A method of adjusting a wind turbine variable pitch system fault assessment value

Abstract

The application discloses a method for adjusting fault evaluation value of a variable pitch system of a wind turbine generator set, obtains SCADA data of normal operation of the wind turbine generator set, extracts common parameters of three blades in the variable pitch system, obtains transformed two-dimensional features, constructs a training model for training, obtains a health threshold, obtains a state index after averaging, compares the state index with the health threshold, adjusts the fault evaluation value when the state index of test data continuously exceeds the threshold for 24 points, and thus an alarm is sent. The application has the advantages of accurate prediction.

Description

Technical Field

[0001] This invention belongs to the field of condition monitoring and fault diagnosis of new energy power generation equipment, and in particular relates to a method for adjusting the fault assessment value of the pitch system of a wind turbine. Background Technology

[0002] As one of the most important systems in a wind turbine, the rotor is responsible for converting wind energy and protecting the unit from damage by strong winds. To better utilize wind energy and control the unit's operation, modern large-scale wind turbines employ active pitch control to manage the blades, ensuring high wind energy conversion efficiency and preventing rotor overspeed even when actual wind speeds exceed rated wind speeds. However, due to frequent changes in actual wind conditions and the blades operating under complex and variable loads year-round, failures are frequent. According to statistics, failures related to the blade pitch system account for 13% of all failures in wind turbines, second only to electrical and sensor failures. SCADA systems are widely used for wind turbine condition monitoring and control, with data acquisition frequencies of once per minute or every 10 minutes. While they cover a wide range of measurement points, the low sampling frequency makes it difficult to diagnose component failures using traditional signal analysis methods. Thanks to the development of artificial intelligence, numerous machine learning algorithms and neural network models have been introduced into the field of industrial inspection. Many fault detection algorithms have emerged for wind turbine generators, but these algorithms primarily focus on components such as gearboxes and generators. Relatively little work utilizes SCADA data for pitch system fault identification. Pitch control refers to achieving maximum power generation by changing the pitch angle of the wind turbine blades.

[0003] Patent document CN115079626 discloses a method and system for early warning of potential operational risks of wind turbine components. This method utilizes historical data of the wind turbine for statistical analysis, identifying data characteristics of corresponding components under abnormal conditions based on historical data and current values. Based on these abnormal data characteristics, it performs data screening of the wind turbine's current operating data to determine whether the component's operating status is normal under different operating conditions and provides early warnings for abnormal operating states. This invention solves the problem in existing technologies that it does not provide early warnings for abnormal component performance caused by the component's operating logic and external wind conditions, even when the component itself is not faulty. Patent document CN103410660B discloses a sliding mode variable pitch control method based on support vector machines. The learning control is divided into two steps. The first step uses a conventional sliding mode controller (SMC) for control. The SVM-SMC controller learns the controller's structure and preliminary parameters through support vector machines. When the learning reaches a certain level and the approximation error of the SVM-SMC to the SMC is less than a threshold, the variable pitch system switches to SVM-SMC control. The second step employs an exploratory mechanism, where the actual control quantity consists of the output of the SVM-SMC controller plus a random disturbance with a zero-mean normal distribution. Learning samples are obtained based on predicted performance indicators, and the control parameters are optimized in real time using an online learning algorithm. This invention not only possesses the advantages of sliding mode control, such as strong anti-interference capabilities, robustness to changing parameters, and high speed, but also ensures stable power output near the rated value while achieving smooth pitch angle adjustment, reducing unit fatigue and wear between components.

[0004] To extract SCADA data and enable monitoring and early warning of pitch systems with high failure rates, this invention proposes a method for adjusting the fault assessment value of wind turbine pitch systems. This method helps wind farm maintenance personnel predict and locate faults, rationally organize maintenance plans, avoid huge downtime losses caused by subsequent maintenance, and improve maintenance efficiency. Summary of the Invention

[0005] To achieve fault identification in the pitch control system of wind turbine generators, reduce unnecessary downtime caused by pitch control system failures, and provide maintenance personnel with maintenance suggestions, this invention discloses a method for adjusting the fault assessment value of the pitch control system of wind turbine generators, comprising the following steps:

[0006] Step 1: Obtain vibration signals from the wind turbine during normal operation. The vibration signals are SCADA data.

[0007] Step 2: Extract the common parameters of the three blades in the pitch system to obtain the transformed two-dimensional features;

[0008] Step 3: Construct a single-class support vector machine model, using the Gaussian kernel function as the kernel function of the model, and train the model using the two-dimensional features obtained in Step 2 as input data;

[0009] Step 4: Substitute normal data into the trained model to calculate its state index value under normal conditions, and then use the 3σ rule to calculate the threshold.

[0010] Step 5: Calculate the health threshold using status indicators from normal data;

[0011] Step 6: Obtain all samples that exceed the threshold, and calculate the cosine similarity between the two-dimensional data of each sample after projection transformation and the corresponding two-dimensional data center of the normal sample set relative to the coordinate axis.

[0012] Step 7: Adjust the blade fault assessment value based on the sum of cosine similarity.

[0013] Furthermore, the normal operation data in step 1 does not include the following data: the aforementioned fault, shutdown, and non-operation data.

[0014] Furthermore, the projection transformation in step 2 refers to: for the common parameters of the three blades in the pitch system, each type of parameter is plotted on three plane coordinate axes that are 120 degrees apart and share the same origin, and the centroid of the triangle formed by the three points is projected onto a plane rectangular coordinate system with equal scale and origin to obtain the transformed two-dimensional feature.

[0015] Furthermore, the single classifier constructed in step 3 is a single-class support vector machine, which uses the Gaussian kernel function as the kernel function of the model and uses all the projected feature values ​​as input data to train the model.

[0016] Furthermore, the calculation of the health threshold in step 5 is determined using the 3σ rule. Specifically, it involves first substituting normal data into the trained model to calculate its state index value under normal conditions, and then calculating the threshold using the following formula: ,in , y i This indicates a normal data status indicator, and n represents the data length.

[0017] Furthermore, following step 5, steps 5.1 and 5.2 are also included:

[0018] Step 5.1: During testing, the test data is processed according to steps 1 and 2, then fed into the trained single-class support vector machine model and the state index is calculated. The state index is smoothed using the exponential weighted moving average method, and the average value is obtained by aggregating the data hourly after smoothing.

[0019] Step 5.2: Compare the smoothed and averaged status indicators with the health threshold. When the test data status indicators exceed the threshold for 24 consecutive points, an alarm is issued.

[0020] In step 5.2, the Euclidean distance between each projected feature point and the decision boundary obtained from model training is used as the state index.

[0021] Furthermore, in the test data diagnosis process described in step 5.1, the calculated state index is smoothed using an exponentially weighted moving average. Specifically, for the original state index y... i Its smoothing value Where λ is the weighting coefficient, i>1, μ is the mean of the state index of normal data, obtained from step 4.

[0022] Furthermore, in step 5.2, when issuing an early warning, an alarm is issued when 24 consecutive status indicators exceed the threshold.

[0023] Furthermore, in step 6, during the fault location process, the out-of-limit data is extracted as a location sample. The cosine similarity of the line vector connecting the center of the two-dimensional feature of each type of parameter and the center of the two-dimensional feature of the corresponding parameter in the normal data with respect to the coordinate axis is calculated, and its absolute value is taken. The aforementioned coordinate axis is the coordinate axis that is 120 degrees apart from each other used in step 2.

[0024] Furthermore, in step 7, after calculating the cosine similarity of all parameters relative to the three coordinate axes, the sum of the cosine similarity under each coordinate axis is calculated, and the fault assessment value of the blade is adjusted according to the sum of the cosine similarity. The fault assessment value of the blade with the highest cosine similarity is adjusted to above the warning value.

[0025] This invention discloses a method for detecting faults in a pitch system using SCADA data. Its advantages include: proposing a feature reconstruction method based on common parameters of the three blades, using the centroid of the triangle formed by these parameters to represent the state, eliminating the influence of operating conditions and improving the data signal-to-noise ratio. Furthermore, based on this parameter reconstruction method, corresponding methods for calculating state indicators and locating faulty blades are proposed, making model diagnosis efficient and interpretable. Attached Figure Description

[0026] Figure 1 The graph showing the change of state indicators over time in an embodiment of the present invention is shown.

[0027] Figure 2 A two-dimensional coordinate scatter plot of the unit drive temperature projection is shown in an embodiment of the present invention. Detailed Implementation

[0028] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the process equipment or apparatus not specifically specified in the following embodiments are all conventional equipment or apparatus in the art. Furthermore, it should be understood that one or more method steps mentioned in the present invention do not exclude the existence of other method steps before or after the combined steps, or the insertion of other method steps between these explicitly mentioned steps, unless otherwise stated; it should also be understood that the combined connection relationship between one or more devices / apparatus mentioned in the present invention does not exclude the existence of other devices / apparatus before or after the combined devices / apparatus, or the insertion of other devices / apparatus between these explicitly mentioned two devices / apparatus, unless otherwise stated. Moreover, unless otherwise stated, the numbering of each method step is only a convenient tool for identifying each method step, and not for limiting the order of the method steps or limiting the scope of the present invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the present invention.

[0029] The unit described in this embodiment of the invention is a 2.5MW doubly-fed wind turbine. During deep maintenance of the pitch system on May 28, 2019, it was discovered that the pitch driver wiring terminals in pitch control cabinet #3 were burnt out. Replacing the pitch driver resolved the fault. The specific implementation process of this invention is as follows:

[0030] Step 1: Check the unit's maintenance records and find that there are no records of pitch system failures before November 2018. Therefore, take the SCADA data from April 2018 to November 2018 as normal data, and then delete the data with power less than or equal to 0 and hub speed less than or equal to 0.

[0031] Step 2: Data Preprocessing. Extract common parameters for the three blades in the pitch system: blade motor temperature, blade control cabinet temperature, driver temperature, and blade radiator temperature. Plot each type of parameter on three mutually perpendicular coordinate axes sharing a common origin. Project the centroid of the triangle formed by these three points onto a Cartesian coordinate system with equal scale and origin to obtain the transformed two-dimensional features. That is, for the parameters... The features after projection transformation are: , After the four parameter transformations are completed, eight newly constructed features are obtained, and these eight features are concatenated into a single vector. This refers to the features after projection transformation.

[0032] Step 3: Construct and train the model. Construct a single-class support vector machine model, using a Gaussian kernel function as the model's kernel function, with the support vector error limited to 0.01. Train the model using all projected feature values ​​as input data, with the input data dimension being 12394×8.

[0033] Step 4: Threshold Calculation. Substitute normal data into the trained model to calculate its state index value under normal conditions, then use 3... σ The rule calculates the threshold, and the calculation formula is as follows: ,in , , y i Indicates status indicators, n This indicates the data length, which is 12394.

[0034] Step 5: Take test data from December 2018 to July 2019, process the data according to steps 1 and 2, then input it into the trained model and calculate the state index.

[0035] Step 5.1: State Indicator Smoothing. The state indicators of the test data are smoothed using an exponentially weighted moving average. For the original test data state indicators... y i Its smoothing value is ,in λ The weighting coefficient is set to 0.2. i ={1,2, …, 12394} After smoothing, an average value is taken for each hour. Figure 1 It shows how the smoothed state index changes over time.

[0036] Step 5.2: Fault Alarm. The processed status indicators from Step 5.1 are compared with the health threshold of 12.969. An alarm is issued when 24 consecutive status indicators exceed the threshold. Figure 1 The algorithm of this invention issued an alarm on January 10, 2019, while the fault was discovered manually on May 29, 2019. The algorithm detected the pitch system fault at least 139 days in advance.

[0037] Step 6: Fault Location. After the alarm was issued on January 10, 2019, data from January 10, 2019 to May 29, 2019 was collected. The vector formed by connecting the centers of the two-dimensional features of each type of parameter in the sample with the corresponding two-dimensional feature centers of the normal data was obtained. The cosine similarity of this vector with respect to the coordinate axis was calculated, and its absolute value was taken. The specific calculation formula is as follows: ,in k Indicates the parameter type; j Indicates the index of the coordinate axis. j={1, 2, 3}; The first in the normal dataset k The center of the parameter characteristics; For the anomaly point k Two-dimensional feature coordinates of the parameter; representing the first parameter. j The unit vectors of the coordinate axes, i.e. , , .

[0038] Step 7: Calculate the sum of cosine similarities for each coordinate axis. Adjust the fault assessment value of the blade corresponding to the highest cosine similarity to above the warning value. ,in Indicates the first j Failure index of each blade K The total parameter type is 4; m The number of points exceeding the limit; The largest corresponding j The faulty blade. Calculations were performed. , , , The value is the largest, therefore it is determined that blade number 3 has failed, consistent with the maintenance record. From Figure 2 As can be seen in the abnormal data, the pitch drive temperature is located in the lower right corner of the normal data, approximately 30 degrees to the horizontal axis of the Cartesian coordinate system, which is consistent with the angle assigned by the feature projection (when performing feature projection, the coordinates of the No. 1 blade parameter are aligned with the vertical axis in the Cartesian coordinate system).

[0039] The above description is merely a specific embodiment of this application. Under the guidance of the above teachings, those skilled in the art can make other improvements or modifications based on the above embodiments. Those skilled in the art should understand that the above specific description is only to better explain the purpose of this application, and the scope of protection of this application should be determined by the scope of the claims.

[0040] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.

Claims

1. A method for adjusting the fault assessment value of a wind turbine pitch system, characterized in that... Includes the following steps: Step 1: Obtain vibration signals during normal operation of the wind turbine. The vibration signals are derived from SCADA data. Step 2: Extract the common parameters of the three blades in the pitch system to obtain the transformed two-dimensional features; Step 3: Construct a single-class support vector machine model, using the Gaussian kernel function as the kernel function of the model, and train the model using the two-dimensional features obtained in Step 2 as input data; Step 4: Substitute normal data into the trained model to calculate its state index value under normal conditions, and then calculate the threshold using the 3σ rule; specifically: first, substitute normal data into the trained model to calculate its state index value under normal conditions, and then calculate the threshold. The calculation formula is thre = μ + 3σ, where... y i This indicates a normal data status indicator, where n represents the data length. Step 5: Calculate the health threshold using status indicators from normal data; Steps 5.1 and 5.2 are included after Step 5: Step 5.1: During testing, the test data is processed according to steps 1 and 2, then fed into the trained single-class support vector machine model to calculate the state index. The state index is smoothed using the exponentially weighted moving average method. Specifically, for the original state index y... i Its smoothing value z i =(1-λ)z i-1 +λy i , where λ is the weighting coefficient, i>1, z1=μ, μ is the mean of the status index of the normal data obtained from step 4; after smoothing, the mean is obtained by aggregating by hour; Step 5.2: Compare the smoothed and averaged status indicators with the health threshold. When the status indicators of the test data exceed the threshold for 24 consecutive points, an alarm is issued. In Step 5.2, the Euclidean distance between each feature point after projection transformation and the decision boundary obtained by model training is used as the status indicator. Step 6: Obtain all samples that exceed the threshold, and calculate the cosine similarity between the two-dimensional data of each sample after projection transformation and the corresponding two-dimensional data center of the normal sample set relative to the coordinate axis. Step 7: Adjust the blade fault assessment value based on the sum of cosine similarity.

2. The method for adjusting the fault assessment value of the wind turbine pitch system according to claim 1, characterized in that: The normal operation data in step 1 does not include the following data: the fault, shutdown, and non-working data.

3. The method for adjusting the fault assessment value of the wind turbine pitch system according to claim 1, characterized in that... The projection transformation mentioned in step 6 refers to: for the common parameters of the three blades in the pitch system, each type of parameter is plotted on three plane coordinate axes that are 120 degrees apart and share the same origin, and the centroid of the triangle formed by the three points is projected onto a plane rectangular coordinate system with the same scale and origin to obtain the transformed two-dimensional feature.

4. The method for adjusting the fault assessment value of the wind turbine pitch system according to claim 1, characterized in that: The single classifier model constructed in step 3 is a single-class support vector machine, which uses the Gaussian kernel function as the kernel function of the model and uses all the feature values ​​after projection transformation as input data to train the model.

5. The method for adjusting the fault assessment value of the wind turbine pitch system according to claim 1, characterized in that: When issuing an early warning in step 5.2, an alarm is issued when 24 consecutive status indicators exceed the threshold.

6. The method for adjusting the fault assessment value of the wind turbine pitch system according to claim 1, characterized in that... In step 7, after calculating the cosine similarity of all parameters relative to the three coordinate axes, the sum of the cosine similarity under each coordinate axis is calculated. The fault assessment value of the blade is adjusted according to the sum of the cosine similarity. The fault assessment value of the blade with the highest cosine similarity is adjusted to above the warning value.

Images