A method for early weak fault detection in pitch systems based on multi-index reconstruction

By using data cleaning and operating condition classification based on multi-index reconstruction, combined with JS divergence reconstruction, the complexity of early weak fault detection in wind turbine pitch system is solved, and more accurate and timely fault identification is achieved.

CN117605619BActive Publication Date: 2026-11-17SHENYANG INST OF ENG
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311146581.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-06
Publication Date
2026-11-17
Estimated Expiration
2043-09-06

AI Technical Summary

Technical Problem

Early minor faults in the pitch system of wind turbines are difficult to separate from frequent switching of operating conditions and noise signals, making detection challenging.

Method used

A multi-index reconstruction approach is adopted, including data cleaning, working condition division, and JS divergence reconstruction, to achieve early anomaly detection through offline modeling and online evaluation.

Benefits of technology

It improves the accuracy and timeliness of detecting early minor faults in the pitch system, reduces the false alarm rate, and enhances the sensitivity to real abnormal signals.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117605619B_ABST
    Figure CN117605619B_ABST
Patent Text Reader

Abstract

This invention belongs to the field of fault detection for key components of wind turbines, specifically relating to a method for detecting early, weak faults in pitch systems based on multi-index reconstruction. This method identifies data deviations caused by sudden changes in pitch system state, noise signals, and actual anomalies. It utilizes data cleaning and operating condition segmentation based on power curves, employing distributed modeling to address the complexity of operating conditions and mitigate false alarms caused by sudden changes. By reconstructing the J-S divergence under approximate operating conditions, it balances the sensitivity and volatility of evaluation indices to address the impact of noise data. This effectively mitigates evaluation index deviations caused by complex operating condition changes and noise data, while simultaneously improving sensitivity to actual anomaly signals. Consequently, it more accurately and promptly detects early, weak anomalies in the pitch system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of fault detection of key components of wind turbines, and specifically relates to an early weak fault detection method for pitch systems based on multi-index reconstruction that can identify data offset problems caused by sudden changes in the state of the pitch system, noise signals, and real anomalies. Background Technology

[0002] Wind energy, with its green and pollution-free advantages, has become one of the fastest-growing renewable energy sources today. However, the continuous expansion of wind power installed capacity has also increased the pressure on wind turbine operation and maintenance. Failures in key components of wind turbine units severely impact power generation efficiency and cause significant economic losses, making it a crucial task for operation and maintenance. Timely and accurate detection of faults in key wind power components is an effective means of reducing operation and maintenance costs. Currently, there is extensive research on wind power fault detection methods, including those based on signal processing, image recognition, machine learning, and deep learning.

[0003] Statistical analysis of wind turbine reliability reveals that the pitch system is a high-risk component for failure. Frequent downtime caused by the pitch system can lead to significant economic losses. Furthermore, pitch system failures can directly result in uncontrollable loads on the turbine, potentially causing serious and destructive accidents. Currently, fault detection methods for wind turbine pitch systems can be categorized into data-based and model-based approaches. Data-based methods have the advantage of not relying on prior process models or domain experts. Moreover, with the rapid development of information acquisition, transmission technologies, and data mining methods, data-driven process fault detection technologies have received widespread attention.

[0004] The pitch control system of a wind turbine is a typical device operating under complex conditions. Subtle changes in the early stages of a fault are often hidden in data offsets caused by sudden changes in operating conditions or noise. Therefore, separating early abnormal signals from errors caused by sudden changes in conditions and noise signals is an important way to improve the performance of early fault detection. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for detecting early weak faults in pitch systems based on multi-index reconstruction, which solves the problem that early abnormal signals of wind turbine pitch systems have weak non-stationarity and are easily confused by the frequent switching of operating conditions, making them difficult to detect in a timely manner.

[0006] This invention is implemented as follows:

[0007] A method for detecting early-stage weak faults in a pitch system based on multi-index reconstruction, comprising the following steps:

[0008] Step (1) Clean the offline data based on the theoretical power curve;

[0009] Step (2) performs data homogenization and working condition division on the cleaned offline data based on secondary clustering of the pitch system to obtain several sub-working conditions;

[0010] Step (3) Use the JS divergence reconstruction of offline samples to perform offline modeling;

[0011] Step (4) Determine the sub-condition threshold by using the offline reconstruction index to assess the preset coverage of normal samples;

[0012] Step (5) The online data is evaluated online using the JS divergence reconstruction index of the samples within the online window to obtain the working condition reconstruction results of the online samples;

[0013] Step (6) uses the sub-condition threshold of step (4) as the condition threshold of the online sample. The condition reconstruction result of the online sample is compared with the condition threshold of the online sample to achieve early anomaly detection.

[0014] Furthermore, step (1) specifically includes:

[0015] When normal operating data is sufficient to cover all operating conditions, isolated points are removed from the wind speed-power curve to make the curve more closely approximate the ideal state, including:

[0016] Sample points below the cut-in wind speed and at zero power are removed. Using the theoretical power curve as the standard, abnormal operating points that are far from it are also removed. The theoretical power curve is discretized at intervals of 0.1 m / s to obtain the theoretical power curve sample set G. I The dataset G collected by SCADA S Calculate the G value for each sample. I The nearest neighbor distance D of the set SI ,Right now

[0017]

[0018] Among them, y i ∈G I For x i ∈G S The nearest neighbor samples, where m is the feature dimension, i = 1, 2, ..., n, and n is the total number of offline samples, for D SI The samples are sorted in ascending order, and the nearest neighbor distance of the α1%×nth sample is used as the threshold α1 for removing abnormal operation samples. Finally, a feasible region is formed centered on the theoretical power curve, and the width of the region is determined by the threshold α1.

[0019] Furthermore, in step (2), the specific steps for dividing the working conditions are as follows:

[0020] The wind speed-power curve in the data acquisition and monitoring control system is divided into several segments based on a width of 0.1 m / s. A k-means clustering operation is then performed on the samples in each segment. The clustering selection for each sub-block is as follows:

[0021]

[0022] Where n E n is the desired amount of experimental data. W To achieve data homogenization, the cluster center data of each sub-block is retained and other data is removed, based on the number of sub-blocks divided according to wind speed.

[0023] For the homogenized sample set Q2, quadratic k-means clustering is used to divide similar working conditions. The clustering validity function is as follows:

[0024]

[0025] disMean out (i) is the average distance from data i to points outside the current class, disMean in (i) is the average distance from data i to points in this class. -1<=S C (i) <= 1, and the closer to 1, the larger the distance between classes and the smaller the distance within classes, and the better the clustering effect.

[0026] Furthermore, in step (3), offline modeling is performed using the JS divergence reconstruction of offline samples, specifically including:

[0027] Step 31: Divide the offline samples into n offline samples according to the specified window width;

[0028] Step 32: Determine all sample sets P within each window i The nearest neighbor sample set Q i This forms a comparative JS distribution data set;

[0029] Step 33: Calculate the sample P within the window i Compared with the comparison dataset Q i JS divergence:

[0030]

[0031] Where KL is the KL divergence, calculated as follows:

[0032]

[0033] Step 34: Calculate the stationarity and volatility of the JS divergence for all offline window samples:

[0034] Specify the window width and define two statistical indicators of the nearest neighbor distance of sample points within the window: mean offset I. d Volatility I v The hypothesis testing function and activation function are used to transform each index into a probabilistic form, and the product of the two is used as the comprehensive evaluation index I. re ,Right now

[0035] I re =P d ×P v

[0036] Where P d P v I d I v The probability form;

[0037] Average offset I d This indicates the degree of deviation from the mean nearest neighbor distance, and its calculation form and probability representation are as follows:

[0038] I d =(μ-μ * ) / σ *

[0039] P d =2*Φ(I d )-1

[0040] Where Φ is the cumulative distribution function of the standard normal distribution;

[0041] Volatility I v The measure of the fluctuation in nearest neighbor distance is as follows:

[0042] I v =σ 2 / σ *2

[0043] P v =F(I v ,n1,n * )

[0044] Where F is the cumulative distribution function of the F distribution, n1 is the number of samples in the window, and n * P represents the number of test samples. d For average offset I d The probability form, P v For volatility I v The probability form.

[0045] Furthermore, in step (4), the index probability reconstruction values ​​of all offline samples are used as evaluation indicators, and the sub-condition threshold D of confidence α is estimated independently using the chilimit function in Matlab within the PLS toolbox for each sub-condition. α 2 .

[0046] Furthermore, step (5) specifically includes:

[0047] Step 51: For the online data window sample, determine its nearest neighbor operating conditions in the offline sample;

[0048] Step 52: Calculate the JS divergence between the online data and the data in the nearest neighbor conditions;

[0049] Step 53: Calculate the average offset and volatility of the JS divergence of the set of windows with all samples as right endpoints within the window;

[0050] Step 54: Reconstruct the metrics of the data within the window.

[0051] Furthermore, in step (6), the threshold of each sample in the sub-condition within the window is used as a dynamic threshold and compared with the reconstructed index. If it is lower than the dynamic threshold, it is a normal state; otherwise, it is an abnormal state.

[0052] Compared with the prior art, the beneficial effects of this invention are as follows:

[0053] The operating states of wind turbine pitch systems are complex, and the distinctions between different operating conditions are blurred. Traditional normal behavior models struggle to avoid the impact of sudden changes in complex operating conditions. The ability to separate data offsets caused by sudden changes in complex operating conditions, noise signals, and genuine weak anomalies is crucial for improving early fault detection performance. This paper proposes a method that utilizes data cleaning and operating condition segmentation based on power curves, employing distributed modeling to address the complexity of operating conditions and mitigate false alarms caused by sudden changes. Furthermore, it uses JS divergence reconstruction under approximate operating conditions to balance the sensitivity and volatility of evaluation indicators, thus addressing the impact of noise data. These two innovative approaches effectively mitigate evaluation indicator offsets caused by sudden changes in complex operating conditions and noise data, while simultaneously improving sensitivity to genuine anomalies. This leads to more accurate and timely detection of early weak anomalies in the pitch system.

[0054] This invention provides a data cleaning and operating condition classification method based on power curves and secondary clustering, which preserves the operating mechanism of the pitch system and improves the authenticity and uniformity of the data.

[0055] Anomaly assessment based on JS divergence reconstruction can effectively balance the volatility and sensitivity of assessment indicators, mitigate false alarms caused by sudden changes in operating conditions and noisy data, and improve the sensitivity to real, subtle anomalies. Attached Figure Description

[0056] Figure 1 The flowchart shows a method for detecting early weak faults in a pitch system based on multi-index reconstruction.

[0057] Figure 2 This is a scatter plot of actual wind speed-power for a 2MW unit in an embodiment of the present invention.

[0058] Figure 3 This is the theoretical power curve of a 2MW wind turbine in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0060] See Figure 1 As shown, a method for detecting early minor faults in a pitch system using multi-index reconfiguration is presented, the method comprising the following steps:

[0061] Step (1) Clean the offline data based on the theoretical power curve;

[0062] Step (2) performs data homogenization and working condition division on the cleaned offline data based on secondary clustering of the pitch system to obtain several sub-working conditions;

[0063] Step (3) Use the JS divergence reconstruction of offline samples to perform offline modeling;

[0064] Step (4) Determine the sub-condition threshold by using the offline reconstruction index to assess the preset coverage of normal samples;

[0065] Step (5) The online data is evaluated online using the JS divergence reconstruction index of the samples within the online window to obtain the working condition reconstruction results of the online samples;

[0066] Step (6) uses the sub-condition threshold from step (4) as the condition threshold for the online sample. The condition reconstruction result of the online sample is compared with the condition threshold of the online sample to achieve early anomaly detection. In step (1), the specific cleaning steps are as follows:

[0067] When normal operating data sufficiently covers all operating conditions, isolated points in the wind speed-power curve may be noise points or anomalies. Therefore, it is necessary to remove isolated points to make the wind speed-power curve more approximate the ideal state. First, sample points below the cutoff wind speed and zero power are removed. Then, using the theoretical power curve as the standard, abnormal operating points that are far from it are removed. The theoretical power curve is discretized at intervals of 0.1 m / s to obtain the theoretical power curve sample set G.I The dataset G collected in SCADA (Supervisory Control and Data Acquisition) system. S Calculate the G value for each sample. I The nearest neighbor distance D of the set SI ,Right now

[0068]

[0069] Among them, y i ∈G I For x i ∈G S The nearest neighbor samples, where m is the feature dimension, i = 1, 2, ..., n, and n is the total number of offline samples, for D SI The samples are sorted in ascending order, and the nearest neighbor distance of the α1%×nth sample is used as the threshold for removing abnormal operation samples. This method forms a feasible region centered on the theoretical power curve, and the width of the region is determined by the threshold α1.

[0070] In step (2), the specific steps for data homogenization and working condition division are as follows:

[0071] In the fault detection process, highly overlapping data can directly affect the judgment of the state. In order to reduce the impact of duplicate data or highly similar data on the subsequent fault detection modeling process, this section proposes a secondary clustering method to achieve data homogenization and working condition classification.

[0072] First, the SCADA wind speed-power mapping map is divided into several segments with a width of 0.1 m / s. Then, a k-means clustering is performed on the samples in each segment. The clustering selection for each sub-block is as follows:

[0073]

[0074] Where n E n is the desired amount of experimental data. W This represents the number of sub-blocks divided according to wind speed. Cluster center data within each sub-block is retained, while other data is removed to achieve data homogenization.

[0075] For the homogenized sample set Q2, quadratic k-means clustering is used to divide similar working conditions. The clustering validity function is as follows:

[0076]

[0077] disMean out (i) is the average distance from data i to points outside the current class, disMean in(i) is the average distance from data i to points in this class. -1<=S C (i) <= 1, and the closer to 1, the larger the distance between classes and the smaller the distance within classes, and the better the clustering effect.

[0078] In step (3), the specific steps of the offline model building stage are as follows:

[0079] Step 31: Divide the offline samples into n offline samples according to the specified window width.

[0080] Step 32: Determine all sample sets P within each window i The nearest neighbor sample set Q i This forms a comparative JS distribution data set;

[0081] Step 33: Calculate the sample P within the window i Compared with the comparison dataset Q i JS divergence:

[0082]

[0083] Where KL is the KL divergence, calculated as follows:

[0084]

[0085] Step 34: Calculate the average bias and volatility of the JS divergence for all offline window samples:

[0086] Specify the window width (selecting a data sampling interval of 10s, therefore setting the window width to 1h), and define two statistical indicators of the nearest neighbor distance of sample points within the window: mean skewness I. d Volatility I v Since the two are independent of each other, a hypothesis testing function and an activation function are used to transform each index into a probabilistic form, and the product of the two is used as the comprehensive evaluation index I. re ,Right now

[0087] I re =P d ×P v

[0088] Where P d P v I d I v The probability form.

[0089] Average Deviation Index I d The deviation of the mean of nearest neighbor distance is represented by the following calculation form and probability representation.

[0090] I d =(μ-μ * ) / σ*

[0091] P d =2*Φ(I d )-1

[0092] Where Φ is the cumulative distribution function of the standard normal distribution.

[0093] Volatility Indicator I v This indicates the degree of fluctuation in nearest neighbor distance, and its measurement form is:

[0094] I v =σ 2 / σ *2

[0095] P v =F(I v ,n1,n * )

[0096] Where F is the cumulative distribution function of the F distribution, n1 is the number of samples in the window, and n * P represents the number of test samples. d Deviation index I d The probability form, P v Volatility Indicator I v The probability form.

[0097] In step (4), the confidence threshold D can be estimated using the chilimit function in Matlab from the PLS toolbox. α 2 .

[0098] In step (5), the specific steps for the online status detection stage of the online data are as follows:

[0099] Step 51: For the online data window sample, determine its nearest neighbor operating conditions in the offline sample;

[0100] Step 52: Calculate the JS divergence between the online data and the data in the nearest neighbor conditions;

[0101] Step 53: Calculate the JS divergence offset and volatility indices for the set of windows with all samples as their right endpoints within the window;

[0102] Step 54: Reconstruct the metrics for the data within the window;

[0103] In step (6), the sub-condition threshold of step (4) is used as the condition threshold of the online sample. The condition reconstruction result of the online sample is compared with the condition threshold of the online sample. If it is lower than the sub-condition threshold of step (4), it is a normal state; otherwise, it is an abnormal state.

[0104] The implementation process of this invention is illustrated using common pitch motor overheating faults, pitch angle abnormality faults, and pitch angle asymmetry faults in wind turbine pitch systems as examples. Fault descriptions are shown in Table 1. SCADA experimental data were collected from a 2MW direct-drive wind turbine at the Dalian Tuoshan Wind Farm, with 25,000 sets of operational data collected between February and March 2014 at 5-minute intervals.

[0105] Table 1. Description of Pitch System Faults

[0106]

[0107] Fifteen monitoring variables related to the pitch system were selected for fault detection in the SCADA system, as shown in Table 2. Among them, wind speed, power, and generator speed are key variables for measuring the operating status of the wind turbine, pitch motor current and temperature are key state variables for monitoring pitch motor over-temperature faults, and encoder position and redundant encoder position are key state variables for monitoring pitch angle abnormal faults and pitch angle asymmetry.

[0108] Table 2. SCADA monitoring variables related to the pitch system

[0109]

[0110] 25,000 sets of operational data were collected from Unit 7 of the Dalian Tuoshan Wind Farm between February and March 2014 at 5-minute intervals. After removing points with wind speeds lower than the cut-in speed and low power, 20,380 sample points remained. The remaining samples were then cleaned using the method proposed in step (1). Figure 2 The image shows a scatter plot of wind speed and power for a 2MW wind turbine operating for 3 months. It contains a large amount of data on shutdown status, such as the sample points in the dashed circles, as well as abnormal operating points and noise data, such as the samples in the solid circles in the image. Figure 3 This is the theoretical power curve for this model.

[0111] This invention detects three types of pitch system faults. Considering that the wind turbine's state parameters may have already become abnormal before the SCADA system issues an alarm, this experiment only counts the false alarm rate of the first 300 samples and the false alarm rate of the last 200 samples. For the three types of pitch system faults shown in Table 1, Tables 3 and 4 compare the false alarm rate and false alarm rate of the fault detection method of this invention with those of the PCA and KPCA fault detection methods, respectively.

[0112] Table 3 Comparison of Fault False Reporting Rates in Pitch System

[0113]

[0114] Table 4 Comparison of false alarm rates for pitch system faults

[0115]

[0116] As shown in Tables 3 and 4, when fault characteristics are detected, the PCA fault detection method has poor handling capabilities for nonlinearities, resulting in unsatisfactory detection results for all three types of faults, with high false alarm and false false alarm rates. The fault detection mechanism proposed in this invention can better handle the data characteristics of wind turbine multi-condition operation modes by adopting a "divide and conquer" approach.

[0117] Furthermore, for fault 1, fault 2, and fault 3, the detection method of this invention is at least 8.2 h, 14.5 h, and 6.4 h earlier than other comparative methods, respectively, verifying the sensitivity of this invention to real, subtle anomalies.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early weak fault detection of a variable pitch system based on multi-index reconstruction, characterized in that, The method includes the following steps: Step (1) Clean the offline data based on the theoretical power curve; Step (2) performs data homogenization and operating condition division on the cleaned offline data based on secondary clustering of the pitch system, resulting in several sub-operating conditions; the specific steps for operating condition division in step (2) are as follows: The wind speed-power curve in the data acquisition and monitoring control system is divided into several segments based on a width of 0.1 m / s. A k-means clustering operation is then performed on the samples in each segment. The clustering selection for each sub-block is as follows: , where n E is the desired amount of experimental data, n W is the number of sub-blocks according to wind speed, the cluster center data in each sub-block is retained, and other data is removed to achieve data uniformization; For the homogenized sample set Q2, quadratic k-means clustering is used to divide similar working conditions. The clustering validity function is as follows: , where disMean out (i) is the average distance of data i to non-class points, disMean in (i) is the average distance of data i to class points, -1 <= S C (i) <= 1, and the closer to 1, the greater the inter-class distance, the smaller the intra-class distance, and the better the clustering effect. Step (3) Use the JS divergence reconstruction of offline samples to perform offline modeling; Step (4) Determine the sub-condition threshold by using the offline reconstruction index to assess the preset coverage of normal samples; Step (5) The online data is evaluated online using the JS divergence reconstruction index of the samples within the online window to obtain the working condition reconstruction results of the online samples; Step (6) uses the sub-condition threshold of step (4) as the condition threshold of the online sample. The condition reconstruction result of the online sample is compared with the condition threshold of the online sample to achieve early anomaly detection.

2. The method of claim 1, wherein, Step (1) specifically includes: When normal operating data is sufficient to cover all operating conditions, isolated points are removed from the wind speed-power curve to make the curve more closely approximate the ideal state, including: The sample points below the cut-in wind speed and zero power are removed, and the non-normal operating points far from the theoretical power curve are removed. The theoretical power curve is discretized at intervals of 0.1 m / s to obtain a theoretical power curve sample set G I The data set G collected by the SCADA S The nearest neighbor distance D of each sample in G to the set G is calculated I SI , that is​ , Among them, y i ∈G I For x i ∈G S The nearest neighbor samples, m is the feature dimension, i=1, 2,…, n, where n is the total number of offline samples, for D SI The samples are sorted in ascending order, and the nearest neighbor distance of the α1%×nth sample is used as the threshold α1 for removing abnormal operation samples. Finally, a feasible region interval centered on the theoretical power curve is formed, and the width of the interval is determined by the threshold α1.

3. The method for detecting early weak faults in a pitch system based on multi-index reconstruction according to claim 1, characterized in that, In step (3), offline modeling is performed using the JS divergence reconstruction of offline samples, specifically including: Step 31: Divide the offline samples into n offline samples according to the specified window width; Step 32: Determine all sample sets P within each window i of the neighborhood sample sets Q i , constituting the contrast JS distribution data set; Step 33: Compute the sample P within the window i JS divergence from the contrast dataset Q i JS divergence: , Where KL is the KL divergence, calculated as follows: ; Step 34: Calculate the stationarity and volatility of the JS divergence for all offline window samples: Specify the window width and define two statistical indicators of the nearest neighbor distance of sample points within the window: mean offset I. d Volatility I v The hypothesis testing function and activation function are used to transform each index into a probabilistic form, and the product of the two is used as the comprehensive evaluation index I. re ,Right now , where P d , P v are the probability forms of I d , I v , respectively; Average deviation I d represents the average deviation of the distance of the nearest neighbors, which is calculated as follows: , , Where Φ is the cumulative distribution function of the standard normal distribution; Volatility I v represents the degree of fluctuation of the distance of the nearest neighbors, which is measured as: , , Where F is the cumulative distribution function of the F distribution, n1 is the number of samples in the window, and n * P represents the number of test samples. d For average offset I d The probability form, P v For volatility I v The probability form.

4. The method of claim 1, wherein, In step (4), the index probability reconstruction values ​​of all offline samples are used as evaluation indicators, and the sub-condition threshold D of confidence α is estimated independently using the chilimit function in Matlab within the PLS toolbox for each sub-condition. α 2 .

5. The method for detecting early weak faults in a pitch system based on multi-index reconstruction according to claim 1, characterized in that, Step (5) specifically includes: Step 51: For the online data window sample, determine its nearest neighbor operating conditions in the offline sample; Step 52: Calculate the JS divergence between the online data and the data in the nearest neighbor conditions; Step 53: Calculate the average offset and volatility of the JS divergence of the set of windows with all samples as right endpoints within the window; Step 54: Reconstruct the metrics of the data within the window.

6. The method for detecting early weak faults in a pitch system based on multi-index reconstruction according to claim 1, characterized in that, In step (6), the threshold of each sample in the sub-condition within the window is used as the dynamic threshold and compared with the reconstructed index. If it is lower than the dynamic threshold, it is a normal state; otherwise, it is an abnormal state.

Citation Information

Patent Citations

  • Wind power variable pitch system fault detection method oriented to complex operation conditions

    CN114237206A

  • Online identification method for measuring point high-dimensional abnormal operation data of key equipment of wind turbine generator

    CN116662916A