CN114237206A - Wind power variable pitch system fault detection method for complex operation conditions - Google Patents

Wind power variable pitch system fault detection method for complex operation conditions Download PDF

Info

Publication number
CN114237206A
CN114237206A CN202111554303.4A CN202111554303A CN114237206A CN 114237206 A CN114237206 A CN 114237206A CN 202111554303 A CN202111554303 A CN 202111554303A CN 114237206 A CN114237206 A CN 114237206A
Authority
CN
China
Prior art keywords
sample
samples
stage
distance
fault detection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111554303.4A
Other languages
Chinese (zh)
Inventor
钱小毅
孙天贺
叶鹏
王宝石
李天岳
杨宏宇
魏靖晓
张政斌
王子赫
邵旸棣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenyang Institute of Engineering
Original Assignee
Shenyang Institute of Engineering
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenyang Institute of Engineering filed Critical Shenyang Institute of Engineering
Priority to CN202111554303.4A priority Critical patent/CN114237206A/en
Publication of CN114237206A publication Critical patent/CN114237206A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
    • G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
    • G05B23/0224—Process history based detection method, e.g. whereby history implies the availability of large amounts of data
    • G05B23/024—Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
    • Y04S10/52—Outage or fault management, e.g. fault detection or location

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Wind Motors (AREA)

Abstract

The invention belongs to the technical field of fault detection of wind power generation assemblies, and particularly relates to a fault detection method for a wind power variable pitch system facing complex operation conditions. The method effectively avoids the intercrossing of complex operation conditions, better describes the operation state of specific conditions, and can effectively reduce the false alarm rate and the missing report rate in the fault detection process. The method comprises the following steps: step 1, cleaning SCADA data of a wind turbine generator; step 2, selecting the associated characteristics of a variable pitch system of the wind turbine generator; step 3, performing off-line modeling by using the normal behavior sample; step 4, determining an early warning threshold value through offline deviation distribution; step 5, carrying out online evaluation through the difference measurement of the similar states; and 6, realizing abnormal identification according to the relation between the online evaluation result and the threshold value.

Description

Wind power variable pitch system fault detection method for complex operation conditions
Technical Field
The invention belongs to the technical field of fault detection of wind power generation assemblies, and particularly relates to a fault detection method for a wind power variable pitch system facing complex operation conditions.
Background
The development of low-carbon economy, the development and utilization of clean renewable energy sources have become key points of human sustainable development and energy strategy. Wind power generation has become the fastest growing green energy globally as a new energy power generation mode with mature technology, large development scale and good commercial development. The method has the advantages that the frequent faults of key components of the wind turbine generator set cause huge economic loss, the fault diagnosis technology grasps the change of the running state of the wind turbine generator set by analyzing the real-time running parameters of the components of the wind turbine generator set, the fault hidden danger is found in time, major accidents are avoided, and the method is an effective way for reducing the operation and maintenance cost of the wind turbine generator set and improving the economic benefit.
The variable pitch system belongs to a fault high-risk component, and frequent shutdown caused by the variable pitch system brings great economic loss, and meanwhile, the fault of the variable pitch system directly causes uncontrollable load borne by a unit and even causes serious destructive accidents. The existing wind turbine generator system variable pitch system fault detection research realizes real-time monitoring of operation states through different methods, wherein the research aiming at complex operation working conditions of wind turbine equipment mainly focuses on dividing the operation working conditions and establishing a distributed anomaly detection model. The clustering-based method is more reasonable for distributed modeling, however, similar to the condition-based method, the selection and the selection of the contribution of each sub-model in the evaluation process of the distributed detection model can cause great influence on the detection result, not only delay the real-time performance of fault detection, but also be a key factor for restricting the further reduction of the false alarm rate and the false missing report rate. In addition, the complexity of local modeling based on working condition division depends on the sample size and the number of working conditions, and the determination of the number of the working conditions lacks theoretical basis. Therefore, the fault detection scheme of the wind turbine generator variable pitch system designed according to complex working conditions has practical application value.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a wind power pitch system fault detection method facing to complex operation conditions.
In order to achieve the purpose, the invention adopts the following technical scheme that the method comprises the following steps:
step 1, cleaning SCADA data of a wind turbine generator;
step 2, selecting the associated characteristics of a variable pitch system of the wind turbine generator;
step 3, performing off-line modeling by using the normal behavior sample;
step 4, determining an early warning threshold value through offline deviation distribution;
step 5, carrying out online evaluation through the difference measurement of the similar states;
and 6, realizing abnormal identification according to the relation between the online evaluation result and the threshold value.
Further, the step 1 is divided into a first stage and a second stage;
the first stage cleaning step is as follows:
step 1.1: computing offline samples xiNearest neighbor distance of
Figure BDA0003418157090000021
Wherein, yiIs xiN is the feature dimension, i is 1,2, …, N is the total number of offline samples, for DiIn ascending order and at the alpha1Neighbor distance of% × N samples
Figure BDA0003418157090000022
As a threshold for similar sample culling;
step 1.2: let i equal to 1, calculate the nearest neighbor distance D for the ith samplei；
Step 1.3: if it is
Figure BDA0003418157090000023
Then the sample is deleted and N-1, otherwise the sample is retained;
step 1.4: i-i +1, and its nearest neighbor distance D is calculated among the remaining N samplesi+1If, if
Figure BDA0003418157090000024
Deleting the sample N-1, otherwise, retaining the sample;
step 1.5: repeating step 1.4 until all remaining samples are traversed;
in the cleaning process of the first stage, the average value of the nearest neighbor distances of all samples is defined as the average crowding degree, and is used as a cleaning evaluation index of the stage:
Figure BDA0003418157090000025
the more the average crowding, the more evenly the sample distribution, with fewer highly similar samples.
Further, the second stage cleaning step is as follows:
calculate all samples xiK nearest neighbor distance of
Figure BDA0003418157090000031
Wherein, yi，jIs a sample xiN is the feature dimension, i is 1,2, …, N is the total number of current samples, for all samples
Figure BDA0003418157090000038
In descending order and in alpha2Neighbor distance of% × N samples
Figure BDA0003418157090000032
As a threshold for similar sample culling; k nearest neighbor distance D for all samplesi kIf, if
Figure BDA0003418157090000033
Rejecting the sample, otherwise retaining the sample;
in the cleaning process of the second stage, defining the mean value of k neighbor distances of all samples as the average dispersion, and taking the mean value as the cleaning evaluation index of the stage
Figure BDA0003418157090000034
Further, in step 3, the off-line model building stage is:
step 3.1: searching k neighbor samples for all samples in the training set;
step 3.2: calculating kNN distance of each sample, and kNN distance D of each sample ii 2Defined as the sum of the distance of sample i to its k neighbors:
Figure BDA0003418157090000035
wherein the content of the first and second substances,
Figure BDA0003418157090000036
the Euclidean distance from the sample i to the jth neighbor thereof;
step 3.3: an abnormal state discrimination threshold is determined.
Further, in step 4, if the k-nearest neighbor distance d between the training samples is assumedijObey a normal distribution with a non-zero mean value, and dijThe square sum is randomly selected to obtain the non-central x2Distributing; since the calculation of the kNN distance is not a random process, the distribution can only be approximated as a non-central χ2Distribution, using the Chilimit function in Matlab in PLS toolset to estimate the threshold D of confidence αa 2。
Further, in step 5, the online status detection stage specifically includes the steps of:
step 5.1: for the online input sample x, its k neighbor samples are found from the training dataset.
Step 5.2: computing kNN distance of x
Figure BDA0003418157090000037
Compared with the prior art, the invention has the beneficial effects.
Firstly, a two-stage cleaning scheme based on neighbor distance is adopted, and in the first stage, samples with higher similarity are evaluated and removed by adopting single-sample circulating nearest neighbor distance; and in the second stage, abnormal data and noise data are removed by adopting a k-nearest neighbor whole sample evaluation-based method, so that the cleaning of the running state data of the wind turbine generator is realized.
According to the method for detecting the fault of the variable pitch system based on the similarity state difference, each step of analysis only depends on the distance relation among a plurality of similarity states, the intercrossing of complex operation conditions is effectively avoided, the operation state of specific conditions is better described, and the false alarm rate of the fault detection process can be effectively reduced.
Drawings
The invention is further described with reference to the following figures and detailed description. The scope of the invention is not limited to the following expressions.
FIG. 1 is a diagram of an overall data cleaning scheme of a pitch system of a wind turbine generator.
FIG. 2 is a pitch system fault detection flow diagram based on approximate state similarity.
FIG. 3 is a wind speed-power relationship graph of SCADA data.
FIG. 4(a) shows a difference of α1Lower average congestion contrast map.
FIG. 4(b) shows a difference of α2Lower dispersion contrast plot.
FIG. 5 is a wind turbine SCADA data cleaning process.
Detailed Description
As shown in fig. 1 to 5, the invention relates to a wind power pitch system fault detection method facing complex operation conditions, which comprises the following steps:
(1) and cleaning SCADA data of the wind turbine generator.
(2) And selecting the associated characteristics of the variable pitch system of the wind turbine generator.
(3) And carrying out off-line modeling by using the normal behavior sample.
(4) And determining an early warning threshold value through offline deviation distribution.
(5) And carrying out online evaluation through the similarity state difference metric.
(6) And realizing the abnormal recognition according to the relation between the online evaluation result and the threshold value.
In the step (1), the specific cleaning steps are as follows:
the first stage is as follows: similar samples were washed.
In the fault diagnosis process, highly overlapped data can directly influence the judgment of the state, and in order to reduce the influence of repeated data or highly similar data in the subsequent fault detection modeling process, a single sample circulation similar sample cleaning method based on nearest neighbor distance is provided, and the method specifically comprises the following steps:
step 1: computing offline samples xiNearest neighbor distance of
Figure BDA0003418157090000051
Wherein, yiIs xiN is the feature dimension, i is 1,2, …, N is the total number of offline samples, for DiIn ascending order and at the alpha1Neighbor distance of% × N samples
Figure BDA0003418157090000052
As a threshold for similar sample culling.
Step 2: let i equal to 1, calculate the nearest neighbor distance D for the ith samplei。
And step 3: if it is
Figure BDA0003418157090000053
The sample is deleted and N-1, otherwise the sample is retained.
And 4, step 4: i-i +1, and its nearest neighbor distance D is calculated among the remaining N samplesi+1If, if
Figure BDA0003418157090000054
The sample N-1 is deleted, otherwise the sample is retained.
And 5: repeat step 4 until all remaining samples are traversed.
The method is implemented by matching a threshold value alpha1The setting of the method can adjust the proportion of rejecting similar samples, and the single-sample circulating mode can avoid the condition that all samples in local similar states are simultaneously rejected to cause sample loss. Due to the adoption of the single-sample circulation strategy, the nearest neighbor distance of the subsequent samples changes along with the reduction of the sample amount, so that the proportion of the actually-rejected data is smaller than alpha1。
In the cleaning process of the first stage, the average value of the nearest neighbor distances of all samples is defined as the average crowding degree, and is used as a cleaning evaluation index of the stage:
Figure BDA0003418157090000055
it is clear that the larger the average crowding, the more evenly the sample distribution, with fewer highly similar samples.
And a second stage: and (4) cleaning an isolated sample.
When the normal operation data sufficiently cover all the operation conditions, the isolated points in the wind speed-power curve graph may be noise points or abnormal points, so that the isolated points need to be removed to make the wind speed-power curve graph more approximate to an ideal state. This section adopts whole sample k to near the means of distance measurement and rejects the relative isolated point in the original data, and the concrete step is as follows:
calculate all samples xiK nearest neighbor distance of
Figure BDA0003418157090000061
Wherein, yi，jIs a sample xiN is the feature dimension, i is 1,2, …, N is the total number of current samples, for all samples
Figure BDA0003418157090000062
In descending order and in alpha2Neighbor distance of% × N samples
Figure BDA0003418157090000063
As a threshold for similar sample culling. K nearest neighbor distance D for all samplesi kIf, if
Figure BDA0003418157090000064
The sample is rejected, otherwise the sample is retained.
In the cleaning process of the second stage, defining the mean value of k neighbor distances of all samples as the average dispersion, and taking the mean value as the cleaning evaluation index of the stage
Figure BDA0003418157090000065
A flow chart for two-stage data cleansing is shown in fig. 1.
When the Relieff algorithm is adopted to select the state parameters of the wind turbine generator in the step (2), firstly, irrelevant state parameters are manually removed, a rough relevant state parameter range is determined, and the total number of the parameters is p; the marked historical operation data is used as training data of a Relieff algorithm; the status type (normal operation or fault type) of each historical sample serves as a classification label for the Relieff algorithm. The state parameter sequencing and the corresponding characteristic weight of the wind turbine generator running state distinguishing capacity can be obtained by setting the proper iteration times and the nearest neighbor sample number.
In the step (3), the off-line model building stage specifically comprises the following steps:
step 1: for all samples in the training set, find their k neighbor samples.
Step 2: calculating kNN distance of each sample, and kNN distance of each sample i
Figure BDA0003418157090000071
Defined as the sum of the distance of a sample i to its k neighbors:
Figure BDA0003418157090000072
wherein the content of the first and second substances,
Figure BDA0003418157090000073
for sample i to its jth nearestThe euclidean distance of the neighbors.
And step 3: an abnormal state discrimination threshold is determined.
In step (4), if the k-nearest neighbor distance d between training samples is assumedijObey a normal distribution with a non-zero mean value, and dijThe square sum is randomly selected to obtain the non-central x2And (4) distribution. Since the calculation of the kNN distance is not a random process, the distribution can only be approximated as a non-central χ2Distribution, therefore, the threshold D of confidence α can be estimated using the ceiling function in Matlab in the PLS toolboxα 2。
In the step (5), the online state detection stage specifically comprises the following steps:
step 1: for the online input sample x, its k neighbor samples are found from the training dataset.
Step 2: computing kNN distance of x
Figure BDA0003418157090000074
In step (6), comparison
Figure BDA0003418157090000077
And a threshold value Dα 2The magnitude relationship of (1). If it is
Figure BDA0003418157090000075
The state is a normal operation state, otherwise, the state is a fault state. Fig. 2 is a flow chart of a similar status based differential fault detection.
The specific embodiment is as follows:
the implementation process of the invention is illustrated by taking common over-temperature faults of a variable pitch motor, abnormal faults of a variable pitch angle and asymmetric faults of the variable pitch angle of a variable pitch system of a wind turbine generator as examples. The fault description is shown in table 1. The SCADA experimental data are collected from a large-continuous-hump wind power plant, a direct-drive wind turbine generator set with the model of golden wind GW77-1.5MW collects 25000 sets of operation data from 2 months to 3 months in 2014 at intervals of 5 minutes. And drawing a wind speed-power scatter plot according to partial SCADA data, as shown in FIG. 3.
TABLE 1 Pitch System Fault description
Figure BDA0003418157090000076
Figure BDA0003418157090000081
The 15 monitoring variables related to the pitch system were selected in the SCADA system for fault detection as shown in table 2. The wind speed, the power and the rotating speed of the generator are key variables for measuring the running state of the wind turbine generator, the current and the temperature of the variable pitch motor are key state variables for monitoring the over-temperature fault of the variable pitch motor, and the position of the encoder and the position of the redundant encoder are key state variables for monitoring the abnormal fault of the variable pitch angle and the asymmetric variable pitch angle.
TABLE 2 SCADA MONITORING VARIABLE FOR VARIABLE-VANE SYSTEM
Figure BDA0003418157090000082
Selecting a No. 7 unit in a great-hump wind field, collecting 25000 groups of running data between 2 months and 3 months in 2014 at intervals of 5 minutes, and removing 20380 groups of sample points which are smaller than cut-in wind speed points and low power points. Then, the two-stage data cleaning method provided by the chapter is adopted for the rest samples, and alpha is set1The average crowding distance of the whole sample is evaluated under each parameter set from 0.1 to 0.5 at intervals of 0.1, as shown in FIG. 4(a), and the comparison result in the figure shows that the average crowding distance is along with alpha1The value increases, the average congestion degree gradually increases, the effectiveness of the proposed method is verified, and alpha is set1Take 0.5 to guarantee the number of samples. For alpha1Taking 0.5 result to carry out second stage cleaning, setting alpha with 5 neighbor number2The average dispersion of the whole sample was evaluated at 0.1 intervals from 0.1 to 0.5, as shown in FIG. 4 (b).
When alpha is1And alpha2Are all taken for 0.5 hourThe cleaning process is shown in fig. 5, in which (a) in fig. 5 is the original data set, similar samples are removed in the first stage, and the result is shown in (b) in fig. 5; the second stage removes the isolated data, and the result is shown in fig. 5 (c); in fig. 5, (d) shows the change in the data amount at each stage.
According to the invention, the faults of the three variable pitch systems are detected, and the condition parameters of the wind turbine generator are considered to be abnormal before the SCADA system gives an alarm, so that the false alarm rate of the front 300 groups of samples and the missing report rate of the rear 200 groups of samples are only counted in the experiment. For the three pitch system faults shown in table 1, table 5 and table 6 are respectively the comparison of the false negative rate and the false positive rate of the fault detection method of the present invention and the PCA fault detection and KPCA fault detection methods.
TABLE 5 Pitch System Fault failure Rate comparison
Figure BDA0003418157090000091
TABLE 6 Pitch System Fault false alarm Rate comparison
Figure BDA0003418157090000092
As can be seen from tables 5 and 6, when the fault characteristics are detected, the detection results of the 3 faults are not ideal, and the false alarm rate are high because the PCA fault detection method has poor nonlinear processing capability. Compared with two methods based on principal component analysis, the method based on approximate state similarity can obtain lower false alarm rate and lower false alarm rate, because the fault detection mechanism of k neighbor can better process the data characteristics of the multi-working-condition operation mode of the wind turbine generator, and the method adopts a 'divide-and-conquer' approach and only relies on a plurality of similar samples to monitor the online samples.
The running state of a variable pitch system of the wind turbine generator is complex, the difference between different running working conditions is fuzzy, and the influence of complex running mutation is difficult to avoid by a traditional normal behavior model. The invention evaluates online samples by using the approximate state difference, determines the allowable threshold value of a normal state by the distribution of the neighbor distance of a historical sample in an offline stage, and evaluates the neighbor difference of each state by the distance of k similar states in an online stage. The simultaneous participation of large-scale data is avoided in both the off-line stage and the on-line stage, each analysis step only depends on the distance relation among a plurality of similar states, the mutual crossing of complex operation conditions is effectively avoided, the operation state of specific conditions is better described, and the false alarm rate in the fault detection process can be effectively reduced.
The operation conditions of a variable pitch system of the wind turbine generator set are frequently switched, and the real-time operation state is difficult to accurately evaluate. Therefore, the invention provides a wind power pitch system fault detection method based on similar state difference.
Firstly, a two-stage wind turbine generator data cleaning method based on neighbor distance measurement is provided: in the first stage, samples with high similarity are evaluated and removed by adopting single-sample circulation type nearest neighbor distance. And in the second stage, an overall sample evaluation method based on k-nearest neighbor distance measurement is adopted to eliminate abnormal data and noise data.
And then, establishing a fault detection model by using the similar operation state so as to solve the fault detection problem under the complex operation condition of the variable-pitch wind turbine generator. The degree of abnormality of the online sample only depends on the adjacent samples so as to deal with the precision influence of complex working conditions on the centralized modeling.
The method can more accurately depict different operation conditions of the wind power variable pitch system, and effectively reduce the false alarm rate and the missing report rate of fault detection.
It should be understood that the detailed description of the present invention is only for illustrating the present invention and is not limited to the technical solutions described in the embodiments of the present invention, and those skilled in the art should understand that the present invention can be modified or substituted equally to achieve the same technical effects; as long as the use requirements are met, the invention is within the protection scope.

Claims (6)

1. A wind power variable pitch system fault detection method for complex operation conditions is characterized by comprising the following steps: the method comprises the following steps:
step 1, cleaning SCADA data of a wind turbine generator;
step 2, selecting the associated characteristics of a variable pitch system of the wind turbine generator;
step 3, performing off-line modeling by using the normal behavior sample;
step 4, determining an early warning threshold value through offline deviation distribution;
step 5, carrying out online evaluation through the difference measurement of the similar states;
and 6, realizing abnormal identification according to the relation between the online evaluation result and the threshold value.
2. The wind power pitch system fault detection method for complex operating conditions according to claim 1, characterized by comprising: the step 1 is divided into a first stage and a second stage;
the first stage cleaning step is as follows:
step 1.1: computing offline samples xiNearest neighbor distance D ofi：
Figure FDA0003418157080000011
Wherein, yiIs xiN is the feature dimension, i is 1,2, …, N is the total number of offline samples, for DiIn ascending order and at the alpha1Neighbor distance of% × N samples
Figure FDA0003418157080000012
As a threshold for similar sample culling;
step 1.2: let i equal to 1, calculate the nearest neighbor distance D for the ith samplei；
Step 1.3: if it is
Figure FDA0003418157080000013
Then the sample is deleted and N-1, otherwise the sample is retained;
step 1.4: i ═ i +1, inCalculating the nearest neighbor distance D of the rest N samplesi+1If, if
Figure FDA0003418157080000014
Deleting the sample N-1, otherwise, retaining the sample;
step 1.5: repeating step 1.4 until all remaining samples are traversed;
in the cleaning process of the first stage, the average value of the nearest neighbor distances of all samples is defined as the average crowding degree, and is used as a cleaning evaluation index of the stage:
Figure FDA0003418157080000015
the more the average crowding, the more evenly the sample distribution, with fewer highly similar samples.
3. The wind power pitch system fault detection method for complex operating conditions according to claim 1, characterized by comprising: the second stage cleaning step is:
calculate all samples xiK nearest neighbor distance Di k：
Figure FDA0003418157080000021
Wherein, yi,jIs a sample xiN is the feature dimension, i is 1,2, …, N is the total number of current samples, for all samples
Figure FDA0003418157080000022
In descending order and in alpha2Neighbor distance of% × N samples
Figure FDA0003418157080000023
As a threshold for similar sample culling; k nearest neighbor distance D for all samplesi kIf, if
Figure FDA0003418157080000024
Rejecting the sample, otherwise retaining the sample;
in the cleaning process of the second stage, defining the mean value of k neighbor distances of all samples as the average dispersion, and taking the mean value as the cleaning evaluation index of the stage
Figure FDA0003418157080000025
4. The wind power pitch system fault detection method for complex operating conditions according to claim 1, characterized by comprising: in step 3, the off-line model establishing stage is as follows:
step 3.1: searching k neighbor samples for all samples in the training set;
step 3.2: calculating kNN distance of each sample, and kNN distance D of each sample ii 2Defined as the sum of the distance of sample i to its k neighbors:
Figure FDA0003418157080000026
wherein the content of the first and second substances,
Figure FDA0003418157080000027
the Euclidean distance from the sample i to the jth neighbor thereof;
step 3.3: an abnormal state discrimination threshold is determined.
5. The wind power pitch system fault detection method for complex operating conditions according to claim 1, characterized by comprising: in step 4, if the k-nearest neighbor distance d between training samples is assumedijObey a normal distribution with a non-zero mean value, and dijThe square sum is randomly selected to obtain the non-central x2Distributing; since the calculation of the kNN distance is not a random process, the distribution can only be approximated as a non-central χ2Distribution, using the Chilimit function in Matlab in PLS toolbox to estimate the threshold D of confidence αα 2。
6. The wind power pitch system fault detection method for complex operating conditions according to claim 1, characterized by comprising: in step 5, the online status detection stage specifically comprises the following steps:
step 5.1: for the online input sample x, its k neighbor samples are found from the training dataset.
Step 5.2: computing the kNN distance D of xx 2。
CN202111554303.4A 2021-12-17 2021-12-17 Wind power variable pitch system fault detection method for complex operation conditions Pending CN114237206A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111554303.4A CN114237206A (en) 2021-12-17 2021-12-17 Wind power variable pitch system fault detection method for complex operation conditions

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111554303.4A CN114237206A (en) 2021-12-17 2021-12-17 Wind power variable pitch system fault detection method for complex operation conditions

Publications (1)

Publication Number Publication Date
CN114237206A true CN114237206A (en) 2022-03-25

Family

ID=80758351

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111554303.4A Pending CN114237206A (en) 2021-12-17 2021-12-17 Wind power variable pitch system fault detection method for complex operation conditions

Country Status (1)

Country Link
CN (1) CN114237206A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116089890A (en) * 2022-12-28 2023-05-09 大唐可再生能源试验研究院有限公司 Fault early warning detection method and system for intelligent wind power plant

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102135995A (en) * 2011-03-17 2011-07-27 新太科技股份有限公司 Extract transform and load (ETL) data cleaning design method
CN103488561A (en) * 2013-07-09 2014-01-01 沈阳化工大学 kNN (k-nearest neighbor) fault detection method for online upgrading master sample model
CN104808648A (en) * 2015-03-09 2015-07-29 杭州电子科技大学 Online and real-time batch process monitoring method based on k nearest neighbor
CN105134510A (en) * 2015-09-18 2015-12-09 北京中恒博瑞数字电力科技有限公司 State monitoring and failure diagnosis method for wind generating set variable pitch system
CN109459409A (en) * 2017-09-06 2019-03-12 盐城工学院 A kind of near-infrared exceptional spectrum recognition methods based on KNN

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102135995A (en) * 2011-03-17 2011-07-27 新太科技股份有限公司 Extract transform and load (ETL) data cleaning design method
CN103488561A (en) * 2013-07-09 2014-01-01 沈阳化工大学 kNN (k-nearest neighbor) fault detection method for online upgrading master sample model
CN104808648A (en) * 2015-03-09 2015-07-29 杭州电子科技大学 Online and real-time batch process monitoring method based on k nearest neighbor
CN105134510A (en) * 2015-09-18 2015-12-09 北京中恒博瑞数字电力科技有限公司 State monitoring and failure diagnosis method for wind generating set variable pitch system
CN109459409A (en) * 2017-09-06 2019-03-12 盐城工学院 A kind of near-infrared exceptional spectrum recognition methods based on KNN

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
申传华;: "数据挖掘过程中的数据清洗研究", 通讯世界, no. 24, pages 81 *
钱小毅: "基于k近邻与规则挖掘的风电机组故障诊断研究", 中国博士学位论文全文数据库（电子期刊）工程科技Ⅱ辑, no. 07, 15 July 2021 (2021-07-15), pages 042 - 3 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116089890A (en) * 2022-12-28 2023-05-09 大唐可再生能源试验研究院有限公司 Fault early warning detection method and system for intelligent wind power plant
CN116089890B (en) * 2022-12-28 2024-03-12 大唐可再生能源试验研究院有限公司 Fault early warning detection method and system for intelligent wind power plant

Similar Documents

Publication Publication Date Title
CN111539553B (en) Wind turbine generator fault early warning method based on SVR algorithm and off-peak degree
WO2023197461A1 (en) Gearbox fault early warning method and system based on working condition similarity evaluation
CN105631596B (en) Equipment fault diagnosis method based on multi-dimensional piecewise fitting
CN104809658B (en) A kind of rapid analysis method of low-voltage distribution network taiwan area line loss
Wang et al. A fast abnormal data cleaning algorithm for performance evaluation of wind turbine
CN110533092B (en) Wind generating set SCADA data classification method based on operation condition and application
CN110097209A (en) A kind of equipment deterioration analysis method based on parameter residual error
CN111260503B (en) Wind turbine generator power curve outlier detection method based on cluster center optimization
CN112101439B (en) High-speed wire quality defect diagnosis and tracing method based on distributed Bayesian network
CN114237206A (en) Wind power variable pitch system fault detection method for complex operation conditions
CN111814848B (en) Self-adaptive early warning strategy design method for temperature faults of wind turbine generator
CN114386718A (en) Wind power plant output power short-time prediction algorithm combined with particle swarm neural network
CN111582406A (en) Power equipment state monitoring data clustering method and system
CN117370744A (en) Dynamic cleaning method and system for abnormal power consumption data of power consumer
CN114563671A (en) High-voltage cable partial discharge diagnosis method based on CNN-LSTM-Attention neural network
Oprime et al. X-bar control chart design with asymmetric control limits and triple sampling
CN111273635B (en) Unknown anomaly detection method for industrial control equipment
CN115271423A (en) Wind generating set running state evaluation method based on hybrid similarity
CN114154567A (en) Wind power plant station operation data anomaly identification method based on machine learning
CN113986892A (en) Coal consumption prediction and data cleaning method and system for thermal power generating unit
CN114033631B (en) Online identification method for wind energy utilization coefficient of wind turbine generator
CN113505818A (en) Aluminum melting furnace energy consumption abnormity diagnosis method, system and equipment with improved decision tree algorithm
Ding et al. Wind power prediction based on multiple support vector machines
Wang et al. Wind turbine abnormal data cleaning method considering multi-scene parameter adaptation
CN114021992B (en) Evaluation method for real-time output performance of large wind turbine generator

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination