are... The normalized deviation from the operating condition is η, where η is the influence coefficient of the deviation. If there is no direct structural coupling relationship between the two structural risk units, then... .

[0030] Furthermore, the cumulative structural damage amount mentioned in step S3 The calculation formula is:

[0031]

[0032] The theoretical life prior The calculation formula is:

[0033]

[0034] In the formula, The failure damage threshold, This represents the cumulative structural damage during the (t-1)th sampling window. The time step of the sampling window, Let ε be the average risk intensity over the most recent few windows, and ε be a small constant to prevent division by zero.

[0035] Furthermore, the multi-source online data mentioned in step S2 includes vibration acceleration signals, oil temperature, oil particle size, rotational speed, power, torque, and ambient temperature; in the feature-structure risk unit mapping matrix B mentioned in step S4, Let be the element in the k-th row and m-th column of matrix B, representing the mapping relationship between the k-th feature and the m-th structural risk unit; when the k-th feature is only related to the m-th structural risk unit, let When the k-th feature is associated with at least two structural risk units, the elements in the k-th row of matrix B are normalized so that the sum of the elements in that row is 1.

[0036] Furthermore, the health indicator modeling network in step S5 is an autoencoder or a GRU encoder, whose output is normalized to the 0~1 range by the Sigmoid function; the temporal encoder in step S7 is a gated recurrent unit, the regression head is a fully connected layer, and the output activation function is Softplus.

[0037] Furthermore, the supervised training loss in the first stage of step S7 is:

[0038]

[0039] in, The remaining lifetime regression error of the source domain. This serves as a flexible consistency constraint between the predicted lifetime of the source domain and the prior theoretical lifetime. Sorting constraints for lifetime sequences;

[0040] The total training loss in the second phase is:

[0041]

[0042] in, Constraints based on health indicators for the target domain. This serves as a flexible consistency constraint between the predicted lifetime and the theoretical lifetime prior in the target domain. For the lifetime sequence sorting constraint of the target domain, β, δ and This is the loss weight.

[0043] Furthermore, the health indicators obtained in step S5 as described in step S8 Determining a health status level specifically includes: when When the value falls below the health baseline threshold and continues to exceed a preset window number, it is determined to enter the degradation stage, and is divided into mild degradation, moderate degradation, and severe degradation according to the health index quantiles; wherein the health baseline threshold is calculated based on samples from the stable healthy operation stage, and the health index quantiles are preset percentiles within the range of health index values.

[0044] Furthermore, the risk index mentioned in step S8 The calculation formula is:

[0045]

[0046] in , , , These are the weighting coefficients. The degree of anomaly of the features associated with the m-th structural risk unit, It is the m-th component of the product of the edge coupling matrix and the node risk vector.

[0047] Furthermore, the health indicator modeling network described in step S5 adopts a self-supervised pre-training method: first, pre-training is performed with the goal of minimizing the reconstruction error, and then pre-training is performed with the inclusion of the consistency constraint term. The training loss function is fine-tuned.

[0048] Furthermore, the training loss function of the health indicator modeling network described in step S5 is:

[0049]

[0050] in, This represents the reconstruction error; For monotonic constraints; For smoothing constraints; , , These are the corresponding weighting coefficients.

[0051] The beneficial effects of this invention are:

[0052] 1. Existing methods typically extract statistical features directly from sensor data, neglecting the complex transmission paths, bearing support relationships, and fault propagation mechanisms within the wind turbine gearbox. This results in extracted features containing a large amount of noise unrelated to degradation. This invention, in step S1, divides the wind turbine gearbox into a finite number of structural risk units and constructs a structural risk map using torque transmission, bearing support, lubrication-thermal coupling, and fault propagation relationships as edges, enabling subsequent calculations to explicitly utilize physical structural knowledge. Based on this, step S4 uses the structural risk map to generate feature gating vectors. This method embeds physical risk information into the feature selection process, enhancing degradation features associated with high-risk structural units while suppressing features unrelated to low-risk units. This mechanism overcomes the shortcomings of traditional feature selection methods that rely on human experience or statistical correlation and fail to perceive the physical degradation mechanism, fundamentally improving the signal-to-noise ratio and causal correlation of degradation features.

[0053] 2. Existing cross-condition remaining lifetime prediction methods typically assume that the degradation characteristics of the source and target domains follow the same distribution. However, in reality, the degradation paths of wind turbine gearboxes differ significantly under different operating conditions, leading to severe inter-domain distribution drift and a substantial decrease in the accuracy of direct migration prediction. This invention constructs failure path consistency constraints in step S6. And combine the degradation feature distribution alignment loss Cross-condition migration loss This constraint forces the degradation features of the source and target domains within the same health status level to align in the structural risk map embedding space, enabling the model to learn shared failure path degradation patterns across operating conditions. Compared to existing methods that only align feature distributions, this invention additionally constrains the consistency of failure paths, effectively solving the problem of migration performance degradation caused by differences in degradation trajectory morphology under different operating conditions.

[0054] 3. Purely data-driven health indicator modeling methods lack physical supervision, and the output health indicators often deviate from the actual mechanical degradation state, and may even exhibit physically unreasonable fluctuations. This invention constructs a consistency constraint term between the health indicators and the cumulative structural damage in step S5. health indicators The downward trend and cumulative structural damage The growth trend aligns on a physical scale. This constraint uses the degradation estimates from the physical model as a weakly supervised signal to guide the learning of health indicators, preserving the data-driven approach's ability to fit complex nonlinear degradation while giving the health indicators a clear physical meaning. Simultaneously, the theoretical lifespan prior in step S3... The introduction of this feature, through flexible consistency constraints L in the two-stage training of step S7, further enhances this effect. life Furthermore, by integrating physical knowledge with data-driven prediction, a two-way collaborative mechanism of "physical model guidance and data-driven correction" is formed, which significantly improves the interpretability and reliability of prediction results.

[0055] 4. In actual wind power operation and maintenance scenarios, full-lifetime failure data for target wind farms is extremely scarce or even completely missing, making traditional supervised learning methods unsuitable for direct application. This invention designs a two-stage training strategy in step S7: the first stage utilizes the full-lifetime labels of the source domain for sufficient supervised pre-training, enabling the model to possess basic remaining lifetime prediction capabilities; the second stage, while maintaining source domain supervision constraints, introduces health indicator constraints from the target domain. Theoretical lifespan flexible constraints and cross-condition migration constraints This allows the model to gradually adapt to the degradation characteristics of the target domain. The core innovation of this strategy lies in: theoretical lifetime flexible constraints. It serves only as a regularization term to guide the prediction direction, without forcibly replacing data-driven results, thus avoiding prediction bias caused by excessive constraints in the case of an unlabeled target domain. Compared to existing semi-supervised or unsupervised domain adaptation methods, this invention achieves prediction accuracy close to that of supervised learning under unlabeled target domain conditions through multi-constraint joint optimization.

[0056] 5. Existing remaining useful life prediction methods typically only output a single remaining useful life value, failing to provide diagnostic information on degradation location and risk sources, which is detrimental to operation and maintenance decisions. This invention constructs a risk index in step S8. This index comprehensively considers four dimensions: node's own risk, degree of feature anomaly, global health level, and coupling propagation risk, enabling it to accurately locate the highest-risk structural unit at the current moment. Simultaneously, in step S8, based on health indicators... By classifying health status levels (mild degradation, moderate degradation, and severe degradation) using health baseline thresholds and quantiles, users can intuitively determine the degradation stage. This three-dimensional output model of "remaining lifetime value + health status level + highest risk unit" provides wind farm operation and maintenance personnel with a more comprehensive and instructive decision-making basis compared to a single lifetime prediction value.

[0057] 6. This invention achieves a close synergistic effect between its various steps through a series of interconnected technical means:

[0058] At the physical modeling level, by quantifying the transmission path, bearing support relationship, lubrication and cooling relationship, and fault propagation relationship of the wind turbine gearbox into specific calculation formulas for node risk value and edge coupling strength, structural topology information and historical failure data are transformed into quantitative risk metrics. This provides a physically interpretable computational foundation for subsequent feature selection and transfer learning, avoiding the uncertainties brought about by purely empirical parameter settings. Simultaneously, by clarifying the mapping relationship and normalization rules between multi-source signals such as vibration, temperature, and granularity and structural risk units, different types of sensor data can be accurately assigned to corresponding risk units according to their physical correlation, ensuring the feasibility of the feature gating mechanism in practical engineering.

[0059] At the model construction level, an autoencoder or GRU encoder is used as the health indicator modeling network, a gated recurrent unit as the temporal encoder, and a fully connected layer as the regression head. This approach effectively controls the number of parameters while ensuring the model's ability to fit complex degradation patterns, enabling efficient deployment on edge computing devices in industrial settings. The health indicator modeling network employs a self-supervised training method, first pre-training with reconstruction errors and then fine-tuning with consistency constraints. Optimization is achieved through a training loss function that includes reconstruction errors, monotonic constraints, consistency constraints, and smoothness constraints. This ensures that the output health indicators possess physical consistency, monotonicity, and smoothness, effectively improving the reliability of health indicators as degradation representations.

[0060] At the training strategy level, by clearly defining the specific composition and weight relationships of each loss function in the two-stage training process, the multi-constraint joint optimization has a clear mathematical expression, ensuring stable model training and rapid convergence. In particular, the design of using the theoretical lifetime flexible constraint only as a regularization term to guide the prediction direction without forcibly replacing the data-driven results effectively avoids prediction bias caused by excessive constraints in the case of an unlabeled target domain.

[0061] At the output application level, continuous health indicators are discretized into clear degradation levels through health baseline thresholds and preset quantiles, enabling maintenance personnel to intuitively determine the current degradation stage of the gearbox and enhancing the operability of degradation diagnosis. The risk index integrates four dimensions: node-specific risk, degree of feature anomaly, global health level, and coupled propagation risk. The weights of each dimension are objectively determined using the analytic hierarchy process (AHP) or entropy weighting method, avoiding biases caused by subjective weighting and ensuring the scientific validity and credibility of the risk assessment results.

[0062] The aforementioned technical approaches support each other and progress in a layered manner: the physical model provides risk guidance for feature gating, feature gating provides high-quality input for health indicators, health indicators provide a hierarchical basis for transfer constraints, and transfer constraints provide cross-domain adaptation capabilities for two-stage training, ultimately achieving accurate remaining lifetime prediction and risk localization. These technical features form an organic whole; the absence of any one link will affect the stability and reliability of the final prediction performance. Attached Figure Description

[0063] Figure 1 This is a flowchart of a method for predicting the remaining life of wind turbine gearboxes based on failure path consistency constraints. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0065] This embodiment takes a typical wind turbine gearbox as an example. The gearbox adopts a structure of two-stage planetary gears plus a single-stage parallel shaft.

[0066] The wind turbine gearbox remaining life prediction method based on failure path consistency constraints in this embodiment includes:

[0067] S1. Constructing a structural risk map of wind turbine gearboxes

[0068] Based on the transmission path, bearing support relationship, lubrication and cooling relationship, and fault propagation relationship of the wind turbine gearbox, it is divided into M structural risk units. In this embodiment, M=8, which are: low-speed planetary gear train unit, medium-speed planetary gear train unit, high-speed parallel shaft gear pair unit, low-speed bearing assembly unit, medium-speed bearing assembly unit, high-speed bearing assembly unit, lubrication and cooling system unit, and gearbox and connecting parts unit.

[0069] Using these eight structural risk units as nodes, and their torque transmission, bearing support, lubrication-thermal coupling, and fault propagation relationships as edges, a structural risk map is constructed. Specifically, torque transmission edges exist between the low-speed planetary gear train unit and the medium-speed planetary gear train unit, and between the medium-speed planetary gear train unit and the high-speed parallel shaft gear pair unit; bearing support edges exist between the low-speed bearing assembly unit and the low-speed planetary gear train unit, the medium-speed bearing assembly unit and the medium-speed planetary gear train unit, and the high-speed bearing assembly unit and the high-speed parallel shaft gear pair unit; lubrication-thermal coupling edges exist between the lubrication and cooling system unit and the low-speed, medium-speed, and high-speed bearing assembly units; and fault propagation edges are set based on historical fault statistics. For example, if tooth surface wear faults in the low-speed planetary gear train unit can propagate to the low-speed bearing assembly unit through the lubrication oil circuit, a fault propagation edge is established between them.

[0070] S2. Collect multi-source online data and perform preprocessing.

[0071] Accelerometers were installed at the bearing housings of the input shaft, intermediate shaft, and output shaft of the wind turbine gearbox to collect vibration acceleration signals at a sampling frequency of 25.6 kHz. Temperature and oil particle size sensors were installed at the inlet and outlet of the gearbox's lubrication oil circuit to collect oil temperature and oil particle size signals. Speed, power, torque, and ambient temperature data were obtained from the wind farm's SCADA system. All sensor data were collected synchronously with a sampling period of 1 second.

[0072] The acquired raw data is preprocessed as follows: First, the vibration acceleration signal is resampled in the angular domain, converting the equal time interval sampling into equal angular interval sampling. The angular domain resampling formula is as follows:

[0073]

[0074] In the formula Rotational speed, in r / min; The angle is then determined. Subsequently, wavelet denoising is performed on the vibration signal using the db4 wavelet basis, with a decomposition layer of 5, and soft threshold denoising. Outlier removal, time synchronization, and normalization are performed on the oil and operating condition signals.

[0075] The preprocessed data is sliced ​​using a sliding window, with a window length P of 30 minutes and an overlap rate of 50% between adjacent windows, forming the t-th window-level multi-source sample. :

[0076]

[0077] In the formula, This is a angular domain vibration signal; These are state parameters, including oil temperature, oil particle size, and bearing temperature, etc. These are operating parameters, including speed, power, torque, and ambient temperature. Each window-level sample corresponds temporally to the failure path prior output by S1, providing a unified data entry point for subsequent feature gating and lifespan modeling.

[0078] S3. Calculate the node risk value, edge coupling strength, cumulative structural damage, and theoretical life prior.

[0079] Based on historical failure statistics of this type of wind turbine gearbox, the basic failure rate of each key component in each structural risk unit was calculated. The calculation formula is:

[0080]

[0081] In the formula, The historical failure count of the j-th critical component in the m-th structural risk unit; α represents the cumulative equivalent operating time (in years) of the j-th critical component; α and β are smoothing parameters, and in this embodiment, α = 0.5 and β = 1.0.

[0082] In the t-th sampling window, calculate the comprehensive operating condition correction coefficient based on the current operating condition data. The calculation formula is:

[0083]

[0084] in , , These represent the current window's speed, power, and torque, respectively. , , For the corresponding rated value, , , The operating condition sensitivity coefficient is set to 0.3, 0.4, and 0.3 in this embodiment. For example, when the speed is 14.5 rpm, the power is 2.8 MW, and the torque is 1.84 MN·m, and the rated speed is 15 rpm, the rated power is 3 MW, and the rated torque is 1.91 MN·m, the calculated value of Km,t is 1.35.

[0085] Node risk value The calculation formula is:

[0086]

[0087] In the formula, Let m be the set of components belonging to the m-th structural risk unit. Based on the failure rate, This is a comprehensive working condition correction factor. The percentage of historical failures is represented by θ, which is the fusion coefficient and is set to 0.6 in this embodiment. The normalization coefficient is the maximum value of the numerator at each node.

[0088] The edge coupling strength The calculation formula is:

[0089]

[0090] In the formula, The initial coupling coefficients are... The normalized deviation from the operating condition is η, where η is the influence coefficient of the deviation. If there is no direct structural coupling relationship between the two structural risk units, then... .

[0091] Edge coupling strength The calculation formula is:

[0092]

[0093] In the formula, The initial coupling coefficient is set according to the structural coupling strength. For example, the bearing support relationship between the low-speed planetary gear train unit and the low-speed bearing assembly unit is set to 0.85. The normalized deviation from the operating condition is calculated based on the deviation between the current power and the rated power; η is the operating condition deviation influence coefficient, which is taken as 0.5 in this embodiment; if there is no direct structural coupling relationship between the two structural risk units, then... .

[0094] The nodal risk values ​​of all structural risk units Combined into a node risk vector Strengthen the edge coupling between all structural risk units Combined into a side coupling matrix , where M is the total number of structural risk units.

[0095] Cumulative structural damage The calculation formula is:

[0096]

[0097] Theoretical lifespan prior The calculation formula is:

[0098]

[0099] In the formula, The failure damage threshold, This represents the cumulative structural damage during the (t-1)th sampling window. The time step of the sampling window, Let ε be the average risk intensity over the most recent few windows, and ε be a small constant to prevent division by zero.

[0100] The above node risk vector Side coupling matrix Cumulative structural damage and theoretical life prior All data are calculated based on the same set of historical failure data and real-time operating condition data, forming the same failure path prior, and are temporally consistent with the window-level multi-source samples obtained in step S2. One-to-one correspondence.

[0101] S4, Feature Gating

[0102] From window-level multi-source samples Extracting the initial degenerate feature vector K represents the total number of initial degradation features. In this embodiment, K is 17, including 6 features in the time domain (mean, variance, kurtosis, peak factor, impulse factor, and margin factor), 3 features in the frequency domain (centroid frequency, frequency dispersion, and sideband energy percentage), and 8 features in the time-frequency domain (energy percentage of each frequency band), for a total of 17 initial degradation features.

[0103] Construct a feature-structure risk unit mapping matrix B with dimensions K×M, which is 17×8 in this embodiment. Let be the element in the k-th row and m-th column of matrix B, representing the mapping relationship between the k-th feature and the m-th structural risk unit; when the k-th feature is only related to the m-th structural risk unit, let When the k-th feature is associated with at least two structural risk units, the elements in the k-th row of matrix B are normalized so that the sum of the elements in that row is 1.

[0104] Taking the kurtosis characteristic of vibration acceleration signal as an example, this characteristic mainly reflects impact-type faults and is related to the low-speed planetary gear train unit, the medium-speed planetary gear train unit, and the high-speed bearing assembly unit. Therefore, the elements in the columns of these three units in the corresponding row of matrix B are all assigned the value 1 / 3, and the remaining columns are 0.

[0105] Generating feature-gated vectors using structural risk maps:

[0106]

[0107] in, ξ is the feature gating vector; ξ is the gating gain coefficient, which is 0.5 in this embodiment; ρ is the coupling risk impact coefficient, which is 0.3 in this embodiment; the feature gating vector and the initial degraded feature vector are multiplied element-wise to obtain the degraded features after structural risk gating:

[0108]

[0109] In the formula, ⊙ represents element-wise multiplication.

[0110] S5, Health Indicator Modeling

[0111] A pre-trained health indicator modeling network is used to gating the degradation features after structural risk. Compressed into one-dimensional health indicators In this embodiment, the health indicator modeling network adopts an autoencoder structure, comprising an encoder and a decoder. The encoder compresses the K-dimensional input into 3-dimensional latent variables, and the decoder reconstructs the 3-dimensional latent variables back into a K-dimensional output. Both the encoder and decoder contain two fully connected layers with ReLU activation function. The output layer uses the sigmoid function to normalize the health indicators to the range of 0 to 1, where 1 represents a completely healthy state and 0 represents a completely ineffective state.

[0112] The health indicator modeling network employs a self-supervised pre-training method: first, it is pre-trained using the minimization of reconstruction error as the training loss function, enabling the autoencoder to learn to extract the main change patterns of degenerative features; then, it is pre-trained using a method incorporating reconstruction error, monotonic constraints, and consistency constraints. The training loss function with smoothing constraints is fine-tuned, and the calculation formula of the training loss function is as follows:

[0113]

[0114] Among them, reconstruction error Mean squared error of input and output for modeling health indicators; monotonic constraints The penalty term for the upward trend of health indicators over time is calculated by summing the positive values ​​of the differences between health indicators at adjacent time points; smoothing constraint. The penalty term for fluctuations in health indicators at adjacent time points is the sum of squares of the differences in health indicators at adjacent time points; , , The corresponding weighting coefficients are 0.2, 0.5, and 0.1 in this embodiment.

[0115] Consistency constraints The calculation formula is as follows:

[0116]

[0117] Wherein, κ is the damage mapping coefficient, which is taken as 1.0 in this embodiment.

[0118] S6, Construction of Cross-Condition Migration Constraints

[0119] Full-life test data or historical failure samples with remaining lifetime labels are used as source domain data. Full-life test data refers to complete operational data of the wind turbine gearbox from healthy operation to complete failure, with each time window having a corresponding remaining lifetime label. Historical failure samples refer to multi-source online data segments recorded during the historical operation of units that have experienced failures in actual wind farms; their remaining lifetime labels are constructed based on the failure time, and the labeling method is as follows:

[0120]

[0121] In the formula, The time of failure is t, and the time of acquisition in the current window is t. To maximize the cutoff lifetime and avoid excessively large lifetime labels for early healthy samples that could lead to training instability, this embodiment uses 730 days (approximately 2 years). The target domain data may not have real RUL labels, but pseudo-labels for the healthy status can be obtained from S4.

[0122] The online data collected from the target wind field is used as the target domain data. The target domain data lacks a true remaining lifetime label, but health indicators can be obtained from step S5. The health status level is used as a pseudo-label. The health indicators obtained from step S5... Stratify the health status of samples in the source and target domains: divide the range of health indicators into normal ( ) based on quantiles. >0.8), mild degradation (0.6< ≤0.8), moderate degradation (0.4< ≤0.6) and severe degradation ( Four levels: ≤0.4.

[0123] Within the same health state, construct the structural risk map embedding vector:

[0124]

[0125] The vector By integrating node-specific risk, coupled propagation risk, and cumulative damage information, and using this embedding vector as a constraint, a cross-condition migration loss is constructed:

[0126]

[0127] in, To align the degradation feature distributions within the same health state, the maximum mean difference is used to measure the distance between the degradation feature distributions in the source and target domains. For failure path consistency constraints within the same health state, the calculation formula is:

[0128]

[0129] Where S is the set of samples within a certain health state in the target domain. The structural risk map embedding vector in the source domain that is in the same health state as the sample. The set of indices of the k nearest neighbors, where k is 5. This constraint forces the target domain sample to remain close to its nearest neighbor in the source domain in the structural risk map embedding space, thereby achieving consistent migration of failure paths. To constrain the weights, in this embodiment Take 0.5.

[0130] S7, Two-Phase Training

[0131] The remaining lifespan prediction model consists of a time-series encoder and a regression head. The time-series encoder employs a gated recurrent unit (GRU), specifically two layers in this embodiment, with 64 hidden units per layer. The input dimension for each time step is K+1+M+1+1 dimensions (K-dimensional degradation features, 1-dimensional health indicators, M-dimensional node risk vectors, 1-dimensional cumulative damage, and 1-dimensional theoretical lifespan prior), with a time step size of 30. The regression head uses a fully connected layer, containing two hidden layers with dimensions of 64 and 32 respectively, and the output layer activation function is Softplus.

[0132] During training, structural risk gating degradation features were used across P consecutive windows. Health indicators HI, nodal risk vector R, cumulative structural damage D, and theoretical life prior. As input to the timing encoder ,Right now:

[0133]

[0134] The output of the timing encoder is used as the input of the regression head, and the regression head outputs the remaining lifetime prediction value. :

[0135]

[0136] In the formula, Let be the predicted remaining lifetime for the t-th time window; the Softplus function is used to ensure that the predicted lifetime is non-negative. This is the weight matrix of the regression head. The two terms are the bias terms of the regression head and are both trainable parameters.

[0137] The training is divided into two phases:

[0138] Phase 1: Supervised training using real remaining lifetime labels from the source domain. The training loss is:

[0139]

[0140] in, The source domain remaining lifetime regression error is calculated using the mean squared error formula, as follows:

[0141]

[0142] in, Let be the predicted remaining lifetime of the i-th sample within the batch. Let N be the true remaining lifetime of the i-th sample in the batch, where i is the sample index in the batch; N is the batch size, which is 64 in this embodiment.

[0143] To provide a flexible consistency constraint between the predicted lifetime and the theoretical lifetime prior in the source domain, the Huber loss function is adopted, and the calculation formula is as follows:

[0144]

[0145] When the deviation is less than c, the squared error is used; otherwise, the linear error is used. In this embodiment, c is set to 1.0. This represents the theoretical lifetime prior for the i-th sample within the batch.

[0146] The lifetime sequence ordering constraint is calculated using the following formula:

[0147]

[0148] Penalize sample pairs whose predicted lifetime ranking is inconsistent with the actual lifetime ranking, where i and j are sample indices within the batch.

[0149] Take 0.3, Set the value to 0.2. Training uses the Adam optimizer with an initial learning rate of 10. −3The batch size is 64, and training stops when the loss on the validation set no longer decreases for 10 consecutive rounds.

[0150] Phase Two: While maintaining the source domain supervision constraints, introduce health indicator constraints, theoretical lifespan flexibility constraints, and cross-condition transfer constraints constructed in step S6 for the target domain. The total training loss function is:

[0151]

[0152] in, Constraints based on health indicators for the target domain; Predicted lifetime and theoretical lifetime prior for the target domain The flexible consistency constraints between them do not force or replace data-driven prediction results; Sorting constraints for the lifetime sequence of the target domain; β, , δ and This is the loss weight. and The calculation method is the same as in the first stage. and The same applies, except that the source domain data is replaced with the target domain data. β , δ and As a loss weight, β is set to 0.5 in this embodiment. Take 0.3, δ take 0.4, Set the learning rate to 0.2. The second phase involves 50 epochs of training, with the learning rate reduced to 10%. −4 .

[0153] S8, Combined Output

[0154] After training, the structural risk gating degradation features of the current window will be determined. Health indicators Node risk vector R t Cumulative structural damage D t and theoretical life prior Input the trained remaining lifetime prediction model to obtain the remaining lifetime prediction value. .

[0155] At the same time, according to health indicators Determining the health status level: A baseline value for the health indicator was calculated using samples from the first three months of the gearbox's stable and healthy operation phase. In this embodiment, the baseline value was 0.95, and the health baseline threshold was set to 0.75 times the baseline value, i.e., 0.71. When the health indicator is below 0.71 and remains below it for more than three consecutive windows, the gearbox is considered to have entered the degradation phase. Based on the health indicator quantiles of 33% and 67%, the degradation phase was divided into mild degradation (health indicator greater than 0.55 and less than or equal to 0.71), moderate degradation (health indicator greater than 0.40 and less than or equal to 0.55), and severe degradation (health indicator less than or equal to 0.40).

[0156] And based on the node risk value , edge coupling matrix Coupling risks and health indicators after transmission And the degree of anomaly of the features associated with the m-th structural risk unit, and calculate the risk index of each structural risk unit. :

[0157]

[0158] in , , , The weighting coefficients are determined using the analytic hierarchy process (AHP). In this embodiment, we take... =0.35, =0.25, =0.25, =0.15. The degree of anomaly of the feature associated with the m-th structural risk unit is obtained by calculating the Mahalanobis distance between the current feature value and the historical normal distribution; Let m be the m-th component of the product of the edge coupling matrix and the node risk vector, and its calculation formula is:

[0159]

[0160] Where M is the total number of structural risk units, j is the summation variable, and the index of all structural risk units is traversed. Side coupling matrix The element in the m-th row and j-th column represents the edge coupling strength from the j-th structural risk unit to the m-th structural risk unit, and its value range is [0,1]. Let be the node risk value of the j-th structural risk unit. This represents the sum of coupling risks passed from all other structural risk units to the m-th structural risk unit through structural coupling relationships.

[0161] Identify the highest risk structural unit The output includes: predicted remaining lifespan, health status level, and name of the highest-risk structural unit.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining life of wind turbine gearboxes based on failure path consistency constraints, characterized in that, include: S1. Based on the transmission path, bearing support relationship, lubrication and cooling relationship and fault propagation relationship of the wind turbine gearbox, the wind turbine gearbox is divided into a finite number of structural risk units. With the structural risk units as nodes and the torque transmission, bearing support, lubrication and thermal coupling and fault propagation relationship as edges, a structural risk map of the wind turbine gearbox is constructed. S2. Collect multi-source online data during the operation of the wind turbine gearbox, including real-time operating data; preprocess and slice the multi-source online data using a sliding window to form the t-th window-level multi-source sample. ; S3. Calculate the risk value of each node in the structural risk map under the t-th sampling window based on historical failure data and real-time operating condition data collected in step S2. and the edge coupling strength between each node And then from Combining the results yields the node risk vector. ,Depend on Combining them yields the edge coupling matrix Based on the node risk vector Calculate the cumulative structural damage and theoretical life prior ;in , , , They constitute the same failure path prior, and correspond one-to-one with the window-level samples in step S2 in time; S4, from window-level multi-source samples Extracting the initial degenerate feature vector Construct a feature-structure risk unit mapping matrix B, and use the node risk vector obtained in step S3. and edge coupling matrix The feature gating vector is generated using the following formula: in, Let ξ be the feature gating vector; ξ be the gating gain coefficient; and ρ be the coupling risk impact coefficient. The feature gating vector is multiplied element-wise with the initial degraded feature vector to obtain the result. and Degradation characteristics after co-modulated structural risk gating : In the formula, ⊙ represents element-wise multiplication; S5. Use a pre-trained health indicator modeling network to process the degradation features obtained in step S4. Compressed into one-dimensional health indicators The training loss function of the health indicator modeling network includes health indicators. The cumulative structural damage amount calculated in step S3 Consistency constraints between : Where κ is the damage mapping coefficient; S6. Use full-life test data or historical fault samples with remaining life tags as source domain data, and online data collected from the target wind field as target domain data; based on the health indicators obtained in step S5... The health status of samples in the source and target domains is stratified, and structural risk map embedding vectors are constructed within the same health status. A cross-condition migration loss is constructed using the structural risk map embedding vector as a constraint. ;in The loss is the alignment loss for the distribution of degradation features within the same health state. Consistency constraints for failure paths within the same health state. To constrain weights; S7. Train the remaining lifetime prediction model consisting of a time-series encoder and a regression head, using structural risk-gated degradation features across P consecutive windows during training. Health indicators HI, nodal risk vector R, cumulative structural damage D, and theoretical life prior. The output of the timing encoder is used as the input to the regression head, and the regression head outputs the remaining lifetime prediction value. ; The training consists of two phases: the first phase uses the real remaining lifetime labels of the source domain for supervised training; the second phase, while maintaining the supervision constraints of the source domain, introduces health index constraints, theoretical lifetime flexibility constraints, and cross-condition transfer constraints constructed in step S6 for the target domain. The theoretical lifetime flexibility constraints refer to the difference between predicted lifetime and theoretical lifetime prior. The flexible consistency constraints between them do not force or replace data-driven prediction results; S8. Output the predicted remaining useful life and health status level of the current wind turbine gearbox, wherein the predicted remaining useful life is output by the trained remaining useful life prediction model, and the health status level is based on the health indicators obtained in step S5. Determine, and based on node risk values , edge coupling matrix Coupling risks and health indicators after transmission And the degree of anomaly of the features associated with the m-th structural risk unit, and calculate the risk index of each structural risk unit. Identify the highest risk structural unit And output it.

2. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The risk value mentioned in step S3 The calculation formula is: In the formula, Let m be the set of components belonging to the m-th structural risk unit. Based on the failure rate, This is a comprehensive working condition correction factor. The percentage of historical failures is represented by θ, where θ is the fusion coefficient. These are the normalization coefficients; The edge coupling strength The calculation formula is: In the formula, The initial coupling coefficients are... The normalized deviation of the operating condition is η, and the deviation of the operating condition is the influence coefficient. If there is no direct structural coupling relationship between two structural risk units, then .

3. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The cumulative structural damage amount mentioned in step S3 The calculation formula is: The theoretical life prior The calculation formula is: In the formula, The failure damage threshold, This represents the cumulative structural damage during the (t-1)th sampling window. The time step of the sampling window, Let ε be the average risk intensity over the most recent few windows, and ε be a small constant to prevent division by zero.

4. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The multi-source online data mentioned in step S2 includes vibration acceleration signals, oil temperature, oil particle size, rotational speed, power, torque, and ambient temperature; in the feature-structure risk unit mapping matrix B mentioned in step S4, Let be the element in the k-th row and m-th column of matrix B, representing the mapping relationship between the k-th feature and the m-th structural risk unit; when the k-th feature is only related to the m-th structural risk unit, let When the k-th feature is associated with at least two structural risk units, the elements in the k-th row of matrix B are normalized so that the sum of the elements in that row is 1.

5. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The health indicator modeling network in step S5 is an autoencoder or a GRU encoder, whose output is normalized to the 0~1 range by the Sigmoid function; the time encoder in step S7 is a gated recurrent unit, the regression head is a fully connected layer, and the output activation function is Softplus.

6. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The supervised training loss in the first stage of step S7 is: in, The remaining lifetime regression error of the source domain. This serves as a flexible consistency constraint between the predicted lifetime of the source domain and the prior theoretical lifetime. Sorting constraints for lifetime sequences; The total training loss in the second phase is: in, Constraints based on health indicators for the target domain; A flexible consistency constraint is imposed between the predicted lifetime of the target domain and the theoretical lifetime prior. For the lifetime sequence sorting constraint of the target domain, β, , δ and This is the loss weight.

7. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The health indicators obtained in step S5 as described in step S8 Determining a health status level specifically includes: when When the value falls below the health baseline threshold and continues to exceed a preset window number, it is determined to enter the degradation stage, and is divided into mild degradation, moderate degradation, and severe degradation according to the health index quantiles; wherein the health baseline threshold is calculated based on samples from the stable healthy operation stage, and the health index quantiles are preset percentiles within the range of health index values.

8. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The risk index mentioned in step S8 The calculation formula is: in , , , These are the weighting coefficients. The degree of anomaly of the features associated with the m-th structural risk unit, It is the m-th component of the product of the edge coupling matrix and the node risk vector.

9. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The health indicator modeling network described in step S5 adopts a self-supervised pre-training method: first, pre-training is performed with the goal of minimizing the reconstruction error, and then pre-training is performed with the inclusion of the consistency constraint term. The training loss function is fine-tuned.

10. The method for predicting the remaining life of a wind turbine gearbox based on failure path consistency constraints according to claim 1, characterized in that: The training loss function of the health indicator modeling network described in step S5 is: in, This represents the reconstruction error; For monotonic constraints; For smoothing constraints; , , These are the corresponding weighting coefficients.

Images