Fault diagnosis method for wind turbine gearbox based on fuzzy mutual k-nearest neighbor

Abstract

The application discloses a wind driven generator gearbox fault diagnosis method based on fuzzy mutual K near neighbor, and comprises the following steps: 1, collecting parallel structure gearbox data, extracting key information in vibration data as a training set through time and frequency domain method; collecting historical wind driven generator operation monitoring data, screening relevant monitoring variables as a training set; 2, using mutual near neighbor rule and mean-value-median judgment rule to screen training set data, and preliminarily eliminating abnormal points; 3, fuzzy training set sample through membership function; 4, collecting online data, finding mutual near neighbor of sample in the training set, and calculating online sample membership degree by using membership degree of training set sample and improved membership function; 5, judging the running state of the gearbox at the current moment. The application reduces the interference caused by abnormal value in the gearbox fault diagnosis process by eliminating abnormal value, and obtains more accurate detection result than traditional machine learning method.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine gearbox fault diagnosis, and specifically relates to a wind turbine gearbox fault diagnosis method based on fuzzy mutual K-nearest neighbors. Background Technology

[0002] With the rapid development of technology, new power generation methods, primarily wind power, have seen rapid growth. As a crucial component of wind turbine generators, the gearbox's kinetic energy transmission mechanism is vital; any malfunction can disrupt the turbine's normal operation, potentially leading to burnout, economic losses, and even personal injury. Therefore, ensuring gearbox proper functioning and performing fault diagnosis are crucial for the turbine's smooth operation. K-Nearest Neighbors (KNN), a machine learning algorithm, has not only successfully handled classification problems but has also been successfully applied to fault diagnosis. However, when diagnosing wind turbine gearbox faults, the harsh operating environment and potential interference during data transmission result in noise and outliers in the collected data, leading to errors in fault diagnosis. Compared to other complex machine learning algorithms, KNN is relatively simple to implement and performs well in fault diagnosis. However, KNN suffers from inaccurate classification for samples with unclear class boundaries; furthermore, outliers significantly impact KNN, potentially leading to the identification of incorrect neighbors and affecting the wind turbine gearbox fault diagnosis results. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing a wind turbine gearbox fault diagnosis method based on fuzzy mutual K-nearest neighbors, so as to improve the accuracy of wind turbine gearbox fault diagnosis.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A method for fault diagnosis of wind turbine gearboxes based on fuzzy mutual K-nearest neighbor (FMKNN) includes the following steps:

[0006] Step 1: Offline training. Collect data on the normal and various fault operation states of parallel structure gearboxes, as well as data on the normal and various fault operation states of planetary structure gearboxes of wind turbine generators, and filter the data. Use the nearest neighbor method to remove individual outliers in the training set that are far from the data center. Then, use the mean-median method to treat samples that exceed the threshold range as outliers that have a nearest neighbor relationship with each other and remove them. Finally, use the fuzzy membership function to fuzzify the samples.

[0007] Step two: in the online monitoring stage, the membership of the online collected samples is calculated by using the training set processed in the offline stage and the membership vector of the training samples, and the running state of the current gearbox is judged according to the membership vector, thereby providing support for later maintenance.

[0008] The training data include two parts, one part is parallel structure gearbox data, including normal gear data, gear tooth breakage fault data and gear wear fault data under different rotating speeds; the other part is data collected by a SCADA (Supervisory Control and Data Acquisition) monitoring system during the operation of the fan planetary structure gearbox, including normal gearbox operation data, gearbox filter core blockage fault data, gearbox low pressure fault data and gearbox temperature control valve failure fault data.

[0009] The fuzzy mutual K nearest neighbor method includes a mutual nearest neighbor method and fuzzification based on the mutual nearest neighbor.

[0010] The method, the mutual nearest neighbor rule is used to eliminate abnormal values in the training set; if a sample x is given, x i is a nearest neighbor in x, and at this time x i is also a nearest neighbor in x, then x and x i are mutual nearest neighbors, denoted as M k (x) = {x i}; if the above conditions are not met, it is considered that x does not have mutual nearest neighbors, denoted as M k (x) = {Φ}, and the formula is:

[0011] M k (x) = {x i ∈T|x i ∈S k (x) ∧ x ∈ S k (x i )}

[0012] Wherein S k (x) represents the set of k nearest neighbors of the sample x, S k (x i ) represents the set of k nearest neighbors of the sample x i , the above formula is applied to find all samples in the training set T that do not have mutual nearest neighbors, and the samples are removed to return

[0013] The mean-median method is:

[0014] Firstly, the Euclidean distance between the ith sample and other samples under the same condition is calculated, the mean value of the distance is calculated, and the median value is set as the threshold value; then the difference between the mean value and the median value is calculated, and it is judged whether it exceeds the threshold line, if it exceeds, the sample i is regarded as an abnormal value and is removed from the training set, and returns The calculation formula is as follows:

[0015]

[0016] Wherein, D i represents the mean value of the distance between the ith sample and other samples under the same condition, d i,j represents the Euclidean distance from sample i to the jth neighbor, and m represents the number of samples under the same condition as i.

[0017] The method, when the training set samples are fuzzed, uses the mutual neighbor-based fuzzy membership function calculation formula:

[0018]

[0019] Wherein, u c (x i ) is the membership of sample x i to class c, and the membership range is u c (x i )∈[0,1]; n j is the number of mutual neighbors of the training sample belonging to class c; is the base of x i , that is, the total number of mutual neighbors of x i .

[0020] The method, for the data collected in the online phase, uses the processed training set, the training set membership and the improved membership function to fuzz the online samples, obtains the fuzzy membership result, and judges the current gearbox running state according to the membership vector; the expression of the improved membership function is:

[0021]

[0022] Wherein, u(x j ) represents the membership of y to the mutual neighbor found in the training set; u(y) represents the membership vector of the online monitoring sample, and the membership of y is represented by the linear combination of the membership of its mutual neighbor; is the base of y, that is, the total number of mutual neighbors of y in the training set; is the weight of the mutual neighbor x j , M∈(1,+∞)(here M=2), ω j is used for normalization to ensure

[0023] The beneficial effects of this invention are:

[0024] This invention minimizes the impact of outliers during wind turbine gearbox inspection, thereby improving the accuracy of gearbox fault detection. Specifically, before training the data, some outliers in the training samples are removed, reducing their influence on fault diagnosis. In the online detection phase, compared to K-nearest neighbors, this invention utilizes more reliable mutual nearest neighbor information, enabling more clear differentiation of similar faults and further improving fault diagnosis performance. Attached Figure Description

[0025] Figure 1 This is a flowchart of the present invention.

[0026] Figure 2 The data source for the parallel shaft structure gearbox includes three parts: (a) equipment connection diagram, (b) QPZZ-Ⅱ rotating machinery vibration fault test platform, and (c) signal testing and analysis system and supporting software.

[0027] Figure 3 It is a graph showing the average distance statistics between a sample and other samples of the same type when a certain proportion of outliers are included.

[0028] Figure 4 This is a screenshot showing the result of outlier removal based on the nearest neighbor rule. Specific implementation methods

[0029] The present invention will be further described below with reference to the accompanying drawings and examples, so that those skilled in the art can implement it based on the description.

[0030] like Figure 1 As shown, the wind turbine gearbox fault diagnosis method based on fuzzy mutual K-nearest neighbors includes the following steps:

[0031] The first step is to acquire data on the parallel gearbox and the planetary gearbox respectively, and collect data on normal operating conditions. and operating data under various fault conditions, As the training set, denoted as (p+1)m1 represents the total number of samples, and n1 represents the number of variables.

[0032] The second step is to normalize each sample in order to avoid the influence of individual large variables on the distance calculation.

[0033]

[0034] Where, x scaleTo normalize the range of the variable, x scale ∈ [0, 1]; x min represents the minimum value in the variable, x max represents the maximum value in the variable, and the return set

[0035] The third step is to determine the value of the parameter k. Different values of k have a great impact on the fault diagnosis results. When the value of k is small, overfitting problems are likely to occur. When the value of k is too large, experimental verification shows that the classification accuracy of the algorithm decreases with the increase of the value of k. By changing the value of k, the value of k with the best performance is selected. Then, the abnormal points in the training set T1 are removed using the mutual nearest neighbor rule, and the return set The method for preliminary removal of abnormal values is as follows: given a sample x, if x i is a neighbor of x, and at this time x i is also a neighbor of x, then x and x i are mutual nearest neighbors, denoted as M k (x) = {x i}; if the above conditions are not met, then x does not have mutual neighbors, denoted as M k (x) = {Φ}, the formula is:

[0036] M k (x) = {x i ∈ T | x i ∈ S k (x) ∧ x ∈ S k (x i )}

[0037] where S k (x) represents the set of k neighbors of sample x, S k (x i ) represents the set of k neighbors of sample x i , and the above formula is used to find all samples in the training set T that do not have mutual neighbors, and the return set

[0038] The fourth step is to calculate the distance between each sample x i in T2 and other samples x b under the same working condition, and the calculation formula is:

[0039]

[0040] where d(y, x i) represents the Euclidean mean between the sample and other samples in the same class, and m represents the number of samples in the same working condition. Since the abnormal points are usually distributed far from the center of the normal sample distribution, the mean and median of the Euclidean distance between the ith sample and other samples in the same working condition can be calculated, and the median is set as the threshold. Then, the difference between the mean and the median is calculated to determine whether it exceeds the threshold line. If it exceeds, the sample i is regarded as an abnormal value and is removed from the training set, and the process returns to step 1. The calculation formula is as follows:

[0041]

[0042] where D i represents the mean distance between the ith sample and other samples in the same working condition, and d i,j represents the Euclidean distance between sample i and the jth nearest neighbor, and m represents the number of samples in the same working condition as i.

[0043] Step 5, for the samples x i in T3, find the mutual nearest neighbors using the MKNN rule, introduce the membership function, and fuzzy the samples to obtain the fuzzy membership of x i . The fuzzy membership calculation formula is as follows.

[0044]

[0045] where u c (x i ) is the membership of sample x i to class c, and the membership range is u c (x i ) ∈ [0, 1]; n j is the number of mutual nearest neighbors of the training samples belonging to class c; is the cardinality of , that is, the total number of mutual nearest neighbors of x i .

[0046] Step 6, for the newly acquired data in the online phase, denoted as , find its mutual nearest neighbors in the set T3. If the sample does not have mutual nearest neighbors, find the k nearest neighbors of the sample. According to the mutual nearest neighbors or the nearest neighbors found, use the improved membership function to assign the membership of y, and the calculation formula is as follows.

[0047]

[0048] where u(y) represents the membership of y to each class, and the membership range is u(y) ∈ [0, 1]; the membership of y is represented by the linear combination of the membership of its mutual nearest neighbors; and the membership of x j) represents the membership of y found in the training set of mutual neighbors; is the weight of mutual neighbor x j M∈(1, +∞) (here M = 2) is used for normalization to ensure

[0049] Step 7, the label of the class to which the sample y belongs is determined, and the calculation formula is as follows.

[0050]

[0051] The class with the largest membership is taken as the class of the sample y, so as to judge the current running state of the gearbox and perform fault classification. Embodiment

[0052] The effectiveness of the method will be described below in combination with parallel structure gearbox data and planetary structure gearbox data.

[0053] Firstly, the parallel structure gearbox data is shown. The QPZZ-II rotating machinery vibration fault experiment platform is used to carry out the experiment (as shown in Figure 2 , wherein the connection between each device in the experiment process is shown in Figure 2 (b part). The vibration signal is collected at a sampling frequency of 2 kHz, and the normal gear signal at 1500 r / min, the broken gear signal and the worn gear signal under the condition of 1500 r / min and 1200 r / min are collected by a piezoelectric sensor. Each kind of state collects 150 signals, and the sampling interval time is 5 s. The characteristics of each signal are extracted and taken as a sample, and the data set T1 contains 750 samples. 30 samples are selected as test samples, and the rest are taken as training samples. The samples with a selection ratio of 4% are superimposed with SNR = -5 dB Gaussian white noise, which replaces the original sample as an abnormal point introduced into the training set.

[0054] Next, the implementation steps of the present application will be described in more detail in combination with the specific process:

[0055] 1. Firstly, the value of the parameter k is determined. Different k values will have a great influence on the fault diagnosis result. When the k value is small, overfitting problem is easy to occur. When the k value is too large, it is verified through experiments that the classification accuracy of the algorithm will decrease with the increase of the k value. After many experiments, the k value with the optimal performance is 6. Then, the mutual neighbor rule is used to remove part of the outliers in T1, and T2 is returned.

[0056] 2、Calculate the distance between each sample in T2 and other samples in the same working condition, and then use the mean-median method to identify and remove abnormal points that are near neighbors of each other, returning T3. Then, the processed training set is fuzzified using the fuzzy membership function. Find the mutual neighbors of the test sample in the training set T3, use the membership of the training set and the improved membership function to get the membership of the test sample, and diagnose the gearbox fault by the class to which the maximum membership belongs.

[0057] First, detect the effect of the fuzzy mutual K-nearest neighbor algorithm on the removal of abnormal values. Take 4% of all samples with added Gaussian white noise as abnormal values. To facilitate observation, add abnormal values to the last sample of each class of training set samples, and use green boxes to circle the red abnormal value samples. Figure 3 The horizontal axis represents the sample, and the vertical axis represents the mean value of the distance between the sample and other samples under the same working condition. Figure 4 The sample distance mean value distribution graph after using the mutual neighbor method and the mean-median method to remove abnormal points. From the graph, it can be clearly observed that abnormal values can be completely removed.

[0058] Tables 1, 2, 3, and 4 respectively show the comparison of the FMKNN method and other machine learning methods when different proportions of abnormal points are added. From the tables, it can be clearly observed that the FMKNN algorithm has better fault classification performance for gearbox data with different proportions of abnormal points. With the increase of the proportion of abnormal points, the performance decreases, but compared with other methods, this method can still maintain high performance.

[0059] Table 1 Algorithm classification accuracy (%) of different proportions of abnormal points when SNR = -5dB

[0060]

[0061] Table 2 Algorithm classification accuracy (%) of different proportions of abnormal points when SNR = -7dB

[0062]

[0063] Table 3 Algorithm classification accuracy (%) of different proportions of abnormal points when SNR = -5dB

[0064]

[0065] Table 4 Algorithm classification accuracy (%) of different proportions of abnormal points when SNR = -7dB

[0066]

[0067] Secondly, the planetary structure gearbox data is shown, and the experimental data is collected by SCADA monitoring system in the running process of a plurality of 2MW wind turbines in a wind farm. Data is collected every 5 minutes, including data collected under four conditions of normal gearbox, blocked filter core of gearbox gear box, low pressure of gearbox and failure of gearbox temperature control valve. Each gearbox state takes 150 consecutive samples, and the data set contains 600 samples, 30% of which are taken as test set and the remaining samples are taken as training set. The variables monitored by the SCADA system include 63 types of data such as wind speed, current, yaw error, bearing temperature, blade position, grid power and torque, and 13 types of data are selected for research, including generator speed, wind speed, gearbox oil temperature and active power.

[0068] Table 5 and Table 6 show the accuracy of the FMKNN method and other machine learning methods in fault classification when different proportions of abnormal points are added in the wind turbine gearbox. From the table, it can be seen that the FMKNN method has better performance than other machine learning algorithms in the fault classification of planetary structure gearbox, and the fault classification accuracy is higher than that of other machine learning methods. From the tables, it can be seen that the FMKNN method not only has good fault classification performance in parallel structure gearbox, but also maintains high classification accuracy in fault classification of planetary structure gearbox.

[0069] Table 5 Fault classification accuracy of wind turbine gearbox when different proportions of abnormal points are added (%) Table 6 Fault classification accuracy of wind turbine gearbox with abnormal points (%)

[0070]

[0071] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for fault diagnosis of wind turbine gearboxes based on fuzzy mutual K-nearest neighbors, characterized in that, Includes the following steps: Step 1: Offline training. Collect data on the normal and various fault operation states of parallel structure gearboxes, as well as data on the normal and various fault operation states of planetary structure gearboxes of wind turbine generators, and filter the data. Use the nearest neighbor method to remove individual outliers in the training set that are far from the data center. Then, use the mean-median method to treat samples that exceed the threshold range as outliers that have a nearest neighbor relationship with each other and remove them. Finally, use the fuzzy membership function to fuzzify the samples. Step 2: In the online monitoring phase, the membership degree of the online collected samples is calculated using the training set processed in the offline phase and the membership degree vector of the training samples. Based on the membership degree vector, the current operating state of the gearbox is determined, providing support for later maintenance. The training data for offline training consists of two parts. One part is parallel structure gearbox data, including normal gear data, as well as gear tooth breakage fault data and gear wear fault data at different speeds. The other part is data collected by the SCADA (Supervisory Control and Data Acquisition) monitoring system during the operation of the wind turbine planetary structure gearbox, including normal gearbox operation data, gearbox filter element blockage fault data, gearbox low pressure fault data, and gearbox temperature control valve failure fault data. The mean-median method is as follows: First, calculate the Euclidean distance between the i-th sample and other samples under the same working conditions, then calculate the mean and median of these distances, and set the median as the threshold. Next, calculate the difference between the mean and the median to determine if the difference exceeds the threshold. If it does, remove the sample from the threshold. Values ​​considered outliers are removed from the training set and returned. The calculation formula is as follows: ； Among them, D i Let d represent the mean distance between the i-th sample and other samples under the same working conditions. i,j Let represent the Euclidean distance from sample i to its j-th nearest neighbor, and m represent the number of samples with the same working conditions as i.

2. The method according to claim 1, characterized in that, Fuzzy mutual K-nearest neighbor methods include mutual nearest neighbor methods and mutual nearest neighbor-based fuzzification.

3. The method according to claim 1, characterized in that, The nearest neighbor rule is used to remove outliers from the training set; given a sample ,if As One of the nearest neighbors, and at this time Too One of the close neighbors is called and Mutual nearest neighbors, represented as If the above conditions are not met, then it is called No mutual nearest neighbors are represented as The formula is: ； in Represented as a sample The set of k nearest neighbors, Represented as a sample Given the set of k nearest neighbors, use the above formula to find the training set. All samples that do not have nearest neighbors are removed and returned. .

4. The method according to claim 1, characterized in that, When fuzzifying the training set samples, the formula for calculating the fuzzy membership function based on mutual nearest neighbors is as follows: ； in, For the sample China belongs to Membership degree of a class, the range of membership degree is ; It belongs to the nearest neighbors of the training samples The number of classes; yes The cardinality, that is The total number of mutual nearest neighbors.

5. The method according to claim 1, characterized in that, For the data collected during the online phase, the online samples are fuzzified using the processed training set, training set membership degrees, and the improved membership function to obtain the fuzzified membership results. The current gearbox operating status is then determined based on the membership vector. The expression for the improved membership function is as follows: ； in, express Membership degree of mutual nearest neighbors found in the training set; The membership vector representing the online monitoring samples. The membership degree of a is represented by a linear combination of the membership degrees of its nearest neighbors; yes The cardinality, that is The total number of mutual nearest neighbors in the training set; as neighbors The weight, , Used for normalization to ensure .

Images