CN109376801A - Blade of wind-driven generator icing diagnostic method based on integrated deep neural network - Google Patents
Blade of wind-driven generator icing diagnostic method based on integrated deep neural network Download PDFInfo
- Publication number
- CN109376801A CN109376801A CN201811468877.8A CN201811468877A CN109376801A CN 109376801 A CN109376801 A CN 109376801A CN 201811468877 A CN201811468877 A CN 201811468877A CN 109376801 A CN109376801 A CN 109376801A
- Authority
- CN
- China
- Prior art keywords
- data
- wind
- driven generator
- self
- blade
- 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.)
- Granted
Links
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
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Wind Motors (AREA)
Abstract
The invention proposes a kind of blade of wind-driven generator icing diagnostic method based on integrated deep neural network mainly solves to need additional additional equipment, problem at high cost, detection accuracy is low when prior art detection blade of wind-driven generator fault condition.Implementation step includes: the historical data of 1) acquisition wind-driven generator operation;2) historical data is pre-processed;3) data balancing processing is carried out to pretreated historical data, obtains training dataset;4) it constructs integrated deep neural network and it is trained using training dataset;5) data prediction is carried out to the new real time data for flowing into wind-driven generator;6) by the integrated deep neural network model after the input training of pretreated real time data, the diagnosis of blade icing condition is carried out.The present invention realizes the real-time monitoring to freeze to blade of wind-driven generator under the premise of not introducing other measuring devices, and substantially increases the accuracy rate of blade of wind-driven generator icing diagnosis.
Description
Technical field
The invention belongs to plant maintenance fields, specially a kind of further to the predictive maintenance of wind power plant
Blade of wind-driven generator icing diagnostic method based on integrated deep neural network.It can be used for rapidly and accurately to wind-driven generator
Carry out fault diagnosis.
Background technique
Wind-driven generator obtains wind energy by blade, and converts wind energy into mechanical energy, so that driven generator generates electricity,
And then realize the conversion of wind energy to electric energy.Therefore, blade is one of the core component for guaranteeing the reliable high-efficiency operation of wind-driven generator.
In different geographical, wind-driven generator can face varying environment, and the wind resource in China enrich region many places in
The lower northwestern of temperature, while with the continuous promotion of modern wind generator design power, blade and tower height
It is continuously increased, lower cloud layer is touched in tending in winter.Under low temperature and high relative humidity environment, fan blade is easy to tie
Ice problem, this is also that current wind power plant safeguards faced one of global problem.Low temperature and high relative humidity environment is caused
It is leaf that blade ice formation issues would generally change blade, destroys blade aerodynamic characteristic, and cause blade material, structural behaviour and load
The change of lotus distribution so as to cause the decline of wind turbine power generation efficiency, and causes greatly to threaten to the safe operation of blower.Therefore,
In order to effectively improve fan operation efficiency and ensure operational safety, freezing to diagnose further investigate to blade of wind-driven generator is
Very necessary.
From the point of view of current disclosed data, diagnosis is frozen mainly by additionally adding on blade to blade of wind-driven generator
Add icing monitoring device.For example, application publication number is CN207701300U, entitled " blade of wind-driven generator freezes supervises online
The patent application of survey device ", discloses a kind of blade of wind-driven generator icing on-Line Monitor Device.The on-Line Monitor Device passes through
It installs sensor additional on leaflet inner faces, blade loading signal and vibration signal is acquired by data acquisition unit, thus to obtain
The running state information of draught fan impeller equipment.Although this method can be realized the monitoring to blade icing condition, however, it is in wind
It installs sensor on power generator blade additional, not only changes the blade construction of wind-driven generator, increases cost;And due to adding
The sensor of dress is exposed under adverse circumstances locating for blade of wind-driven generator for a long time, and meeting acceleration sensor function is moved back
Change, cause its service life shorter, if replacing the accuracy that can also seriously affect the diagnosis that freezes not in time.
SCADA system is the standard configuration of wind-driven generator, and the SCADA data amount of acquisition is huge, and contains and leaf
Piece freezes closely related valuable information.The valuable information in SCADA data is excavated, only manually method can not almost be done
It arrives.Currently, the rise of deep learning method is that fault diagnosis field brings new research direction.Wherein, stacking-type encodes certainly
Device model is widely used because of the features such as its structure is simple, characterization ability is strong.For example, application publication number is
CN106919164A, the patent application of entitled " the water conservancy unit failure analysis methods based on storehouse autocoder " are open
A kind of water conservancy unit failure analysis methods based on storehouse autocoder.This method is using stacking-type self-encoding encoder to water conservancy
The data set of unit carries out deep learning and training, and adaptive learning and the water conservancy unit for realizing water conservancy set state feature are strong
The assessment of health state.However, its deficiency is that the generalization ability of stacking-type self-encoding encoder model is weaker, assessment result is resulted in
Accuracy is lower, limits the practical engineering application of stacking-type self-encoding encoder model.
Summary of the invention
It is an object of the invention to overcome the problems of the above-mentioned prior art, provide a kind of based on integrated depth nerve
The blade of wind-driven generator icing diagnostic method of network, it is intended under the premise of not introducing other measuring devices, realize and wind-force is sent out
The real-time monitoring that motor blade freezes, and improve the accuracy rate of blade of wind-driven generator icing diagnosis.
To achieve the above object, the technical solution that the present invention takes includes the following steps:
(1) historical data of acquisition wind-driven generator operation;
(2) data prediction is carried out to collected historical data, obtains pretreated historical data;
(3) data balancing processing is carried out to pretreated historical data, obtains training dataset x_label;
(4) it constructs integrated deep neural network and it is trained using training dataset x_label, after obtaining training
Integrated deep neural network model, it is specific as follows:
(4a) builds the basic structure of stacking-type self-encoding encoder model:
Set the number L of stacking-type self-encoding encoder model, the input layer number u of each self-encoding encoderlAnd hidden layer section
Count vl, wherein l=1,2 ..., L, L >=3 and be odd number, the number of k-th of stacking-type self-encoding encoder model is lk, and k=1,
2,...,L;The output layer number of nodes u of first self-encoding encoder1It is self-editing after first equal to the Characteristic Number of training dataset
The input layer number u of code devicetWith the node in hidden layer v of its previous self-encoding encodert-1It is equal, wherein t=2,3 ..., L;
(4b) trains stacking-type self-encoding encoder model:
(4b1) enables k=1;
(4b2) training number is lkSelf-encoding encoder model:
(4b2a) initializes weight matrix W, learning rate η and offset vector b', b ", and sets maximum cycle as I;
(4b2b) removes the class label of training dataset x_label, obtains no label training dataset x:
X={ xi, wherein x ∈ Ru, i ∈ { 0,1,2 ... ..., n-1, n } },
Wherein, n is the number of samples of training dataset, and i is sample serial number, and u is the dimension that each sample corresponds to space vector
Degree;
Number l will be used as without label training dataset xkSelf-encoding encoder model input;
It is l that number, which is calculated, in (4b2c) according to the following formulakSelf-encoding encoder model hidden layer output yk:
It is l that number is calculated according to the following formulakSelf-encoding encoder model output layer output:
Wherein, WTFor the transposition of weight matrix W;
(4b2d) enables j=1;
(4b2e) is according to formula Undated parameter, wherein To compile
Number be lkSelf-encoding encoder model output layer outputThe vector of middle sample serial number i;Wj+1,b′j+1,b″j+1For jth time ginseng
The result that number updates;
(4b2f) judges whether j is equal to I, if so, number is l after the result for taking I subparameter to update is trainedk's
Self-encoding encoder model l'kAnd enter step (4b3);Otherwise return step (4b2e) after adding 1 to j;
(4b3) will be numbered after previous step training as lkSelf-encoding encoder model l'kHidden layer hkIt is l as numberk+1's
The input layer of self-encoding encoder model repeats step (4b2c)-(4b2f) and obtains number lk+1Self-encoding encoder model l after training
'k+1;
(4b4) judges whether k is equal to L-1, if the stacking-type self-encoding encoder model Q after then being trained;Conversely, to k
Return step (4b3) after adding 1;
(4c) builds integrated deep neural network model and is trained:
(4c1) enables k=1;
(4c2) extracts preceding k self-encoding encoder model in the stacking-type self-encoding encoder model Q after training, is only included k
The stacking-type self-encoding encoder model N of a self-encoding encoder modelk, in model NkThe last one self-encoding encoder model hidden layer hk
A Softmax classifier m is connected belowk, and by model NkWith Softmax classifier mkAs integrated deep neural network mould
The l of typekA base classifier;
(4c3) trains lkA base classifier:
Function h is assumed in definitionθ(X(i)) expression formula is as follows:
Wherein:TIt is operated for transposition,For the value of class label,Value for class label isWhen
Data sample belongs to such probability value;
Using training dataset x_label as lkThe input of a base classifier, using BP algorithm to the base classifier into
Row training, then obtains self-encoding encoder model and the parameter θ by fine tuning, to complete lkThe training of a base classifier;
(4c4) judges whether k is equal to L, if so, completing the training to L base classifier to get integrated to after training
Deep neural network model;Conversely, return step (4c2) after adding 1 to k;
(5) data prediction is carried out to the new real time data for flowing into wind-driven generator, obtains pretreated real time data
data;
(6) by the integrated deep neural network model after pretreated real time data data input training, lead to
The model is crossed to diagnose the blade icing condition of wind-driven generator, the specific steps are as follows:
(6a) sets the blade state of wind-driven generator as r and r ∈ { 0,1 }, and definition status 0 is icing condition, state 1 is not
Icing condition;
Real time data data is inputted integrated deep neural network model by (6b), obtains the output h of each base classifierθ, λ
(X(i))=[Pλ(y'i=r | x'i;θ)], wherein λ be base classifier number, λ=1,2 ..., L, pλ(y'i=r | x'i) indicate
I-th of sample of real time data belongs to the probability of state r;
(6c) takes λ=1,2 respectively ..., L obtains p1,p2,…,pL, i.e. the L respective output of base classifier;
(6d) finally judges the output result of all base classifiers using ballot method:
(6d1) counts p1(y'i=0 | x'i), p2(y'i=0 | x'i) ..., pL(y'i=0 | x'i) in numerical value greater than 0.5
Number, and its number is denoted as s;
(6d2) ifThen it is determined as that the blade of wind-driven generator is in icing condition, otherwise determines that it is normally not
It freezes.
Compared with the prior art, the invention has the following advantages:
The first, since the present invention is by the way of having operation information to wind-driven generator and being analyzed and processed, obtaining should
Blade of wind-driven generator icing diagnostic result, without adding additional icing measuring device on wind power plant;This
Invention, which does not introduce additional measuring device, can substantially reduce the working strength of plant maintenance personnel, and save unnecessary hardware at
Sheet and cost of labor;
The second, blade of wind-driven generator icing diagnosis is realized using integrated deep neural network model due to the present invention,
During diagnosis, integrated deep neural network model adaptively learns the implicit information of wind-driven generator operation data, and instructs
Practice multiple base classifiers, comprehensive analysis is carried out by the result to multiple classifiers, obtains examining for blade of wind-driven generator icing
Disconnected result;There is compared with using the classifiers such as stacking-type self-encoding encoder stronger generalization ability and diagnosis essence using the model
Degree, therefore, the present invention largely improve the accuracy rate of blade of wind-driven generator icing diagnosis.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention;
Fig. 2 is the structural schematic diagram of the integrated deep neural network of the present invention.
Specific embodiment
Below in conjunction with the drawings and specific embodiments, invention is further described in detail:
The continuous bimestrial SCADA data and blade knot of the same typhoon power generator in the data set source that this example uses
Ice daily record data, it is wherein specifically as shown in table 1 comprising parameter in SCADA data.This example is made using the data set of first month
For the operation data of acquisition, integrated deep neural network model is established, it is real-time using the data set of second month as what is newly flowed into
Data test model.
Table 1
Referring to Fig.1, a kind of blade of wind-driven generator icing diagnostic method based on integrated deep neural network, including it is as follows
Step:
Step 1: the historical data of acquisition wind-driven generator operation, SCADA data and leaf including wind-driven generator operation
Piece icing daily record data;
Step 2: data prediction is carried out to collected historical data, obtains pretreated historical data;
(2a) removes the abnormal point in collected historical data, obtains the historical data for being no different constant value;
(2b) carries out nondimensionalization processing to the historical data for being no different constant value, obtains pretreated historical data.
Step 3: using SMOTE algorithm, carries out data balancing processing to pretreated historical data, obtains training number
According to collection x_label;In this step data balancing processing also can by balance random forests algorithm, SMOTEBoost algorithm,
Metacost algorithm or Adacost algorithm are realized.
Step 4: it constructs integrated deep neural network and it is trained using training dataset x_label, obtain instruction
Integrated deep neural network model after white silk, specific as follows:
The basic structure of integrated deep neural network is made of stacking-type self-encoding encoder link sort device, wherein each from
The hidden layer of encoder is all connected to a classifier, it may be assumed that several self-encoding encoders stack the stacking-type self-encoding encoder to be formed with
It is connected to the classifier after each self-encoding encoder hidden layer and has collectively constituted integrated deep neural network.
(4a) builds the basic structure of stacking-type self-encoding encoder model:
Set the number L of stacking-type self-encoding encoder model, the input layer number u of each self-encoding encoderlAnd hidden layer section
Count vl, wherein l=1,2 ..., L, L >=3 and be odd number, the number of k-th of stacking-type self-encoding encoder model is lk, and k=1,
2,...,L;The output layer number of nodes u of first self-encoding encoder1It is self-editing after first equal to the Characteristic Number of training dataset
The input layer number u of code devicetWith the node in hidden layer v of its previous self-encoding encodert-1It is equal, wherein t=2,3 ..., L;
The present embodiment takes L=3, and the node in hidden layer of 3 self-encoding encoders is 173,115,59 respectively.
(4b) trains stacking-type self-encoding encoder model:
(4b1) enables k=1;
(4b2) training number is lkSelf-encoding encoder model:
(4b2a) initializes weight matrix W using random number method, takes learning rate η=0.001, offset vector b', b " setting
ForAnd set maximum cycle I=200;
(4b2b) removes the class label of training dataset x_label, obtains no label training dataset x:
X={ xi, wherein x ∈ Ru, i ∈ { 0,1,2 ... ..., n-1, n } },
Wherein, n is the number of samples of training dataset, and i is sample serial number, and u is the dimension that each sample corresponds to space vector
Degree;This example shares 186626 training dataset samples, and the dimension that each sample corresponds to space vector is 17 dimensions;
Number l will be used as without label training dataset xkSelf-encoding encoder model input;
It is l that number, which is calculated, in (4b2c) according to the following formulakSelf-encoding encoder model hidden layer output yk:
It is l that number is calculated according to the following formulakSelf-encoding encoder model output layer output
Wherein, WTFor the transposition of weight matrix W;
(4b2d) enables j=1;
(4b2e) is according to formula Undated parameter, wherein To compile
Number be lkSelf-encoding encoder model output layer outputThe vector of middle sample serial number i;Wj+1,b'j+1,b”j+1For jth time ginseng
The result that number updates;
(4b2f) judges whether j is equal to 200, if so, number is after the result for taking the 200th subparameter to update is trained
lkSelf-encoding encoder model l'kAnd enter step (4b3);Otherwise return step (4b2e) after adding 1 to j;
(4b3) will be numbered after previous step training as lkSelf-encoding encoder model l'kHidden layer hkIt is l as numberk+1's
The input layer of self-encoding encoder model repeats step (4b2c)-(4b2f) and obtains number lk+1Self-encoding encoder model l after training
'k+1;
(4b4) judges whether k is equal to 2, if the stacking-type self-encoding encoder model Q after then being trained;Conversely, adding 1 to k
Return step (4b3) afterwards;
(4c) builds integrated deep neural network model and is trained:
Integrated deep neural network model is made of L base classifier, after each base classifier is constructed and trained
Obtain integrated deep neural network model.
(4c1) enables k=1;
(4c2) extracts preceding k self-encoding encoder model in the stacking-type self-encoding encoder model Q after training, is only included k
The stacking-type self-encoding encoder model N of a self-encoding encoder modelk, in model NkThe last one self-encoding encoder model hidden layer hk
A Softmax classifier m is connected belowk, and by model NkWith Softmax classifier mkAs integrated deep neural network mould
The l of typekA base classifier;
(4c3) trains lkA base classifier:
Function h is assumed in definitionθ(X(i)) expression formula is as follows:
Wherein:TIt is operated for transposition,For the value of class label,Value for class label isWhen number
Belong to such probability value according to sample;
Using training dataset x_label as lkThe input of a base classifier, using BP algorithm to the base classifier into
Row training, then obtains self-encoding encoder model and the parameter θ by fine tuning, to complete lkThe training of a base classifier;
(4c4) judges whether k is equal to 3, if so, completing the training to 3 base classifiers to get integrated to after training
Deep neural network model, as shown in Figure 2;Conversely, return step (4c2) after adding 1 to k;
Step 5: carrying out data prediction to the new real time data for flowing into wind-driven generator, obtains pretreated real-time
Data data, it is specific as follows
The abnormal point in real time data that (5a) removal newly flows into, obtains the real time data for being no different constant value;
(5b) carries out nondimensionalization processing to the real time data for being no different constant value, obtains pretreated real time data
Data:
Data={ (x'i,y'i), wherein x' ∈ Ru, i ∈ { 0,1,2 ... ..., n'-1, n'} },
Wherein, n' is the number of samples of data set data, and u is the dimension that each sample corresponds to space vector;This example is total
There are 187521 data set data samples, the dimension that each sample corresponds to space vector is 17 dimensions;
Step 6: by the integrated deep neural network model after pretreated real time data data input training
In, it is diagnosed by blade icing condition of the model to wind-driven generator, with the first data of in December, 2015 data
For, specific judgment step is as follows:
(6a) sets the blade state of wind-driven generator as r and r ∈ { 0,1 }, and definition status 0 is icing condition, state 1 is not
Icing condition;
Real time data data is inputted integrated deep neural network model by (6b), obtains the output h of each base classifierθ, λ
(X(i))=[Pλ(y'i=r | x'i;θ)], wherein λ be base classifier number, λ=1 in the present embodiment, 2,3, pλ(y'i=r |
x'i) indicate that i-th of sample of real time data belongs to the probability of state r;
(6c) takes λ=1,2,3 to obtain p respectively1,p2,p3, i.e. 3 respective outputs of base classifier;
(6d) finally judges the output result of all base classifiers using ballot method:
(6d1) counts p1(y'i=0 | x'i),p2(y'i=0 | x'i),p3(y'i=0 | x'i) in numerical value greater than 0.5
Number, and its number is denoted as s, the s=3 for the data;
(6d2) judge s withSize relation, ifThen it is determined as that the blade of wind-driven generator is in icing condition,
Otherwise it determines that it is and does not freeze normally.For the data, due to 3 > 1.5, therefore determine that blade is in icing condition.
Effect of the invention can be further illustrated by following simulation result:
1. simulated conditions and emulation content:
It is Intel (R) Core (TM) i5-7500 3.40GHz, memory 8G, WINDOWS7 operating system in central processing unit
On, it is emulated with MATLAB R2014b software.The wind-driven generator related data that in December, 2015 acquires is input to mould
In type, and model parameter, accordingly set value is input in model, obtains the output result of data sample simultaneously
It is analyzed.
2. interpretation of result:
The present invention is applied to be evaluated and tested using the accuracy rate and MCC coefficient that diagnose to blade icing.Due to wind-driven generator
The real time data newly flowed into is unbalanced dataset, and MCC coefficient is suitable for the evaluation of result of unbalanced dataset.Therefore using quasi-
True rate and MCC coefficient jointly assess result, accuracy with higher.Accuracy rate and MCC coefficient can be by obscuring square
Battle array, which calculates, to be obtained, and confusion matrix is as shown in table 2.
Table 2
Accuracy rate and MCC coefficient formulas are
Performance of the invention, specific comparative experiments are verified using multiple groups comparative experiments are as follows:
First group, blade of wind-driven generator icing is diagnosed using stacking-type self-encoding encoder model;
Second group, blade of wind-driven generator icing is diagnosed using random forest and support vector machines.
Input data used in training pattern is the operation data of the present embodiment acquisition, input number used in assessment models
According to being real time data that the present embodiment newly flows into, experimental result is as shown in table 3.
Table 3
Algorithm title | Accuracy rate | MCC coefficient |
Integrated deep neural network | 0.9511 | 0.7335 |
Stacking-type self-encoding encoder | 0.9112 | 0.6022 |
Random forest | 0.9229 | 0.3838 |
Support vector machines | 0.9251 | 0.4568 |
Analytical table 4 is carried out by accuracy rate result it is found that being frozen using integrated deep neural network to blade of wind-driven generator
Diagnosis has highest accuracy.By MCC coefficient results it is found that using integrated deep neural network to blade of wind-driven generator
Icing is diagnosed, and has highest MCC coefficient value, and compared to other algorithms, MCC coefficient value is greatly improved,
Show that the present invention freezes diagnosis performance well to blade of wind-driven generator, accuracy with higher and generalization ability.
Unspecified part of the present invention belongs to common sense well known to those skilled in the art.
Above description is only example of the present invention, does not constitute any limitation of the invention.Obviously for
It, all may be without departing substantially from the principle of the invention, structure after having understood the content of present invention and principle for one of skill in the art
In the case where, carry out various modifications and variations in form and details, but these modifications and variations based on inventive concept
Still within the scope of the claims of the present invention.
Claims (6)
1. a kind of blade of wind-driven generator icing diagnostic method based on integrated deep neural network, it is characterised in that including as follows
Step:
(1) historical data of acquisition wind-driven generator operation;
(2) data prediction is carried out to collected historical data, obtains pretreated historical data;
(3) data balancing processing is carried out to pretreated historical data, obtains training dataset x_label;
(4) it constructs integrated deep neural network and it is trained using training dataset x_label, the collection after obtaining training
It is specific as follows at deep neural network model:
(4a) builds the basic structure of stacking-type self-encoding encoder model:
Set the number L of stacking-type self-encoding encoder model, the input layer number u of each self-encoding encoderlAnd the number of hidden nodes
vl, wherein l=1,2 ..., L, L >=3 and be odd number, the number of k-th of stacking-type self-encoding encoder model is lk, and k=1,
2,...,L;The output layer number of nodes u of first self-encoding encoder1It is self-editing after first equal to the Characteristic Number of training dataset
The input layer number u of code devicetWith the node in hidden layer v of its previous self-encoding encodert-1It is equal, wherein t=2,3 ..., L;
(4b) trains stacking-type self-encoding encoder model:
(4b1) enables k=1;
(4b2) training number is lkSelf-encoding encoder model:
(4b2a) initializes weight matrix W, learning rate η and offset vector b', b ", and sets maximum cycle as I;
(4b2b) removes the class label of training dataset x_label, obtains no label training dataset x:
X={ xi, wherein x ∈ Ru, i ∈ { 0,1,2 ... ..., n-1, n } },
Wherein, n is the number of samples of training dataset, and i is sample serial number, and u is the dimension that each sample corresponds to space vector;
Number l will be used as without label training dataset xkSelf-encoding encoder model input;
It is l that number, which is calculated, in (4b2c) according to the following formulakSelf-encoding encoder model hidden layer output yk:
It is l that number is calculated according to the following formulakSelf-encoding encoder model output layer output
Wherein, WTFor the transposition of weight matrix W;
(4b2d) enables j=1;
(4b2e) is according to formula Undated parameter, wherein For number
For lkSelf-encoding encoder model output layer outputThe vector of middle sample serial number i;Wj+1,b'j+1,b”j+1For jth subparameter
The result of update;
(4b2f) judges whether j is equal to I, if so, number is l after the result for taking I subparameter to update is trainedkIt is self-editing
Code device model l'kAnd enter step (4b3);Otherwise return step (4b2e) after adding 1 to j;
(4b3) will be numbered after previous step training as lkSelf-encoding encoder model l'kHidden layer hkIt is l as numberk+1It is self-editing
The input layer of code device model repeats step (4b2c)-(4b2f) and obtains number lk+1Self-encoding encoder model l' after trainingk+1;
(4b4) judges whether k is equal to L-1, if the stacking-type self-encoding encoder model Q after then being trained;Conversely, after adding 1 to k
Return step (4b3);
(4c) builds integrated deep neural network model and is trained:
(4c1) enables k=1;
(4c2) extracts preceding k self-encoding encoder model in the stacking-type self-encoding encoder model Q after training, is only included k a certainly
The stacking-type self-encoding encoder model N of encoder modelk, in model NkThe last one self-encoding encoder model hidden layer hkBelow
Connect a Softmax classifier mk, and by model NkWith Softmax classifier mkAs integrated deep neural network model
LkA base classifier;
(4c3) trains lkA base classifier:
Function h is assumed in definitionθ(X(i)) expression formula is as follows:
Wherein:TIt is operated for transposition,For the value of class label,Value for class label isWhen data sample
Originally belong to such probability value;
Using training dataset x_label as lkThe input of a base classifier instructs the base classifier using BP algorithm
Practice, then obtain self-encoding encoder model and the parameter θ by fine tuning, to complete lkThe training of a base classifier;
(4c4) judges whether k is equal to L, if so, completing the training to L base classifier to get the integrated depth to after training
Neural network model;Conversely, return step (4c2) after adding 1 to k;
(5) data prediction is carried out to the new real time data for flowing into wind-driven generator, obtains pretreated real time data data;
(6) by by pretreated real time data data input training after integrated deep neural network model in, by this
Model diagnoses the blade icing condition of wind-driven generator.
2. a kind of blade of wind-driven generator icing diagnostic method based on integrated deep learning according to claim 1,
It is characterized in that: the historical data of the operation of wind-driven generator described in step (1), the SCADA data including wind-driven generator operation
With blade icing daily record data.
3. a kind of blade of wind-driven generator icing diagnostic method based on integrated deep learning according to claim 1,
Be characterized in that: the data preprocessing operation in step (2) includes the following steps:
(2a) removes the abnormal point in collected historical data, obtains the historical data for being no different constant value;
(2b) carries out nondimensionalization processing to the historical data for being no different constant value, obtains pretreated historical data.
4. a kind of blade of wind-driven generator icing diagnostic method based on integrated deep learning according to claim 1,
Be characterized in that: the processing of data balancing described in step (3) is realized using SMOTE algorithm.
5. a kind of blade of wind-driven generator icing diagnostic method based on integrated deep learning according to claim 1,
Be characterized in that: the data preprocessing operation in step (5) includes the following steps:
The abnormal point in real time data that (5a) removal newly flows into, obtains the real time data for being no different constant value;
(5b) carries out nondimensionalization processing to the real time data for being no different constant value, obtains pretreated real time data data:
Data={ (x'i,y'i), wherein x' ∈ Ru, i ∈ { 0,1,2 ... ..., n'-1, n'} },
Wherein, n' is the number of samples of data set data, and u is the dimension that each sample corresponds to space vector.
6. a kind of blade of wind-driven generator icing diagnostic method based on integrated deep learning according to claim 1,
It is characterized in that: the blade icing condition of wind-driven generator being diagnosed by integrated deep neural network model in step (6)
Specific step is as follows:
(6a) sets the blade state of wind-driven generator as r and r ∈ { 0,1 }, and definition status 0 is icing condition, state 1 is not freeze
State;
Real time data data is inputted integrated deep neural network model by (6b), obtains the output h of each base classifierθ, λ(X(i))
=[Pλ(y'i=r | x'i;θ)], wherein λ be base classifier number, λ=1,2 ..., L, pλ(y'i=r | x'i) indicate number in real time
According to i-th of sample belong to the probability of state r;
(6c) takes λ=1,2 respectively ..., L obtains p1,p2,…,pL, i.e. the L respective output of base classifier;
(6d) finally judges the output result of all base classifiers using ballot method:
(6d1) counts p1(y'i=0 | x'i),p2(y'i=0 | x'i),…,pL(y'i=0 | x'i) in numerical value be greater than 0.5 number,
And its number is denoted as s;
(6d2) ifThen it is determined as that the blade of wind-driven generator is in icing condition, otherwise determines that it is and do not freeze normally.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811468877.8A CN109376801B (en) | 2018-12-04 | 2018-12-04 | Wind driven generator blade icing diagnosis method based on integrated deep neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811468877.8A CN109376801B (en) | 2018-12-04 | 2018-12-04 | Wind driven generator blade icing diagnosis method based on integrated deep neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109376801A true CN109376801A (en) | 2019-02-22 |
CN109376801B CN109376801B (en) | 2021-07-20 |
Family
ID=65375373
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811468877.8A Active CN109376801B (en) | 2018-12-04 | 2018-12-04 | Wind driven generator blade icing diagnosis method based on integrated deep neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109376801B (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110162888A (en) * | 2019-05-24 | 2019-08-23 | 北京天泽智云科技有限公司 | A method of the fan blade based on semi-supervised learning, which freezes, to be detected |
CN110609229A (en) * | 2019-09-17 | 2019-12-24 | 电子科技大学 | Wind driven generator blade imbalance fault detection method based on deep learning |
CN111144499A (en) * | 2019-12-27 | 2020-05-12 | 北京工业大学 | Fan blade early icing fault detection method based on deep neural network |
CN111737923A (en) * | 2020-07-28 | 2020-10-02 | 湘电风能有限公司 | Fan blade icing prediction method |
CN111860658A (en) * | 2020-07-24 | 2020-10-30 | 华北电力大学(保定) | Transformer fault diagnosis method based on cost sensitivity and integrated learning |
CN111988237A (en) * | 2020-07-31 | 2020-11-24 | 中移(杭州)信息技术有限公司 | Flow identification method and device, electronic equipment and storage medium |
CN112215281A (en) * | 2020-10-12 | 2021-01-12 | 浙江大学 | Fan blade icing fault detection method |
CN112729783A (en) * | 2020-12-04 | 2021-04-30 | 西人马联合测控(泉州)科技有限公司 | Fan blade fault diagnosis method, device, equipment and computer storage medium |
CN112832960A (en) * | 2020-12-23 | 2021-05-25 | 国电南瑞南京控制系统有限公司 | Fan blade icing detection method based on deep learning and storage medium |
CN112949744A (en) * | 2021-03-23 | 2021-06-11 | 华中科技大学无锡研究院 | Blade icing detection method and system for wind driven generator and storage medium |
CN113719425A (en) * | 2021-05-14 | 2021-11-30 | 北京智慧空间科技有限责任公司 | Fan blade fault prediction method and system and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103323772A (en) * | 2012-03-21 | 2013-09-25 | 北京光耀能源技术股份有限公司 | Wind driven generator operation state analyzing method based on neural network model |
CN103982378A (en) * | 2014-04-25 | 2014-08-13 | 广东电网公司电力科学研究院 | Method for diagnosing surface icing faults of wind power generator blade of power system based on machine visual images |
CN105464912A (en) * | 2016-01-27 | 2016-04-06 | 国电联合动力技术有限公司 | Method and device for detecting freezing of wind generating set blades |
US9911290B1 (en) * | 2015-07-25 | 2018-03-06 | Gary M. Zalewski | Wireless coded communication (WCC) devices for tracking retail interactions with goods and association to user accounts |
EP3361412A1 (en) * | 2017-02-10 | 2018-08-15 | Fujitsu Limited | Black ice detection system, program, and method |
-
2018
- 2018-12-04 CN CN201811468877.8A patent/CN109376801B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103323772A (en) * | 2012-03-21 | 2013-09-25 | 北京光耀能源技术股份有限公司 | Wind driven generator operation state analyzing method based on neural network model |
CN103982378A (en) * | 2014-04-25 | 2014-08-13 | 广东电网公司电力科学研究院 | Method for diagnosing surface icing faults of wind power generator blade of power system based on machine visual images |
US9911290B1 (en) * | 2015-07-25 | 2018-03-06 | Gary M. Zalewski | Wireless coded communication (WCC) devices for tracking retail interactions with goods and association to user accounts |
CN105464912A (en) * | 2016-01-27 | 2016-04-06 | 国电联合动力技术有限公司 | Method and device for detecting freezing of wind generating set blades |
EP3361412A1 (en) * | 2017-02-10 | 2018-08-15 | Fujitsu Limited | Black ice detection system, program, and method |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110162888B (en) * | 2019-05-24 | 2022-12-09 | 北京天泽智云科技有限公司 | Fan blade icing detection method based on semi-supervised learning |
CN110162888A (en) * | 2019-05-24 | 2019-08-23 | 北京天泽智云科技有限公司 | A method of the fan blade based on semi-supervised learning, which freezes, to be detected |
CN110609229A (en) * | 2019-09-17 | 2019-12-24 | 电子科技大学 | Wind driven generator blade imbalance fault detection method based on deep learning |
CN111144499A (en) * | 2019-12-27 | 2020-05-12 | 北京工业大学 | Fan blade early icing fault detection method based on deep neural network |
CN111860658A (en) * | 2020-07-24 | 2020-10-30 | 华北电力大学(保定) | Transformer fault diagnosis method based on cost sensitivity and integrated learning |
CN111737923A (en) * | 2020-07-28 | 2020-10-02 | 湘电风能有限公司 | Fan blade icing prediction method |
CN111988237A (en) * | 2020-07-31 | 2020-11-24 | 中移(杭州)信息技术有限公司 | Flow identification method and device, electronic equipment and storage medium |
CN112215281A (en) * | 2020-10-12 | 2021-01-12 | 浙江大学 | Fan blade icing fault detection method |
CN112729783A (en) * | 2020-12-04 | 2021-04-30 | 西人马联合测控(泉州)科技有限公司 | Fan blade fault diagnosis method, device, equipment and computer storage medium |
CN112832960A (en) * | 2020-12-23 | 2021-05-25 | 国电南瑞南京控制系统有限公司 | Fan blade icing detection method based on deep learning and storage medium |
CN112949744A (en) * | 2021-03-23 | 2021-06-11 | 华中科技大学无锡研究院 | Blade icing detection method and system for wind driven generator and storage medium |
CN112949744B (en) * | 2021-03-23 | 2023-09-05 | 元始智能科技(南通)有限公司 | Wind driven generator blade icing detection method, system and storage medium |
CN113719425A (en) * | 2021-05-14 | 2021-11-30 | 北京智慧空间科技有限责任公司 | Fan blade fault prediction method and system and storage medium |
CN113719425B (en) * | 2021-05-14 | 2022-11-22 | 北京智慧空间科技有限责任公司 | Fan blade fault prediction method and system and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109376801B (en) | 2021-07-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109376801A (en) | Blade of wind-driven generator icing diagnostic method based on integrated deep neural network | |
CN106777984B (en) | A method of photovoltaic array Working state analysis and fault diagnosis are realized based on density clustering algorithm | |
CN102944418B (en) | Wind turbine generator group blade fault diagnosis method | |
CN111753893A (en) | Wind turbine generator power cluster prediction method based on clustering and deep learning | |
CN109751206A (en) | fan blade icing failure prediction method, device and storage medium | |
CN107392304A (en) | A kind of Wind turbines disorder data recognition method and device | |
CN105891546A (en) | Wind vane fault diagnosis method in wind turbine yaw system based on big data | |
CN110609229B (en) | Wind driven generator blade imbalance fault detection method based on deep learning | |
CN109522599A (en) | Transmission line of electricity catastrophic failure method for early warning caused by a kind of typhoon | |
CN107133632A (en) | A kind of wind power equipment fault diagnosis method and system | |
CN109299507A (en) | A kind of different degrees of Analysis on Fault Diagnosis method under the coal pulverizer multi-state based on dynamic modeling | |
CN104182816A (en) | Method for evaluating power quality comprehensively based on the Vague sets and the improved technique for order preference by similarity to ideal solution and application thereof | |
CN105279612A (en) | Poisson distribution-based power transmission line tripping risk assessment method | |
CN105065212A (en) | Checking method and system of wind generation sets of wind power plant | |
CN109307854A (en) | A kind of electric energy metering device comprehensive estimation method and system | |
CN106650977A (en) | Short-term power prediction method used for newly-built wind farm | |
CN106768933A (en) | A kind of wind power generation unit blade method for diagnosing faults based on genetic algorithm | |
CN104102951A (en) | Short-term wind power prediction method based on EMD (Empirical Mode Decomposition) historical data preprocessing | |
CN110378549A (en) | A kind of transmission tower bird pest grade appraisal procedure based on FAHP- entropy assessment | |
CN110738253A (en) | short-term wind power prediction method based on FCM and AFSA-Elman | |
CN116306139A (en) | Intelligent monitoring method and system for service life of wind turbine blade | |
CN109595130A (en) | A kind of fan blade icing failure prediction method and system | |
CN113689053B (en) | Strong convection weather overhead line power failure prediction method based on random forest | |
CN111950131B (en) | Wind power plant output equivalent aggregation model construction method considering electricity limiting factors | |
CN111047121B (en) | Fan fault prediction method and system |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |