A wind turbine gearbox fault early warning and state evaluation method

Abstract

The application discloses a wind turbine gearbox fault early warning and state evaluation method, an improved integrated KNN model is used as a modeling core to improve model estimation precision, similarity measurement is improved to consider feature weight relationship affinity distance, and the reliability of the similarity is improved; in view of the problem that the difference of the integrated KNN model using only random sampling with replacement is not significant, leading to poor model robustness, feature level sampling is added, and the difference of the self-learner is increased; in view of the problem that one-way evaluation of feature relationship leads to modeling feature redundancy, a feature priority sorting method based on forward search is proposed, and the 'dimension disaster' of the model is reduced. The application effectively solves the problems of poor robustness, modeling feature redundancy and easy to fall into local optimum of the traditional KNN integrated method in dealing with high-dimensional variable modeling, has high precision, strong generalization and strong robustness, and improves the real-time performance and reliability of the wind turbine gearbox fault early warning and state evaluation.

Description

Technical Field

[0001] This invention pertains to the condition monitoring and fault diagnosis of wind power generation equipment, specifically relating to a method for early warning and condition assessment of wind turbine gearbox faults. Background Technology

[0002] Wind turbines are typically deployed in harsh environments, such as offshore platforms or mountainous areas, where transportation is often inconvenient and natural conditions are severe. When critical components of wind turbines, such as gearboxes and generators, fail, maintenance faces significant challenges due to high technical difficulty and costs. Industry statistics show that the operation and maintenance costs of wind turbines account for approximately 25%-30% of their total lifespan, with downtime due to failures causing substantial economic losses. Therefore, achieving online fault early warning and condition assessment for wind turbines is of great significance for reducing operation and maintenance costs and improving the economic efficiency of wind farms.

[0003] Currently, wind turbine fault early warning technologies mainly include vibration analysis, oil analysis, acoustic emission technology, and temperature monitoring. Vibration analysis monitors the vibration characteristics of components such as bearings and gears, but requires numerous additional sensors, resulting in high costs, and the signals are easily affected by environmental noise. Oil analysis determines wear conditions by analyzing the content and morphology of metal particles in lubricating oil, but the sampling cycle is long, making real-time monitoring impossible. Acoustic emission technology monitors the sound wave signals generated by components under stress, is sensitive to environmental noise, and requires specialized equipment. Temperature monitoring only monitors temperature changes in key areas, has low sensitivity, and struggles to detect early faults. In contrast, condition monitoring methods based on SCADA (Supervisory Control and Data Acquisition) systems have significant advantages because wind turbines are generally equipped with SCADA systems that record operating parameters every 5-10 minutes, including wind speed, power, rotational speed, temperature, and other variables, acquiring abundant operating data without the need for additional sensors.

[0004] State monitoring methods based on SCADA data are mainly implemented by establishing Normal Behavior Modeling (NBM). Common modeling methods include parametric and non-parametric methods. Parametric methods include linear regression and autoregressive moving average models, while non-parametric methods include the K-Nearest Neighbor (KNN) algorithm, multivariate state estimation techniques, and autoassociative kernel regression algorithms. The KNN algorithm, as a classic non-parametric method, does not rely on pre-trained models and directly uses stored data to solve classification or regression problems. It is flexible in handling large amounts of data and suitable for fault early warning in complex wind turbine operating conditions. However, the traditional KNN algorithm still has four core algorithmic defects in actual industrial scenarios, making it difficult to meet the requirements for early warning accuracy, real-time performance, and reliability.

[0005] 1. Feature processing algorithms are one-sided and redundancy removal is incomplete: During equipment operation, multiple dimensions of state parameters such as temperature, speed, and power are involved. Existing feature selection algorithms generally suffer from the defect of "single-dimensional correlation preference." For example, they only calculate the correlation between input and output features using indicators such as mutual information and Pearson coefficient, but ignore the redundancy analysis between input features (such as the strong correlation of temperature sensor data at different locations, and the coupling relationship between power and speed). This one-sidedness leads to the retention of a large number of highly redundant features (correlation coefficient > 0.95), which not only causes the "curse of dimensionality" (the model parameter size increases by more than 50%), but also amplifies noise interference due to information duplication, causing early fault features to be submerged.

[0006] 2. Limited accuracy of state estimation algorithms: Traditional non-parametric models such as KNN often use Euclidean distance to measure sample similarity. This distance does not take into account feature weights and data distribution characteristics, and is easily affected by outliers and gets trapped in local optima, resulting in poor accuracy of equipment state estimation models and difficulty in identifying early minor faults in industrial equipment.

[0007] 3. Insufficient generalization ability of traditional ensemble KNN algorithm: Existing ensemble learning mostly uses a single bootstrap sampling strategy to generate base training sets. However, as a stable learner, KNN has low sensitivity to single data sampling, resulting in insufficient differentiation of base learners. This limits the improvement of the generalization ability of the ensemble model and makes it poorly adaptable to complex working conditions (such as high temperature, high vibration, and load fluctuation).

[0008] 4. Lagging fault warning and status assessment algorithms and high false alarm rate: Most methods adopt a single early warning mechanism of "single residual exceeding the threshold", without combining the dynamic change trend of fault probability. On the one hand, the early warning of early faults is lagging, and it is impossible to reserve a sufficient window period for maintenance. On the other hand, the high false alarm rate caused by instantaneous fluctuations of sensors will interfere with operation and maintenance decisions.

[0009] Existing technologies have failed to achieve collaborative optimization of the entire process of "feature optimization - state estimation - integration and fusion - early warning assessment," and in particular, lack efficient integrated algorithm solutions for early fault identification under complex operating conditions. Therefore, there is an urgent need to construct a data-driven method that balances feature processing efficiency, state estimation accuracy, generalization ability, and early warning reliability to solve the technical challenges of delayed fault early warning and inaccurate state assessment in industrial equipment. Summary of the Invention

[0010] To address the shortcomings of existing technologies, achieve efficient data redundancy removal, enhance model generalization ability, accurately capture early and subtle fault trends of equipment, dynamically assess equipment health status, and provide accurate decision support for predictive maintenance of wind turbines, this invention adopts the following technical solution:

[0011] A method for early warning and condition assessment of wind turbine gearbox faults includes the following steps:

[0012] Step S1: Obtain wind turbine gearbox operating data;

[0013] Step S2: Using the running data as input data and the key parameters selected from the running data as output data, compare the importance of different features in the running data through mutual information, and sort the features based on the forward search strategy to construct the training sample set and test sample set of the state estimation model.

[0014] Step S3: Traditional KNN algorithms typically use Euclidean distance as a similarity metric. However, Euclidean distance is sensitive to outliers and fails to fully consider the global distribution characteristics of the feature space, easily leading to the model getting trapped in local optima. To address this, this invention constructs an improved ensemble KNN state estimation model. Based on the affinity distance between the test sample and each training sample, it selects a set of nearest neighbor samples most similar to the test sample from the training sample set. Based on the output data of these nearest neighbor samples, it predicts the output data of the test sample. This fundamentally improves the traditional KNN algorithm by introducing affinity distance as a new similarity metric, significantly enhancing the model's accuracy and robustness. A set of base learners for the ensemble KNN state estimation model is constructed through a two-layer sampling mechanism and trained using the training sample set. For a new test sample, the key parameter estimates are predicted based on the output results of all base learners.

[0015] Step S4: Collect running data in real time, estimate key parameters in real time using the trained ensemble KNN state estimation model, calculate the residuals between the estimated key parameters and the actual observation data of the sensor, and obtain the real-time residual sequence;

[0016] Step S5: Based on the verification sample set of the gearbox in historical healthy operating conditions, calculate the residual sequence of the verification sample set in order to set a dynamic early warning threshold through SPC technology;

[0017] Step S6: Evaluate the gearbox status based on the number of samples in the residual sequence that exceed the warning threshold.

[0018] Furthermore, in step S2, the importance of different features is compared by calculating mutual information values. The formula for mutual information values ​​is as follows:

[0019]

[0020]

[0021]

[0022] in, Indicates the characteristics of the input runtime data. This indicates the key parameter characteristics of the output. The information entropy of the feature is represented by p(·), which represents the probability density function and is solved using the kernel density estimation method. Representing input features With output features Mutual information between them, min{·} represents the minimum value operation.

[0023] Furthermore, the forward search strategy in step S2 is as follows:

[0024] In the initial iteration of the state estimation model, the input data with the highest mutual information with the output data is selected from the input data and placed into an empty ordered dataset as the primary data.

[0025] In subsequent iterations of the state estimation model, a priority is constructed based on the correlation between the input data and the output data, as well as the average redundancy between the input data and the selected data in the ordered dataset, to select the input data and obtain a data set arranged in descending order of priority. The data with the highest priority is selected from the data set and added to the ordered dataset, finally obtaining an ordered dataset arranged in descending order of priority.

[0026] Furthermore, the priority is measured by a priority score, as shown in the following formula:

[0027]

[0028] in, Indicates priority score, This indicates the features of the selected input data. This represents the number of features in an ordered dataset. This score represents the selected features in the ordered dataset. It takes into account both the correlation between the feature and the output variable (the higher the correlation between the feature and the output variable, the higher the model accuracy) and the average redundancy with the selected feature set (the lower the redundancy between input features, the more lightweight the model is and the more it avoids the "curse of dimensionality").

[0029] Furthermore, in step S3, the construction process of the training sample set for the base learner is as follows:

[0030] Feature sampling is performed based on priority. A set of high-priority data is randomly selected from the ordered dataset to form a subset of training samples for the base learners. This ensures that each base learner focuses on the globally optimal feature combination, while also introducing differences by randomly adjusting the number of features.

[0031] Using the Bootstrap method, perform effective replacement random sampling from the original training sample set or the subset of training samples to extract a set of base learner training samples with the same number of samples as the original training set.

[0032] Furthermore, for a new test sample, the output data of all base learners are aggregated, and their average is calculated as the final predicted output data of the ensemble KNN state estimation model.

[0033] Furthermore, the affinity distance in step S3 is calculated using the following formula:

[0034]

[0035]

[0036] Where d(,) represents the affinity distance, m represents the number of test samples, and x tj Let x represent the j-th test sample. ij This represents the i-th training sample corresponding to the j-th test sample. This represents the adaptive feature weights, where n represents the number of training samples, and the numerator... Indicates test sample In features The total absolute distance between the test sample and all training samples reflects the "periphery" of the test sample relative to the training set as a whole. The denominator contains... Indicates training samples In features The total absolute distance between the weight and all other training samples reflects the "intrinsic dispersion" of the training set on this feature. The physical meaning is that when a test sample differs significantly from the training set in a certain feature (large numerator), and the training set itself is highly concentrated on that feature (small second term in the denominator), that feature may carry abnormal information or noise. The value will be adaptively lowered, thereby reducing the negative impact of this anomalous feature on the overall distance calculation. This mechanism makes the distance metric more global and enhances the model's ability to combat outliers and local perturbations.

[0037] Further, in step S5, using the verification sample set of the gearbox under historical healthy operating conditions, the sample mean and sample standard deviation of the residual sequence of the verification sample set are calculated. Based on the normal distribution characteristics, the gearbox fault early warning threshold is set, as shown in the following formula:

[0038]

[0039] in, This represents a coefficient set based on the system's tolerance for false alarms, i.e., the warning threshold. For example, when... At this point, corresponding to a 99% confidence level of the standard normal distribution, it means that under purely random fluctuations, the residual has only a 1% probability of naturally exceeding this threshold. This threshold serves as the primary standard for judging whether the equipment status is abnormal.

[0040] Further, in step S6, by setting a sliding window, the number of samples whose corresponding residual values ​​exceed the warning threshold is counted at each window position. Based on the ratio of the number of samples exceeding the threshold to the length of the sliding window, the gearbox failure rate is obtained. Based on the long-term trend curve of the failure rate changing with time or the sliding window number, the gearbox life is divided. The failure rate curve fluctuating steadily at a low level without a significant upward trend is classified as a healthy period. The failure rate curve showing a significant and continuous upward trend and starting to oscillate widely in a higher range is classified as an unstable state. The failure rate curve maintaining a high level and no longer falling is classified as a serious failure. The critical point where the failure rate curve is first confirmed to change from the healthy period to the early degradation period is used as the early warning point.

[0041] Furthermore, historical multi-dimensional cross-sectional data of the gearbox under known healthy operating conditions are acquired, including but not limited to environmental parameters (such as ambient temperature and humidity), operating status parameters (such as speed, power, pressure, and flow rate), and key component status parameters (such as bearing temperature and gearbox oil temperature), and data preprocessing is performed, including the following steps:

[0042] Step S1.1: Use the Laida criterion to remove sample points with missing data records to eliminate missing values;

[0043] Step S1.2: Perform data cleaning based on wind turbine characteristics and data distribution, including but not limited to:

[0044] Data for wind turbines in shutdown or abnormal operating mode, such as active power less than 0, wind speed greater than cut-out wind speed, and wind speed less than cut-in wind speed;

[0045] Significantly unreasonable data caused by sensor malfunction, such as data significantly exceeding the measurement range;

[0046] Statistical methods are used to identify and remove noise points that are far from the normal distribution;

[0047] Step S1.3: To eliminate the influence of differences in feature units and numerical ranges on the model, the cleaned data is normalized and scaled to the [0,1] interval. The formula used is:

[0048]

[0049] in, This is the original data. This is the minimum value of the state parameter. This is the maximum value of the state parameter.

[0050] The advantages and beneficial effects of this invention are as follows:

[0051] 1. Improved prediction accuracy and robustness

[0052] Fundamentally improved distance metric: By introducing affinity distance to replace the traditional Euclidean distance, this invention enables the similarity metric to have global adaptive capability. This distance function can automatically identify and suppress the interference of abnormal feature values, effectively overcoming the defect of traditional methods that are prone to getting trapped in local optima, thereby significantly improving the individual prediction accuracy of each base learner.

[0053] A Creative Solution to the Ensemble Challenge: Addressing the core pain point of KNN as a stable learner being insensitive to traditional Bootstrap sampling, leading to poor ensemble performance, this invention designs a dual-layer sampling mechanism. This mechanism significantly enhances the diversity among base learners by simultaneously introducing structured randomness into both the feature space and the sample space. This diversity, combined with high-precision base learners, enables the final ensembled KNN model to achieve generalization ability and prediction accuracy surpassing conventional data-driven models (such as single KNN, GPR, and LS-SVM), providing unprecedented high-fidelity state simulation for digital twins.

[0054] 2. Achieve true "early" warnings and empower predictive maintenance.

[0055] This invention goes beyond simply detecting existing faults; its core value lies in its ability to keenly capture the slow degradation trend of equipment performance. By transforming a continuous residual sequence into an instantaneous failure rate curve based on a sliding window, the system can quantify the gradual change in equipment health.

[0056] This invention can accurately identify the critical point at which equipment transitions from a "healthy period" to an "early degradation period" dozens of days before traditional equipment monitoring systems trigger alarms. This advanced warning window completely changes the passive situation of traditional "reactive maintenance" or "preventive maintenance," providing sufficient data support and decision-making time for implementing "predictive maintenance" strategies, thereby effectively avoiding unplanned downtime and reducing the risk of major accidents.

[0057] 3. Provides intuitive health status assessments, greatly improving decision-making efficiency.

[0058] This invention creatively defines the metric of "failure rate," transforming complex and obscure residual signal sequences into a simple and intuitive percentage-based health metric.

[0059] Maintenance personnel do not need a strong data science background to easily determine the current lifespan stage (healthy, early degradation, or serious failure) of equipment by observing the position and trend of the failure rate curve. This intuitive visualization greatly lowers the barrier to entry for condition monitoring technology, improves the efficiency and accuracy of decision-making, and makes maintenance actions more targeted.

[0060] 4. Possesses strong versatility and scalability, with broad application prospects.

[0061] The technical framework of this invention is entirely data-driven, relying not on physical models of specific equipment or expensive dedicated sensors, but only on routine operational data of the equipment (SCADA, etc.). This makes the method highly versatile and can be quickly deployed in various critical rotating equipment such as wind turbines, industrial motors, pumps, compressors, and gearboxes.

[0062] This invention can be used as a core engine and seamlessly embedded into various industrial internet platforms and digital twin systems to build advanced equipment health management applications, with extremely high scalability.

[0063] 5. The technical solution is complete, forming a closed loop, and has high practical value in engineering.

[0064] This invention provides a complete technical closed loop, from data preprocessing, feature engineering, model building, online monitoring to final health status assessment and early warning. Each step features targeted innovative design, and the solution is rigorous and self-contained.

[0065] This invention directly utilizes widely available monitoring data, requiring no additional hardware investment, resulting in low implementation costs, significant economic benefits, and extremely high engineering application and promotion value. Attached Figure Description

[0066] Figure 1 This is a flowchart of the method in an embodiment of the present invention.

[0067] Figure 2 This is a schematic diagram of the structure of a traditional ensemble KNN model.

[0068] Figure 3 This is a schematic diagram of the structure of the improved integrated KNN model in an embodiment of the present invention.

[0069] Figure 4 This is a schematic diagram of the sliding window state evaluation in an embodiment of the present invention.

[0070] Figure 5 This is a flowchart illustrating the lubrication process of the wind turbine gearbox in an embodiment of the present invention. Detailed Implementation

[0071] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0072] To overcome the shortcomings of existing technologies, such as one-sided feature selection, low state estimation accuracy, weak ensemble generalization ability, and insufficient early warning reliability, this invention provides a high-precision, high-generalization-ability, and reliable early warning method for wind turbine gearbox fault early warning and state assessment based on an improved ensemble algorithm. This method can efficiently remove redundant features from SCADA data and enhance the generalization ability of the ensemble learning model by increasing the difference, accurately capturing early subtle fault trends of equipment, dynamically assessing equipment health status, and providing accurate decision support for predictive maintenance of wind turbines. Specifically, it involves wind turbine gearbox state monitoring based on a Supervisory Control and Data Acquisition (SCADA) system, especially employing an improved two-layer ensemble K-nearest neighbor (EKNN) algorithm to achieve early fault warning. This fully considers the shortcomings of traditional ensemble KNN algorithms, such as insensitivity to sampling methods and the tendency of Euclidean distance to get trapped in local optima. Figure 1 As shown, the specific steps include:

[0073] Step S1: Data Preprocessing. Extract historical multi-dimensional cross-sectional data of the gearbox under known healthy operating conditions from the SCADA system's historical database. This data covers environmental parameters (such as ambient temperature and humidity), operating status parameters (such as speed, power, pressure, and flow rate), and key component status parameters (such as bearing temperature and gearbox oil temperature). Specifically, this includes the following steps:

[0074] Step S1.1: Remove missing values ​​by deleting sample points with missing data records;

[0075] Step S1.2: Perform data cleaning based on wind turbine characteristics and data distribution, including but not limited to:

[0076] (1) Data of wind turbines in shutdown or non-normal working mode, such as active power less than 0, wind speed greater than cut-out wind speed, and wind speed less than cut-in wind speed.

[0077] (2) Obviously unreasonable data caused by sensor failure, such as significantly exceeding the range.

[0078] (3) Use statistical methods to identify and remove noise points that are far from the normal distribution.

[0079] Outlier removal uses the Laida criterion algorithm, which specifically involves calculating the mean and standard deviation of the data, and removing outliers exceeding the specified range. "The range of samples, where" The mean of the dataset. denoted as the standard deviation of the sample.

[0080] Step S1.3: To eliminate the influence of differences in feature units and numerical ranges on the model, the cleaned data is normalized and scaled to the [0,1] interval. The formula used is:

[0081]

[0082] in, This is the original data. This is the minimum value of the state parameter. This is the maximum value of the state parameter.

[0083] Step S2: Feature priority ranking based on regularized mutual information (RMI), including the following steps:

[0084] Step S2.1: Based on expert experience, select a key parameter that can directly and sensitively reflect the core health status of the gearbox as the output variable of the state estimation model. For example, the lubricating oil temperature of a wind turbine gearbox.

[0085] Step S2.2: Use core parameters such as gearbox lubricating oil temperature as output variables of the state estimation model. Input features are quantized using RMI. and output variables The correlation for the training set Each input feature With output variables The formula for their RMI values ​​is as follows:

[0086]

[0087]

[0088]

[0089] in, It is a feature With output Mutual information between them; State parameters Information entropy State parameters The probability density function is obtained by kernel density estimation, where I(;) is the state parameter. and RMI is a standardized mutual information with a value range of [0,1], which can more fairly compare the importance of different features.

[0090] Step S2.3: Design a forward search strategy to prioritize feature parameters:

[0091] (1) Initialize an empty sorted set Used to store the selected features;

[0092] (2) Selection and output variables The highest eigenvalue of the RMI is denoted as It is included in the feature set as the primary feature. and from the feature candidate set Remove;

[0093] (3) In each subsequent iteration, features are selected according to the comprehensive priority score of “correlation with output - average redundancy of selected features” until the candidate features are empty, and a feature set is obtained in descending order of priority.

[0094] For each remaining feature in the candidate feature set Calculate its selection priority score The formula is as follows:

[0095]

[0096] This score takes into account both the correlation between the feature and the output variable (the higher the correlation between the feature and the output variable, the higher the model accuracy) and the average redundancy with the selected feature set (the lower the redundancy between input features, the more lightweight the model is and the more it avoids the "curse of dimensionality").

[0097] (4) Select the feature with the highest priority score and add it to the feature set. At the same time, it is removed from the candidate set;

[0098] (5) Repeat steps (3) and (4) until the candidate feature set is empty, and finally obtain the feature set sorted in descending order of priority. .

[0099] Step S3: Construct an improved ensemble KNN state estimation model. The KNN state estimation algorithm makes predictions by finding the K training samples that are closest to the test samples. For example... Figure 2 As shown, traditional KNN algorithms typically use Euclidean distance as a similarity metric. However, Euclidean distance is sensitive to outliers and fails to fully consider the global distribution characteristics of the feature space, easily leading the model to get trapped in local optima. Therefore, this invention fundamentally improves the traditional KNN algorithm by introducing affinity distance as a new similarity metric, significantly enhancing the model's accuracy and robustness. Figure 3 As shown, the improved steps are as follows:

[0100] Step S3.1: Improve the KNN state estimation algorithm; given the training set ,in and a test sample The state estimation process is as follows:

[0101] (1) For the test sample With each training sample Instead of calculating the traditional Euclidean distance, sample similarity is calculated using affinity distance, defined by the formula:

[0102]

[0103] in, It is an adaptive feature weight, determined by the distribution of the entire training set, and is calculated using the following formula:

[0104]

[0105] molecular Indicates test sample In features The total absolute distance to all training samples reflects the "periphery" of the test sample relative to the entire training set; the denominator contains... Indicates training samples In features The total absolute distance between the weight and all other training samples reflects the "intrinsic dispersion" of the training set on this feature. The physical meaning is that when a test sample differs significantly from the training set in a certain feature (large numerator), and the training set itself is highly concentrated on that feature (small second term in the denominator), that feature may carry abnormal information or noise. The value will be adaptively lowered, thereby reducing the negative impact of this anomalous feature on the overall distance calculation. This mechanism makes the distance metric more global and enhances the model's ability to combat outliers and local perturbations.

[0106] (2) Based on the calculated affinity distance, select samples from the training set that are similar to the test samples. The K most similar nearest neighbors form a set. .

[0107] (3) For the test sample, the average value of the output values ​​of the K nearest neighbors is used to obtain the test sample. The final estimated output :

[0108]

[0109] in, Let K be the set of nearest neighbor samples.

[0110] Step S3.2: Construct an integrated KNN model using a two-layer sampling mechanism.

[0111] The training set for each base learner is constructed through double-layer sampling, for the first... The steps for constructing the training set of each base learner are as follows:

[0112] (1) Feature sampling based on priority

[0113] First, generate a random number. ,satisfy ( (total number of features); subsequently, from the ordered feature set Before being selected These features constitute the feature subset of the base learner. This strategy ensures that each base learner focuses on the globally optimal combination of features, while introducing diversity by randomly adjusting the number of features.

[0114] (2) Sample sampling based on Bootstrap

[0115] In the defined feature subset From the original training set The Bootstrap method is used to perform random sampling with replacement in (or feature subsets) to draw a set of samples with the same number of samples as the original training set, forming the final training sample set for the base learner. .

[0116] (3) Repeat the above (1) and (2) double-layer sampling steps. Next, forming Different base training sets and training Each base learner.

[0117] Step S3.3: For a new test sample, aggregate all The outputs of each base learner are averaged to serve as the final prediction output of the ensemble KNN model, i.e., the key parameter estimate. :

[0118]

[0119] in, For the first The output of each base learner.

[0120] Step S4: Online real-time status monitoring and residual sequence generation.

[0121] The real-time collected device operation data stream, after undergoing the same preprocessing procedure as in step S1, is input into the pre-trained ensemble model to obtain key health parameters. Real-time estimates Calculate the estimated value and compare it with the actual sensor observation. residuals between :

[0122]

[0123] The residuals from continuous operation constitute the real-time residual sequence. When the gearbox is in a healthy state, the sequence should fluctuate slightly around zero.

[0124] Step S5: Setting dynamic early warning thresholds based on Statistical Process Control (SPC) technology. Calculate the sample mean of the validation set residual sequence using validation set data reserved from the gearbox's historical healthy operating conditions. and sample standard deviation Based on the normal distribution characteristics, and since gearbox failures often manifest as increased oil temperature, a gearbox failure early warning threshold is set. The calculation formula is:

[0125]

[0126] in, This is a coefficient set based on the system's tolerance for false alarm rates, i.e., the warning threshold. For example, when... At this point, corresponding to a 99% confidence level of the standard normal distribution, it means that under purely random fluctuations, the residual has only a 1% probability of naturally exceeding this threshold. This threshold serves as the primary standard for judging whether the equipment status is abnormal.

[0127] Step S6: Lifetime trend prediction and visualization based on sliding window failure rate, such as... Figure 4 As shown, it includes the following steps:

[0128] Step S6.1: Set a fixed length as A sliding time window that slides across the residual sequence by one sampling interval at a time.

[0129] Step S6.2: For each window position Calculate the residual values ​​within this window. Exceeding the warning threshold Number of samples The failure rate of the device corresponding to that window. Defined as:

[0130]

[0131] in, If the residual within the window exceeds the warning threshold The number of samples, This is the window length.

[0132] Step S6.3: Lifespan Stage Division and Early Warning

[0133] (1) Plot the failure rate A curve that changes over time or with the sliding window number;

[0134] (2) By analyzing the long-term trend of the curve, the equipment lifespan (health) is divided into three typical stages:

[0135] Healthy period: The failure rate curve fluctuates steadily at a low level (e.g., <5%), without a significant upward trend;

[0136] Early degradation stage: The failure rate curve shows a significant and continuous upward trend, and begins to oscillate widely in a high range (e.g., 20%~60%), indicating that the equipment has entered an unstable state;

[0137] Severe failure period: The failure rate curve remains at a high level (e.g., >80%) and does not fall back.

[0138] (3) Early warning realization: The critical point at which the failure rate curve first confirms the transition from the "healthy period" to the "early degradation period" is marked as the early warning point.

[0139] The above steps form a complete link from data processing to decision support, effectively solving the problems of insufficient accuracy, early warning and practicality in gearbox fault early warning and condition assessment.

[0140] The working process of the wind turbine gearbox lubrication system is as follows: Figure 5 As shown, this invention is applicable to various types of wind turbine generators, especially for condition monitoring and predictive maintenance of wind turbines in offshore or mountainous wind farms where the working environment is harsh and the maintenance cost is high.

[0141] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning and condition assessment of wind turbine gearbox faults, characterized in that... Includes the following steps: Step S1: Obtain wind turbine gearbox operating data; Step S2: Using the running data as input data and the key parameters selected from the running data as output data, compare the importance of different features in the running data through mutual information, and sort the features based on the forward search strategy to construct the training sample set and test sample set of the state estimation model. Step S3: Construct a state estimation model. Based on the affinity distance between the test sample and each training sample, select a set of nearest neighbor samples that are most similar to the test sample from the training sample set. Based on the output data of the nearest neighbor samples, predict the output data of the test sample. Construct a set of base learners for the integrated state estimation model through a two-layer sampling mechanism and train them through the training sample set. For a new test sample, predict the key parameter estimates based on the output results of all base learners. Step S4: Collect operational data in real time, estimate key parameters in real time using the trained ensemble state estimation model, calculate the residuals between the estimated key parameters and the actual sensor observation data, and obtain the real-time residual sequence; Step S5: Based on the verification sample set of the gearbox in historical healthy operating conditions, calculate the residual sequence of the verification sample set in order to set a dynamic early warning threshold; Step S6: Evaluate the gearbox status based on the number of samples in the residual sequence that exceed the warning threshold.

2. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 1, characterized in that: In step S2, the importance of different features is compared by calculating the mutual information value. The formula for the mutual information value is as follows: in, Indicates the characteristics of the input runtime data. This indicates the key parameter characteristics of the output. The information entropy of the feature is represented by p(·), and the probability density function is represented by p(·). Representing input features With output features Mutual information between them, min{·} represents the minimum value operation.

3. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 1, characterized in that: The forward search strategy in step S2 is as follows: In the initial iteration of the state estimation model, the input data with the highest mutual information with the output data is selected from the input data and placed into an empty ordered dataset as the primary data. In subsequent iterations of the state estimation model, a priority is constructed based on the correlation between the input data and the output data, as well as the average redundancy between the input data and the selected data in the ordered dataset, to select input data and obtain a data set arranged by priority. The data with the highest priority is selected from the data set and added to the ordered dataset, finally obtaining an ordered dataset arranged by priority.

4. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 3, characterized in that: The priority is measured by a priority score, as shown in the following formula: in, Indicates priority score, This represents the characteristics of the selected input data. This represents the number of features in an ordered dataset. This represents the features that have been selected in an ordered dataset.

5. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 3, characterized in that: In step S3, the construction process of the training sample set for the base learner is as follows: Feature sampling is performed based on priority, and a set of high-priority data is randomly selected from the ordered dataset to form a training sample subset for the base learner; Perform effective replacement random sampling from the original training sample set or the subset of training samples to extract a set of base learner training samples with the same number of samples as the original training sample set.

6. The method for early warning and status assessment of wind turbine gearbox faults according to claim 5, characterized in that: For a new test sample, the output data of all base learners are aggregated, and their average is calculated as the final predicted output data of the ensemble state estimation model.

7. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 1, characterized in that: The affinity distance in step S3 is calculated using the following formula: Where d(,) represents the affinity distance, m represents the number of test samples, and x tj Let x represent the j-th test sample. ij This represents the i-th training sample corresponding to the j-th test sample. represents the adaptive feature weights, and n represents the number of training samples.

8. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 1, characterized in that: In step S5, the mean and standard deviation of the residual sequence of the verification sample set are calculated using the verification sample set of the gearbox under historical healthy operating conditions. Based on the normal distribution characteristics, the gearbox fault warning threshold is set.

9. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 1, characterized in that: In step S6, by setting a sliding window, the number of samples whose corresponding residual values ​​exceed the warning threshold is counted at each window position. Based on the ratio of the number of samples exceeding the threshold to the length of the sliding window, the gearbox failure rate is obtained. Based on the long-term trend curve of the failure rate changing with time or the sliding window number, the gearbox life is divided. The failure rate curve fluctuating steadily at a low level without a significant upward trend is classified as a healthy period. The failure rate curve showing an upward trend and starting to oscillate widely in the high range is classified as an unstable state. The failure rate curve maintaining a high level and no longer falling is classified as a serious failure. The critical point where the failure rate curve is first confirmed to change from the healthy period to the early degradation period is used as the early warning point.

10. The method for early warning and condition assessment of wind turbine gearbox faults according to claim 1, characterized in that: The historical multidimensional cross-sectional data of the gearbox under known healthy operating conditions are obtained and preprocessed, including the following steps: Step S1.1: Remove missing values; Step S1.2: Perform data cleaning based on wind turbine characteristics and data distribution, including but not limited to: Data when the wind turbine is in a shutdown or abnormal operating mode; Clearly unreasonable data caused by sensor malfunction; Identify and remove noise points that are far from the normal distribution; Step S1.3: Normalize the cleaned data.

Images