CN114139638A - Fan blade icing fault diagnosis method considering multivariable correlation - Google Patents

Fan blade icing fault diagnosis method considering multivariable correlation Download PDF

Info

Publication number
CN114139638A
CN114139638A CN202111470935.2A CN202111470935A CN114139638A CN 114139638 A CN114139638 A CN 114139638A CN 202111470935 A CN202111470935 A CN 202111470935A CN 114139638 A CN114139638 A CN 114139638A
Authority
CN
China
Prior art keywords
data
fan blade
variables
feature
steps
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
CN202111470935.2A
Other languages
Chinese (zh)
Inventor
赵平
杨立
赵健
陈晋
吴姗
龚学海
冯钰玮
龙新
犹珀玉
刘奇
李贤烨
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.)
PowerChina Guizhou Electric Power Engineering Co Ltd
Original Assignee
PowerChina Guizhou Electric Power Engineering Co Ltd
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 PowerChina Guizhou Electric Power Engineering Co Ltd filed Critical PowerChina Guizhou Electric Power Engineering Co Ltd
Priority to CN202111470935.2A priority Critical patent/CN114139638A/en
Publication of CN114139638A publication Critical patent/CN114139638A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a method for diagnosing icing fault of a fan blade by considering multivariate correlation, which comprises the steps of obtaining SCADA original data, and extracting environmental variables and equipment variables influencing fan power generation; extracting depth related information: considering the correlation among variables and deeply exploring implication information in data as characteristic data information; using a principal component analysis method to reduce the dimension of the characteristic data; initializing model parameters, inputting a training data set into a Gentle Adaboosts algorithm, continuously training the model to obtain a strongest fault diagnosis model, and testing the performance of the model in a test data set to obtain a final blade icing fault diagnosis result; the technical problems that in the prior art, mutual relations among a plurality of variables are not considered or deeply studied in the process of diagnosing the icing fault of the fan blade, data characteristics are not fully utilized, and the precision of the fault diagnosis result is low are solved.

Description

Fan blade icing fault diagnosis method considering multivariable correlation
Technical Field
The invention belongs to the technical field of fan blade fault diagnosis, and particularly relates to a fan blade icing fault diagnosis method considering multivariable correlation.
Background
Under the large energy background of 'carbon peak reaching and carbon neutralization', the installed capacity of a wind turbine generator is continuously increased. Wind energy as a green pollution-free clean energy has higher and higher occupation ratio in energy market[1]. Therefore, the key technology for the operation and maintenance of the wind turbine generator is widely concerned by researchers at home and abroad. Wind power installations are generally installed offshore or onshore, and the installation location is usually an area where wind resources are abundant. While these areas provide wind power generation resources, their environment also places higher demands on the maintenance of the wind turbine. Generally, a wind turbine generator is in a high-altitude and low-temperature environment, so that the blades of the wind turbine generator are easy to generate an icing condition, the running performance of the wind turbine generator is reduced, huge hidden dangers are generated on equipment safety, and even the blades are deformed or broken. The method for diagnosing the icing fault of the fan blade is researched, so that the hidden danger of the icing of the blade can be eliminated in time, the negative influence of environmental factors on a wind turbine generator is reduced, and the stable operation of the wind turbine generator is ensured. Therefore, the fan blade icing fault diagnosis method is deeply researchedHas important significance.
At present, methods for diagnosing the icing fault of the fan blade comprise an ultrasonic method, a thermal infrared radiation method, a hyperspectral imaging method and other methods formed according to an icing mechanism. However, these methods require a large number of sensors, which increases economic investment. Meanwhile, the aging and maintenance of the sensor not only needs to invest capital again, but also increases the time cost of diagnosis. An icing fault diagnosis model is established by mining potential information of operation and maintenance data, and the data is developed as an algorithm-driven fault diagnosis method. Real-time Data of the operation of the wind turbine generator is acquired And stored by a Supervisory Control And Data Acquisition (SCADA) system. Because the SCADA data records a large amount of environmental, electrical and mechanical information related to the wind turbine, many researches analyze the state data of the wind turbine from the perspective of machine learning or deep learning to diagnose whether the icing fault occurs to the wind turbine blade. The Neural Network (NN) has the advantages of processing a nonlinear problem, and researches show that the NN-based fan fault early warning method has applicability, but the characteristic extraction capability in mass data is still insufficient, so that the NN model is not high in precision when the problem is processed and diagnosed. According to the traditional method, the applicability and the precision of the fan blade icing fault diagnosis technology are improved through combined optimization of classification models, however, the characteristic data redundancy is caused due to the lack of processing of data characteristics. For high-dimensional feature data, unnecessary components are not removed, and the performance of a diagnosis model is reduced. A classical method also considers the formation mode of the feature vector, a classical Support Vector Machine (SVM) is improved through an ion cluster optimization algorithm, and the diagnosis precision of the icing fault of the wind turbine generator is improved. Or before the ice coating fault diagnosis is carried out by using the classical SVM method, the features in the training data are selected by using the improved RF method, and compared with a single SVM method, the fault diagnosis accuracy is better. However, these methods only use the feature information of a single variable in the data, and do not consider or deeply study the interrelationship between multiple variables, and the data features are not fully utilized, resulting in low accuracy of the fault diagnosis result.
The invention content is as follows:
the technical problems to be solved by the invention are as follows: the method for diagnosing the icing fault of the fan blade by considering the multivariable correlation is provided, and aims to solve the problems that in the prior art, the mutual relation among a plurality of variables is not considered or deeply studied in the process of diagnosing the icing fault of the fan blade, the data characteristics are not fully utilized, and the accuracy of the icing fault diagnosis result is low.
The technical scheme of the invention is as follows:
a method for diagnosing icing fault of a fan blade by considering multivariable correlation comprises the following steps:
step 1, data preprocessing: acquiring SCADA original data, and extracting environment variables and equipment variables influencing the power generation of a fan;
step 2, extracting depth related information: considering the correlation among variables and deeply exploring implication information in data as characteristic data information;
step 3, feature data dimension reduction: using a principal component analysis method to reduce the dimension of the characteristic data;
and 4, initializing model parameters, inputting the training data set into a Gentle Adaboots algorithm, and continuously training the model until the number of training cycles is reached to obtain the blade icing fault diagnosis model. And testing the performance of the model in the test data set to obtain a final blade icing fault diagnosis result and verifying the precision of the result.
The specific method for preprocessing the data in the step 1 comprises the following steps: acquiring SCADA original data, extracting environmental variables and equipment variables influencing fan power generation, and judging whether the fan blade has an icing fault or not by analyzing data results of different variables under the conditions of fan blade fault and normal condition; dividing the processed data into a training data set and a testing data set according to the ratio of 8: 2; respectively normalizing the training data and the test data by a formula (1) to eliminate errors of magnitude and dimension;
Figure BDA0003392145570000031
in the formula
Figure BDA0003392145570000032
Is normalized value, X is original sample data, XmaxAnd XminThe maximum and minimum values for the column in which the data is located.
The specific method for extracting the depth related information in the step 2 comprises the following steps: and (4) applying the 1D-CNN to the normalized data to realize extraction of multivariate correlation information. The feature vector of the data input is [ x ]i1 xi2…xi26]The feature vector dimension is 26, and i represents the size of the sample size. Setting 3 convolution kernels with different sizes, wherein the sizes are 1 multiplied by 7, 1 multiplied by 5 and 1 multiplied by 3 respectively; the number of convolution kernels of each size is 30; the pooling kernel size was 1 × 2.
The 1D-CNN acts directly on the feature vector. The purpose of extracting the features is achieved by setting the sizes and the number of convolution kernels in the convolution layer and the pooling layer; the method specifically comprises the following steps: the convolution layer performs convolution operation on the input feature vector and the convolution kernel to obtain feature information in different directions among variables in an evaluable vector, and the operation expression of the convolution layer is as follows:
Xj＝f(W＊Xi+bi) (2)
wherein W is the i-th layer convolution kernel convolution matrix, x is the convolution operator, and b is the offset term. The relu function is used as an activation function f (z);
Figure BDA0003392145570000041
after the operation of the convolutional layer, a feature vector with a large dimension is formed, the feature vector is subjected to down-sampling by the pooling layer, and the significant features in the extracted data are enhanced by setting the size of the pooling kernel. Filtering unimportant data information, wherein the concrete expression of the pooling layer is as follows:
Figure BDA0003392145570000042
in the formula, Xj(k) Elements in a pooling region for layer j feature matricesAnd element, P is the characteristic matrix element after the pooling treatment, and D is the pooling core coverage area.
The specific method for characterizing data dimension reduction in step 3 is as follows: reducing the dimension of the feature data by using a principal component analysis method, reducing the dimension of the feature vector obtained in the step 2 by using a PCA algorithm, extracting the feature vector of 95 percent of principal components, retaining the significant information of the feature data, and removing redundant data in the feature; the method specifically comprises the following steps:
after the fan operation data collected by the SCADA system is processed by 1D-CNN, the matrix data B of m multiplied by n is formed as follows:
Figure BDA0003392145570000051
the data comprises m sample vectors, the dimension of each vector is n, and the covariance C of the matrix is calculated by a formula (6);
Figure BDA0003392145570000052
obtaining the eigenvalue and eigenvector of the matrix C through a diagonalization matrix; v is an eigenvector, and S is an eigenvalue diagonal matrix of the matrix C; s diagonal matrix elements are the variance of the principal components, the expression of S is shown in a formula (7), and the elements are arranged from large to small; calculating the contribution rate according to the sum of the variances; selecting 95% contribution rate principal component to form a new eigenvector, and representing the icing state of the fan blade
S＝V1CV (7)。
The Gentle Adaboots algorithm includes:
there are n training samples, labeled { (x)1,y1),(x2,y2),…(xn,yn) In which y isi∈{-1,1}，i∈{1…n}。yiExpressed as an icing fault sample, yi-1 is expressed as a normal run sample;
the sample weights are initialized to:
Figure BDA0003392145570000053
j∈{1…n}；
for K is 1,2, … K (K is cycle number);
a. using wjAs the weight of each sample, using as weighted least squares regression to obtain weak classifier fj(x)；
b. Update the weight wj+1→wj×exp(-y×fj(x) J ∈ {1 … n }, renormalized such that ΣjDj＝1；
Outputting a final strong classifier:
Figure BDA0003392145570000061
environmental variables and equipment variables affecting the wind turbine generation are shown in table 1:
TABLE 1
Figure BDA0003392145570000062
The invention has the beneficial effects that:
the invention provides a Gentle Adaboost fan blade icing fault diagnosis method considering multivariate correlation for further improving the precision of fan blade icing fault diagnosis. Meanwhile, the dimensionality of the characteristic vector is reduced by adopting PCA, key components influencing the fault diagnosis precision in the vector are reserved, and the influence of non-key components on a diagnosis model is weakened. And introducing the final characteristic vector into a Gentle Adaboost algorithm with strong robustness to obtain an icing fault diagnosis result.
The invention sufficiently explores the characteristic information of SCADA data and provides a Gentle Adaboost fan blade icing fault diagnosis considering multivariate correlation. The existing fault diagnosis method is used for singly utilizing the variables of the collected data, and the relevance information among the variables is not considered enough, so that the invention utilizes the one-dimensional convolutional neural network to extract the characteristic information in the SCADA collected data, thereby expressing the relevance among the variables and forming the characteristic vector with strong expression capability. Reducing the dimensionality of the feature vector by a Principal Component Analysis (PCA), extracting key components in the vector which can influence a diagnosis result, retaining significant information of data, and weakening the interference of redundant information on a diagnosis model; and finally, importing the feature vector into an integrated learning algorithm GentleAdaboost with strong robustness to learn, obtaining an icing fault diagnosis model with high generalization capability, and outputting a high-precision diagnosis result.
The method solves the technical problems that in the prior art, the mutual relation among a plurality of variables is not considered or deeply studied in the fan blade icing fault diagnosis, the data characteristics are not fully utilized, the fault diagnosis result precision is low and the like. The method can further improve the precision of the icing fault diagnosis of the wind generating set, assist the operation and maintenance personnel of the wind power plant to remove the blade icing in time, and shorten the period of eliminating the hidden danger.
Description of the drawings:
FIG. 1 is a schematic flow chart of the present invention.
FIG. 2 is a schematic diagram of the structure arrangement of 1D-CNN.
FIG. 3 is a schematic diagram of 1D-CNN feature extraction.
The specific implementation mode is as follows:
step 1: and (4) preprocessing data. The method comprises the steps of obtaining SCADA original data, extracting environment variables and equipment variables influencing fan power generation, and judging whether the fan blade has an icing fault or not by analyzing data results of different variables under the conditions of fan blade faults and normal conditions. The processed data is divided into a training data set and a testing data set according to the ratio of 8: 2. The training data and the test data are respectively normalized by a formula (1), and errors of magnitude and dimension are eliminated.
Figure BDA0003392145570000081
In the formula
Figure BDA0003392145570000082
Is normalized value, X is original sample data, XmaxAnd XminThe maximum and minimum values for the column in which the data is located.
TABLE 1 SCADA Fan monitoring variables data
Figure BDA0003392145570000083
Figure BDA0003392145570000091
Step 2: and extracting depth related information. In order to consider the interrelation among a plurality of variables and deeply discover the implication information in the data, the invention applies 1D-CNN to the normalized data to realize the extraction of multivariable correlation information. The feature vector of the data input is [ x ]i1 xi2…xi26]The feature vector dimension is 26, and i represents the size of the sample size. 3 convolution kernels of different sizes are set, and the sizes of the convolution kernels are 1 × 7, 1 × 5 and 1 × 3 respectively. The number of convolution kernels per size is 30. The pooled kernel size is 1 × 2, and its feature automatic extraction structure is set up as shown in fig. 2.
The CNN is mainly used for automatically extracting characteristic information in data and is one of the most widely applied models in the field of deep learning. The CNN structure is constructed by a convolution layer and a pooling layer, and dense and complete feature vectors in data are extracted in a local connection and weight sharing mode. In the invention, related information among multiple variables in the feature vector is extracted, and a 1D-CNN method is adopted to directly act on the feature vector. The purpose of extracting the features is achieved by setting the sizes and the number of convolution kernels in the convolution layer and the pooling layer. The 1D-CNN feature extraction structure is shown in FIG. 3.
The convolution layer performs convolution operation on the input feature vector and the convolution kernel to obtain feature information in different directions among variables in an evaluable vector, and the operation expression of the convolution layer is as follows:
Xj＝f(W＊Xi+bi) (2)
wherein W is the i-th layer convolution kernel convolution matrix, x is the convolution operator, and b is the offset term. The relu function is taken as the activation function f (z).
Figure BDA0003392145570000101
After the operation of the convolutional layer, a feature vector with a large dimension is formed, the feature vector is subjected to down-sampling by the pooling layer, and the significant features in the extracted data are enhanced by setting the size of the pooling kernel. Filtering out unimportant data information. The concrete expression of the pooling layer is as follows:
Figure BDA0003392145570000102
in the formula, Xj(k) The elements of the j-th layer feature matrix in the pooling area are shown, P is the feature matrix elements after the pooling process, and D is the pooling core coverage area.
And step 3: and (5) reducing the dimension of the feature data. The invention uses principal component analysis method to reduce the dimension of the feature data, and PCA algorithm is used to reduce the dimension of the feature vector obtained in step 2, extract the feature vector of 95% principal component, and keep the significant information of the feature data. Reducing the effect of redundant noise data on diagnostic model accuracy by removing redundant data from features
The principal component analysis method is a dimensionality reduction method based on a mathematical statistics theory. The dimensionality of the data is reduced by means of orthogonal transformation, while preserving significant information in the data. Assuming that after the fan operation data collected by the SCADA system is processed by 1D-CNN, the matrix data B of m × n is formed as follows:
Figure BDA0003392145570000103
the data contains m sample vectors, each of which has dimensions n. The covariance C of the matrix is calculated from equation (6).
Figure BDA0003392145570000104
And obtaining the eigenvalue and the eigenvector of the matrix C by the diagonalized matrix. V is the eigenvector and S is the eigenvalue diagonal matrix of matrix C. The S diagonal matrix elements are the principal component variance, the S expression is shown in formula (7), and the elements are arranged from large to small. The contribution rate is calculated from the sum of variances. And selecting the 95% contribution rate principal component to form a new eigenvector, and representing the icing state of the fan blade.
S＝V1CV (7)
And 4, step 4: and (5) training and testing the model. Initializing model parameters, inputting a training data set into Gentle Adaboosts, and continuously training by the model until reaching the training cycle number to obtain a strong fault diagnosis model, namely a blade icing fault diagnosis model. And testing the performance of the model in the test data set to obtain a final blade icing fault diagnosis result and verifying the precision of the result.
Gentle Adaboost is an improvement of the adaptive boosting (AdaBoost) of the classical ensemble learning algorithm. The Adaboost algorithm trains different weak classifiers from a training set, and then combines the weak classifiers to finally form a strong classifier. The traditional Adaboost is easy to cause overfitting, and the generalization capability of the model is weak. Gentle Adaboost is one of Adaboost's improved algorithms that has the best robustness and good classification performance. The difference between the Adaboost algorithm and the Adaboost algorithm is that the weight of training samples of the weak classifiers is continuously updated, the weight is smoothed, and the weak classifier with the minimum error rate is selected to form the strong classifier. The Gentle Adaboots algorithm mainly comprises the following steps:
1) there are n training samples, labeled { (x)1,y1),(x2,y2),…(xn,yn) In which y isi∈{-1,1}，i∈{1…n}。yiExpressed as an icing fault sample, yi-1 is expressed as a normal run sample;
2) the sample weights are initialized to:
Figure BDA0003392145570000111
j∈{1…n}；
3) for K is 1,2, … K (K is cycle number);
a. using wjAs the weight of each sample, using as weighted least squares regression to obtain weak classifier fj(x)；
b. Update the weight wj+1→wj×exp(-y×fj(x) J ∈ {1 … n }, renormalized such that ΣjDj＝1。
4) Outputting a final strong classifier:
Figure BDA0003392145570000121
the method of the invention processes the characteristic vector by 1D-CNN, and deeply excavates the relevant information among variables, so that the information among the variables is fully utilized. Meanwhile, in order to weaken data redundancy caused by the high-dimensional feature vector, the PCA is used for extracting the significant features in the high-dimensional feature vector. The 1D-CNN characteristic vector processing mode can effectively extract the relevant information among a plurality of variables and deeply mine the implicit characteristics of data. The dimensionality of redundant data can be reduced by adopting the PCA method, the data significant information is further expressed and utilized, and the precision of a fault diagnosis model is optimized.

Claims (7)

1. A method for diagnosing icing fault of a fan blade by considering multivariable correlation comprises the following steps:
step 1, data preprocessing: acquiring original data of an SCADA fan data acquisition and monitoring system, and extracting environment variables and equipment variables influencing fan power generation;
step 2, extracting depth related information: considering the correlation among variables and deeply exploring implication information in data as characteristic data information;
step 3, feature data dimension reduction: using a principal component analysis method to reduce the dimension of the characteristic data;
step 4, initializing model parameters, inputting a training data set into a Gentle Adaboots algorithm, and continuously training the model until the number of training cycles is reached to obtain a blade icing fault diagnosis model; and testing the performance of the model in the test data set to obtain a final blade icing fault diagnosis result and verifying the precision of the result.
2. The method for diagnosing the icing fault of the fan blade considering the multivariable correlation according to claim 1, wherein the method comprises the following steps: the specific method for preprocessing the data in the step 1 comprises the following steps: acquiring SCADA original data, extracting environmental variables and equipment variables influencing fan power generation, and judging whether the fan blade has an icing fault or not by analyzing data results of different variables under the conditions of fan blade fault and normal condition; dividing the processed data into a training data set and a testing data set according to the ratio of 8: 2; respectively normalizing the training data and the test data by a formula (1) to eliminate errors of magnitude and dimension;
Figure FDA0003392145560000011
in the formula
Figure FDA0003392145560000012
Is normalized value, X is original sample data, XmaxAnd XminThe maximum and minimum values for the column in which the data is located.
3. The method for diagnosing the icing fault of the fan blade considering the multivariable correlation according to claim 1, wherein the method comprises the following steps: the specific method for extracting the depth related information in the step 2 comprises the following steps: and (4) applying the 1D-CNN to the normalized data to realize extraction of multivariate correlation information. The feature vector of the data input is [ x ]i1 xi2…xi26]The feature vector dimension is 26, and i represents the size of the sample size. Setting 3 convolution kernels with different sizes, wherein the sizes are 1 multiplied by 7, 1 multiplied by 5 and 1 multiplied by 3 respectively; the number of convolution kernels of each size is 30; the pooling kernel size was 1 × 2.
4. The method for diagnosing the icing fault of the fan blade considering the multivariable correlation according to claim 3, wherein the method comprises the following steps: the 1D-CNN acts directly on the feature vector. The purpose of extracting the features is achieved by setting the sizes and the number of convolution kernels in the convolution layer and the pooling layer; the method specifically comprises the following steps: the convolution layer performs convolution operation on the input feature vector and the convolution kernel to obtain feature information in different directions among variables in an evaluable vector, and the operation expression of the convolution layer is as follows:
Xj＝f(W＊Xi+bi) (2)
wherein, W is the i-th layer convolution kernel convolution matrix, x is the convolution operator, and b is the bias term; the relu function is used as an activation function f (z);
Figure FDA0003392145560000021
after the operation of the convolutional layer, a feature vector with a large dimension is formed, the feature vector is subjected to down-sampling by the pooling layer, and the significant features in the extracted data are enhanced by setting the size of the pooling kernel. Filtering unimportant data information, wherein the concrete expression of the pooling layer is as follows:
Figure FDA0003392145560000022
in the formula, Xj(k) The elements of the j-th layer feature matrix in the pooling area are shown, P is the feature matrix elements after the pooling process, and D is the pooling core coverage area.
5. The method for diagnosing the icing fault of the fan blade considering the multivariable correlation according to claim 1, wherein the method comprises the following steps: step 3, the specific method for reducing the dimension of the feature data comprises the following steps: reducing the dimension of the feature data by using a principal component analysis method, performing dimension reduction processing on the feature vectors obtained in the step 2 by using a PCA algorithm, extracting the feature vectors of 95% of principal components, retaining the significant information of the feature data, and removing redundant data in the features; the method specifically comprises the following steps:
after the fan operation data collected by the SCADA system is processed by 1D-CNN, the matrix data B of m multiplied by n is formed as follows:
Figure FDA0003392145560000031
the data comprises m sample vectors, the dimension of each vector is n, and the covariance C of the matrix is calculated by a formula (6);
Figure FDA0003392145560000032
obtaining the eigenvalue and eigenvector of the matrix C through a diagonalization matrix; v is an eigenvector, and S is an eigenvalue diagonal matrix of the matrix C; s diagonal matrix elements are the variance of the principal components, the expression of S is shown in a formula (7), and the elements are arranged from large to small; calculating the contribution rate according to the sum of the variances; selecting 95% contribution rate principal component to form a new eigenvector, and representing the icing state of the fan blade
S＝V1CV (7)。
6. The method for diagnosing the icing fault of the fan blade considering the multivariable correlation according to claim 1, wherein the method comprises the following steps: the Gentle Adaboots algorithm includes:
there are n training samples, labeled { (x)1,y1),(x2,y2),…(xn,yn) In which y isi∈{-1,1}，i∈{1…n}。yiExpressed as an icing fault sample, yi-1 is expressed as a normal run sample;
the sample weights are initialized to:
Figure FDA0003392145560000041
for K is 1,2, … K (K is cycle number);
a. using wjAs the weight of each sample, using as weighted least squares regression to obtain weak classifier fj(x)；
b. Update the weight wj+1→wj×exp(-y×fj(x) J ∈ {1 … n }, renormalized such that ΣjDj＝1；
Outputting a final strong classifier:
Figure FDA0003392145560000042
7. the method for diagnosing the icing fault of the fan blade considering the multivariable correlation according to claim 1, wherein the method comprises the following steps: environmental variables and equipment variables affecting the wind turbine generation are shown in table 1:
TABLE 1
Figure FDA0003392145560000043
Figure FDA0003392145560000051
CN202111470935.2A 2021-12-03 2021-12-03 Fan blade icing fault diagnosis method considering multivariable correlation Pending CN114139638A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111470935.2A CN114139638A (en) 2021-12-03 2021-12-03 Fan blade icing fault diagnosis method considering multivariable correlation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111470935.2A CN114139638A (en) 2021-12-03 2021-12-03 Fan blade icing fault diagnosis method considering multivariable correlation

Publications (1)

Publication Number Publication Date
CN114139638A true CN114139638A (en) 2022-03-04

Family

ID=80387881

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111470935.2A Pending CN114139638A (en) 2021-12-03 2021-12-03 Fan blade icing fault diagnosis method considering multivariable correlation

Country Status (1)

Country Link
CN (1) CN114139638A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117892213A (en) * 2024-03-18 2024-04-16 中国水利水电第十四工程局有限公司 Diagnosis method for icing detection and early warning of wind driven generator blade
CN117992859A (en) * 2024-04-03 2024-05-07 华侨大学 Early-stage fault early-warning and identifying method and device for electromechanical equipment provided with SCADA system

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117892213A (en) * 2024-03-18 2024-04-16 中国水利水电第十四工程局有限公司 Diagnosis method for icing detection and early warning of wind driven generator blade
CN117992859A (en) * 2024-04-03 2024-05-07 华侨大学 Early-stage fault early-warning and identifying method and device for electromechanical equipment provided with SCADA system
CN117992859B (en) * 2024-04-03 2024-06-07 华侨大学 Early-stage fault early-warning and identifying method and device for electromechanical equipment provided with SCADA system

Similar Documents

Publication Publication Date Title
CN106682814B (en) Wind turbine generator fault intelligent diagnosis method based on fault knowledge base
CN111275288B (en) XGBoost-based multidimensional data anomaly detection method and device
CN111582392B (en) Multi-working-condition health state online monitoring method for key components of wind turbine generator
Wen et al. A new ensemble convolutional neural network with diversity regularization for fault diagnosis
CN110929847A (en) Converter transformer fault diagnosis method based on deep convolutional neural network
CN110685868A (en) Wind turbine generator fault detection method and device based on improved gradient elevator
CN110262450B (en) Fault prediction method for cooperative analysis of multiple fault characteristics of steam turbine
CN106933778A (en) A kind of wind power combination forecasting method based on climbing affair character identification
CN114139638A (en) Fan blade icing fault diagnosis method considering multivariable correlation
CN117290800B (en) Timing sequence anomaly detection method and system based on hypergraph attention network
CN105425779A (en) ICA-PCA multi-working condition fault diagnosis method based on local neighborhood standardization and Bayesian inference
CN112418277A (en) Method, system, medium, and apparatus for predicting remaining life of rotating machine component
Pan et al. Research on gear fault diagnosis based on feature fusion optimization and improved two hidden layer extreme learning machine
CN111680875B (en) Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model
CN109298633A (en) Chemical production process fault monitoring method based on adaptive piecemeal Non-negative Matrix Factorization
CN114048688A (en) Method for predicting service life of bearing of wind power generator
CN109184821A (en) A kind of on-line monitoring method of the closed-loop information analysis towards intelligent power plant's Generator Set steam turbine
CN114897138A (en) System fault diagnosis method based on attention mechanism and depth residual error network
CN114330881A (en) Data-driven fan blade icing prediction method and device
CN110033126A (en) Shot and long term memory network prediction technique based on attention mechanism and logistic regression
CN112163474B (en) Intelligent gearbox diagnosis method based on model fusion
Yang et al. Detection of wind turbine blade abnormalities through a deep learning model integrating VAE and neural ODE
Jiang et al. Recurrence plot quantitative analysis-based fault recognition method of rolling bearing
CN108459585A (en) Power station fan method for diagnosing faults based on sparse locally embedding depth convolutional network
CN115310746A (en) State evaluation method and system for main transmission system of wind generating set

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