CN111237134A

CN111237134A - Offshore double-fed wind driven generator fault diagnosis method based on GRA-LSTM-stacking model

Info

Publication number: CN111237134A
Application number: CN202010038529.8A
Authority: CN
Inventors: 魏书荣; 张鑫; 符杨; 任子旭; 缪舒馨; 闫梦飞
Original assignee: Shanghai Electric Power University
Current assignee: Shanghai University of Electric Power; Shanghai Electric Power University; University of Shanghai for Science and Technology
Priority date: 2020-01-14
Filing date: 2020-01-14
Publication date: 2020-06-05
Anticipated expiration: 2040-01-14
Also published as: CN111237134B

Abstract

The invention relates to a fault diagnosis method for an offshore double-fed wind driven generator based on a GRA-LSTM-stacking model, which comprises the following steps of: step 1: analyzing the SCADA state variable through GRA, and screening out the state variable related to the temperature of the generator as the input of the LSTM network; step 2: predicting the generator temperature in a normal state through the LSTM to obtain a predicted value; and step 3: calculating the absolute value of the residual error between the actual value and the predicted value, setting an alarm threshold value by using a statistical method, and accordingly identifying the early fault of the generator and extracting a fault sample; and 4, step 4: and (4) performing data processing on the extracted fault sample through a stacking fusion algorithm and outputting a final accurate diagnosis result. Compared with the prior art, the method has the advantages of accurate fault diagnosis, strong universality, simple acquisition of fault samples and the like.

Description

Offshore double-fed wind driven generator fault diagnosis method based on GRA-LSTM-stacking model

Technical Field

The invention relates to the field of offshore wind turbine generator fault diagnosis, in particular to a fault diagnosis method for an offshore double-fed wind turbine generator based on a GRA-LSTM-stacking model.

Background

Offshore wind power has become the focus of global renewable energy development. With the rapid development of offshore wind power technology, all offshore wind power strong countries in Europe begin to advance to large-scale and deep seas. The offshore wind power plant distance of the Dogger Bank currently planned in the UK is 123-289km, and the total installed capacity reaches 4.8 GW. In Germany, 13 offshore wind farms with a total installed capacity of 21.3GW are planned within 370.4km outside the territory. China offshore wind farms are also advancing towards large-scale, deep-offshore. The doubly-fed asynchronous generator, as one of the mainstream models of the offshore wind turbine, faces the problems of severe operating environment, poor accessibility, high failure rate and the like. Statistically, generator failure is one of the major factors causing wind turbine shutdown, accounting for 37% of all downtime. In order to reduce the great economic loss caused by the fault shutdown of the generator, the fault of the offshore doubly-fed wind generator needs to be accurately diagnosed urgently.

China offshore wind power development has a decade history, a large-batch SCADA (Supervisory Control And Data Acquisition) system of the wind turbine can acquire massive operation Data of the whole wind power plant And perform remote or local monitoring, And the massive time sequence Data records the real-time operation condition of the offshore wind turbine, so that the system has the characteristics of quick signal change And numerous operation parameters. The fault characteristics of the wind turbine generator are hidden in the SCADA variable which can represent the operation state of the wind turbine generator, so that how to fully apply the SCADA data to identify, diagnose and early warn the state of the wind turbine generator becomes a research hotspot in the wind turbine field and is widely concerned by scholars at home and abroad. The existing offshore wind turbine generator fault diagnosis research is developed from a traditional mathematical method to an artificial intelligence direction, for example, the method of a support vector machine, an artificial neural network, a fault tree and the like is adopted to carry out fault diagnosis on important parts of the wind turbine generator, a gearbox and the like. The diagnosis methods have shallow learning level and single structure on data, and the diagnosis precision needs to be further improved.

In recent years, artificial intelligence and deep learning are rapidly developed, and a brand new thought is provided for fault diagnosis of the wind turbine generator. If the wind turbine generator fault is diagnosed by adopting methods such as a convolutional neural network, a long and short memory network and a stacking self-coding method, the result shows that the fault diagnosis method based on deep learning has higher accuracy and generalization compared with a shallow learning method. The deep learning algorithm is mainly used for fault diagnosis by training with known fault samples, so that the purpose of fault identification is achieved. However, if the warning time of the SCADA system is not enough, the collected fault samples are not enough, and the fault diagnosis cannot be performed. In addition, when the wind driven generator has an early slight fault, the monitored state quantity usually does not exceed a system threshold value, and at the moment, the SCADA system cannot give effective early warning, so that a fault sample cannot be obtained.

Disclosure of Invention

The invention aims to provide a fault diagnosis method of an offshore double-fed wind driven generator based on a GRA-LSTM-stacking model aiming at the special operating conditions of an offshore wind turbine generator and the current situation and existing problems of fault diagnosis based on SCADA data. And then, carrying out accurate fault diagnosis on the fault sample with the threshold out-of-limit through a stacking fusion algorithm. And finally, verifying the effectiveness of the method provided by the invention through actual data of a certain offshore wind farm.

The purpose of the invention can be realized by the following technical scheme:

a fault diagnosis method for an offshore doubly-fed wind generator based on a GRA-LSTM-stacking model comprises the following steps:

step 1: analyzing the SCADA state variable through GRA, and screening out the state variable related to the temperature of the generator as the input of the LSTM network;

step 2: predicting the generator temperature in a normal state through the LSTM to obtain a predicted value;

and step 3: calculating the absolute value of the residual error between the actual value and the predicted value, setting an alarm threshold value by using a statistical method, and accordingly identifying the early fault of the generator and extracting a fault sample;

and 4, step 4: and (4) performing data processing on the extracted fault sample through a stacking fusion algorithm and outputting a final accurate diagnosis result.

Further, the step 1 comprises the following sub-steps:

step 11: extracting the normal operation state data of the fan acquired by the SCADA system, and carrying out normalization processing on the data;

step 12: calculating grey correlation coefficients of the generator temperature and each state variable;

step 13: optimizing the weight by using an entropy weight theory, and calculating the grey correlation degree of each state variable and the temperature of the generator;

step 14: the grey relevance is ranked and the relevant state variables are selected as inputs to the LSTM network.

Further, the gray correlation coefficient in step 12 is calculated by the following formula:

in the formula, ξ_ijFor grey correlation coefficient, ρ is resolution, Δ x is the amount of change between the state variable as reference sequence and the generator temperature as comparison sequence, i is 1,2, …, m, j is 1,2, …, n, m is the number of sample samples, n is the number of state variables.

Further, the gray correlation degree in step 13 is calculated by the following formula:

in the formula, r_jThe degree of association is a gray color,

the entropy weight theory is used for optimization, namely the value is 1.

Further, the step 2 specifically includes: and taking the state variable extracted by the GRA model as input and the generator temperature as output, training the LSTM network, and iteratively updating the weight and the offset to minimize the error to obtain a generator temperature prediction model and output a generator temperature prediction value.

Further, the step 3 comprises the following sub-steps:

step 31: acquiring a residual absolute value of an actual value and a predicted value of the temperature of the generator;

step 32: setting a residual absolute value alarm threshold value by adopting a statistical method;

step 33: therefore, early faults of the generator are identified, and once the absolute value of the residual error of the generator temperature exceeds an alarm threshold value, state data after alarm is extracted and used as a fault sample for next accurate fault diagnosis.

Further, the alarm threshold in step 32 is calculated by the following formula:

where f (Re) is the probability density function, Re is the absolute value of the residual, T is the alarm threshold, and α is the significance level.

Further, the step 4 specifically includes: and 3, carrying out data set division on the fault sample extracted in the step 3, training a first-layer base learner of a stacking fusion algorithm by adopting K-fold cross validation, outputting probability output serving as a diagnosis result, combining the probability output by the first-layer base learner and serving as a new data set, training the probability output by the first-layer base learner and serving as the input of a second-layer element learner, and outputting a final accurate diagnosis result.

Further, the relevant algorithm model selected in the first-layer base learner comprises: the method comprises the following steps of supporting a vector machine, a K nearest classification algorithm, a random forest integrated in a bagging mode, a gradient boosting decision tree and an extreme gradient boosting tree integrated in a boosting mode, wherein after each base learner in the first layer of base learners is independently diagnosed, the diagnosis result is further subjected to the sperman relevance analysis, and the calculation formula of the sperman relevance analysis is as follows:

in the formula, ρ_xyFor the spearman correlation analysis corresponding to the outcome measures,

and

are the average values of the respective elements, x_iAnd y_iRespectively, the actual diagnostic value of each vector element.

Further, the relevant algorithm model selected in the second-level meta learner includes GBDT.

Compared with the prior art, the invention has the following advantages:

(1) the method can effectively identify the early fault of the generator and accurately diagnose the specific fault type, solves the problems of insufficient warning time of the SCADA system and difficulty in obtaining fault samples, and effectively avoids huge economic loss caused by fault deterioration.

(2) According to the invention, the stacking fusion algorithm is applied to the field of fault diagnosis for the first time, and different algorithms are fully utilized to deeply mine the SCADA data from multiple angles, so that various algorithms can make up for deficiencies, and the fault diagnosis precision is effectively improved.

(3) The method has universality and can be popularized and applied to other wind turbines provided with SCADA systems.

Drawings

FIG. 1 is a flow chart of the wind turbine generator fault diagnosis of the present invention;

FIG. 2 is a stacking fusion algorithm framework in the present invention;

FIG. 3 is a graph comparing the results of GRA-LSTM and other models;

FIG. 4 is a graph comparing the results of the GRA-LSTM and other models;

FIG. 5 is a diagram showing changes in absolute values of residuals of a fault F1 in three generator fault cases;

FIG. 6 is a diagram showing changes in absolute values of residuals of a fault F2 in three generator fault cases;

FIG. 7 is a diagram showing changes in absolute values of residuals of a fault F3 in three generator fault cases;

FIG. 8 is a diagram showing the correlation analysis result of the stacking diagnostic model.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.

The technical scheme of the invention is shown in figure 1:

a fault diagnosis method for an offshore double-fed wind driven generator based on a GRA-LSTM-stacking model comprises the following steps:

(1) the GRA state quantity extraction process comprises the following steps: firstly, extracting the normal operation state data of the fan collected by an SCADA system, and carrying out normalization processing on the data; secondly, calculating grey correlation coefficients of the generator temperature and each state variable; then, optimizing the weight by using an entropy weight theory, and calculating the grey correlation degree of each state variable and the generator temperature; finally, the grey relevance degrees are sorted, and highly relevant state variables are selected as the input of the LSTM network.

(2) LSTM temperature prediction process: the state variables extracted by the GRA model are used as input, and the generator temperature is used as output. And then training the LSTM network, and updating the weight and the offset by iteration to minimize the error, thereby obtaining a generator temperature prediction model and outputting a generator temperature prediction value.

(3) An early fault identification process: firstly, calculating the residual absolute value of the actual value and the predicted value of the temperature of the generator. Then, a residual absolute value alarm threshold is set by adopting a statistical method. And once the absolute value of the residual error of the temperature of the generator exceeds an alarm threshold value, indicating that the generator has a fault, and extracting state data after alarm as a data set for next accurate fault diagnosis.

(4) And (3) a stacking accurate fault diagnosis process: and dividing a data set according to the state data of the generator fault time period intercepted in the last step. And training a first-layer base learner of a stacking fusion algorithm by adopting K-fold cross validation, and outputting a diagnosis result (probability output). And combining the probabilities output by the first layer into a new data set, and training the new data set as the input of the second-layer meta learner to output the final accurate diagnosis result.

The specific process of the step 4 is as follows:

as an innovation and a key point of the method, a stacking fusion algorithm framework is designed to realize accurate diagnosis of multiple faults of the wind driven generator. The stacking fusion algorithm is one of ensemble learning, and is different from a bagging algorithm and a boosting algorithm which are fused with a plurality of same algorithms, and combines machine learning algorithms which are not used in a certain fusion mode to obtain excellent performance which cannot be obtained by a single machine learning algorithm.

The training process of the first-layer diagnosis model of the stacking fusion algorithm adopts K-fold cross validation, and the specific training mode is as follows: dividing the labeled data set S into K sub-data sets S randomly₁,S₂,…,S_K}. Taking the base learner 1 as an example, each subdata set is respectively made into a primary verification set, the rest K-1 subdata sets are used as training sets, diagnosis results (probability output) under K models are obtained, and a set L is formed₁，L₁The length is the same as S. To other n-1 radicalsThe learner performs the same operation to obtain the set L₂,L₃,…,L_nCombining the diagnostic results of the n basis learners into a new data set L ═ L₁,L₂,…,L_n}. The new data set L is the input data for the second layer diagnostic model meta-learner. The second layer of diagnostic algorithm can find and correct errors in the first layer of diagnostic model in time, and the purpose of improving the accuracy of the diagnostic model is achieved.

The stacking fusion algorithm integrates diversified algorithms, and can fully utilize different algorithms to analyze data from multiple angles. Based on this, the first layer diagnostic model based learner selects not only the algorithm with superior performance, but also a different type of algorithm. The second layer diagnosis model meta-learner selects an algorithm with strong generalization capability, which is beneficial to correcting the first layer diagnosis error and achieving the optimal diagnosis effect. In the method, a first-layer base learner initially selects a Support Vector Machine (SVM), a K-Nearest Neighbor classification algorithm (KNN), a Random Forest (RF) and a Gradient Boost Decision Tree (GBDT) which are integrated in a bagging mode, and an eXtreme Gradient boost Tree (XGboost) which is integrated in a Boosting mode. The second layer selects GBDT as the meta-learner. In order to select the optimal base learner as the first-layer diagnosis model, the method provided by the invention is used for designing an experiment, each base learner is diagnosed independently, and the diagnosis result obtained by the diagnosis is subjected to the spearman correlation analysis. The higher the degree of correlation between algorithms, the higher the diagnosis precision after fusion. Wherein, the spaerman calculation formula is as follows:

and

are respectively in each directionAverage value of quantity elements, x_iAnd y_iRespectively, the actual diagnostic value of each vector element.

The invention is described in detail below with reference to the figures and specific embodiments.

The specific implementation method of the invention can be roughly divided into the following steps: firstly, analyzing the SCADA state variable through GRA, and screening out the state variable highly related to the temperature of the generator; then, predicting the generator temperature in a normal state through LSTM, calculating the absolute value of the residual error between the actual value and the predicted value, setting an alarm threshold value by using a statistical method, and accordingly identifying the early-stage fault of the generator and extracting a fault sample; and finally, carrying out accurate fault diagnosis on the extracted three generator fault samples through a stacking fusion algorithm.

And extracting variables highly related to the temperature of the generator in the SCADA state variables, and qualitatively reflecting the weight of the influence of each state variable on the temperature of the generator by using the GRA value. GRA's thinking and calculation formula are as follows:

1) setting the state variable as a reference sequence, and recording as:

X₀＝{x₀(1),x₀(2),…,x₀(n)}

2) setting the generator temperature as a comparison sequence, and recording as follows:

X_i＝{x_i(1),x_i(2),…,x_i(n)}

in the above two formulas, i is 1,2, …, m; m is the number of sampling samples; n is the number of state variables.

3) And calculating the correlation coefficient. The correlation coefficient of the reference sequence and the ith comparison sequence at the point j is as follows:

in the formula, ξ_ijFor grey correlation coefficients, ρ is the resolution, the value is between 0 and 1, in order to reduce the influence of the pole on the calculation, 0.5 is generally taken in the fault diagnosis, Δ x is the change of the generator temperature as a reference sequence and as a comparison sequenceQuantity, Δ x ═ x₀(j)-x_i(j) I is 1,2, …, m, j is 1,2, …, n, m is the number of sampling samples, n is the number of state variables.

4) And calculating the grey correlation degree.

In the formula, r_jFor grey correlation, in order to fully consider the influence weight of the comparison sequence on the reference sequence, the invention uses the entropy weight theory to carry out optimization, namely

The entropy weight theory is used for optimization, namely the value is 1.

The generator temperature is predicted using the LSTM network. The key of the LSTM network is the cell tuple with memory function, which selectively updates the cell state information by controlling the 3-gate structure. The specific calculation process of the LSTM network is as follows:

1) forget the door state f_t: inputting x from the current layer_tAnd the output h of the previous layer_t-1And (4) jointly determining.

f_t＝σ(W₁ ^fgx_t+W_h ^fgh_t-1+b_f)

2) Input door status i_t: inputting x from the current layer as in the forgetting gate_tAnd the output h of the previous layer_t-1And (4) jointly determining. However, the input gate adds one step cell state C_tAnd (6) updating.

i_t＝σ(W₁ ⁱgx_t+W_h ⁱgh_t-1+b_i)

3) Output gate state o_t: like the input gate, x is input from the current layer_tAnd the output h of the previous layer_t-1Jointly determining, adding one step output h_tAnd (6) updating.

o_t＝σ(W₁ ^ogx_t+W_h ^ogh_t-1+b_o)

h_t＝o_t×tanh(C_t)

In the above formulas, W₁ ^f、W₁ ⁱ、W₁ ^C、W₁ ^oAre respectively the current input x_tThe connection weights of the forgetting gate, the input gate and the cell tuple input and output gate of the previous layer; w_h ^f、W_h ⁱ、W_h ^C、W_h ^oRespectively outputs h for the previous layer_t-1The connection weights of the forgetting gate, the input gate and the cell tuple input and output gate of the previous layer; b_f、b_i、b_C、b_oThe offset of the forgetting gate, the input gate and the output gate of the last layer of cell tuple are respectively; sigma is a nonlinear sigmoid activation function.

When the wind turbine generator normally operates, the absolute value of the residual error between the actual temperature value and the predicted value of the generator is very small. Once an abnormal situation occurs, the absolute value of the residual deviates from the normal level. As the degree of the fault increases, the absolute value of the residual error of the generator temperature gradually increases, and a significant jitter or a climbing condition occurs. Thus, early generator failure may be identified by monitoring whether the absolute value of the generator temperature residual exceeds an alarm threshold.

The method of the invention adopts a statistical method for setting the alarm threshold value. The temperature residual absolute value of the generator in normal operation is fitted to be subjected to certain statistical distribution, and the corresponding probability density function is f (Re), so that the alarm threshold value can be calculated through the following formula.

In the formula, f (Re) is a probability density function, Re is a residual absolute value, T is an alarm threshold value, α is a significance level, the value of the significance level is between 0 and 1, and in order to eliminate the interference of abnormal working conditions, the method selects 0.01.

And according to the state data of the generator fault time period intercepted in the last step, carrying out accurate fault diagnosis on the fault sample with the threshold out of limit through the stacking fusion algorithm framework designed in the previous step.

In order to verify the prediction effect of the LSTM network on the temperature of the generator, the method selects Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Relative Error (MRE) and goodness of fit (R, R) according to the method²) The calculation formula of each evaluation index is as follows:

of the above formulae, Y (i), Y' (i) and

the actual value, the predicted value and the average value of the generator temperature at the ith moment are respectively; 1,2, …, n 1; n1 is the length of the test set time series.

In the prediction model evaluation indexes, RMSE represents the accuracy of the prediction result of the temperature of the generator, and the more accurate the prediction is, the smaller the RMSE is; MAE and MRE represent the consistency of the temperature prediction results of the generator, and the smaller the prediction deviation is, the smaller the MAE and the MRE are; r²Representing the degree of fit of the predicted curve, the better the fitting effect R²The closer to 1.

Meanwhile, in order to evaluate the performance of the stacking diagnosis model, the method adopts the accuracy rate R_TAlarm-missing rate R_FAnd F₁The scores are evaluated. The calculation process of each index is as follows:

in the above formulas, TP represents the number of correctly diagnosed specific faults; FP represents the number of misdiagnoses of other states as a particular fault; FN represents the number of false diagnoses of a particular fault as other states; f₁Is R_TAnd R_FRepresents the overall performance of the classifier.

In the evaluation index of the diagnostic model, R_TAnd F₁The larger the value of (A), R_FThe smaller the value of (a), the better the fault diagnosis model performance.

Practical embodiment

Fig. 1 is a flow chart of a fault diagnosis of a wind power generator according to the present invention, and the present invention is applied to a specific example according to the procedure of the flow chart. Case research is carried out by using data of certain offshore wind power plant in China to verify the feasibility of the method. The wind power plant has 34 fans, and the single machine capacity is 3 MW. Selecting state data collected by an SCADA system of No. 18 wind turbine generator set from 2012 to 2016 every 10min, firstly screening the collected state data, and filtering working condition parameters smaller than cut-in wind speed and larger than cut-out wind speed.

Fig. 2 is a framework diagram of a stacking fusion algorithm, wherein a first layer of the stacking fusion algorithm integrates five algorithms of SVM, KNN, RF, GBDT and XGBoost, and a second layer selects GBDT as a meta-learner. 5-fold cross validation was applied to the initial data set.

Fig. 3 extracts the GRA with 9 state variables X ═ { T ═ T₁,v₁,T₂,I,P,T₃,T₄,v₂,T₅And the LSTM temperature prediction model is used as an input variable of the LSTM temperature prediction model, and the generator temperature is used as an output variable. The training set of the LSTM model is normal operation data of No. 18 engine group from 1 month to 6 months in 2012, and is 17066 groups in total; the verification set is normal operation data of 18 machine set from 7 months to 8 months in 2012, and 5053 sets are calculated. Setting the input dimension of the LSTM model as 9, the output dimension as 1, the number of hidden layer neurons as 15, optimizing by using a random gradient descent Adam algorithm, and taking a loss function as a cross entropy function. An optimal temperature prediction model is obtained through multiple iterations of the training set, the model is used for predicting the verification set, and an obtained prediction curve is shown in fig. 3.

In order to verify and verify that the model has excellent performance in wind turbine temperature prediction, the model is compared and analyzed with the classical time series prediction models BPNN, ARIMA and PCA-LSTM respectively, and the result is shown in FIG. 3, since the horizontal axis time span is large, the comparison effect is not obvious, and the segment (blue dotted frame) in FIG. 3 is extracted, as shown in FIG. 4. The errors of the 4 temperature prediction models are shown in table 1.

TABLE 1 error analysis of prediction results

As can be seen from FIG. 4, the maximum error of the BPNN model for predicting the temperature reaches 3.5 ℃, and the prediction result of the GRA-LSTM model proposed by the present specification is closer to the actual value and is the minimum prediction error of the 4 models. As can be seen from Table 1, the values of RMSE, MAE and MRE of the GRA-LSTM model were reduced by 1.16, 0.72 and 0.01, respectively, on average, compared to the other models; r²The value of (a) is increased by 0.02 on average. Therefore, the model provided by the method is higher in prediction precision, and a foundation is laid for the next early fault identification.

Fig. 5, 6 and 7 are three generator failure cases. And predicting the data of the verification set by using the GRA-LSTM model to obtain a predicted value of the temperature of the generator, and then calculating a residual absolute value Re of the predicted value and the actual value. Then, distribution fitting was performed on Re using a statistical method, and it was found that Re obeyed a log normal distribution. The absolute value of the temperature residual error when the generator normally operates is fitted to obtain mu-6.04741 and sigma-1.19714, and the alarm threshold T in the embodiment of the method is obtained according to a correlation formula and is 0.0516.

And analyzing early fault recognition capability of the GRA-LSTM model by using three generator fault cases of the wind power generation set No. 18 of the wind farm.

1) Failure F1: the 18 # unit stops at 10 months and 4 days 09:55 in 2012 due to the abrasion of the carbon brush of the slip ring of the generator. According to the prediction of the GRA-LSTM model, the absolute value of the residual error of the generator temperature exceeds the alarm threshold value in 1/09/25/10/2012, and as shown in FIG. 5, the alarm is carried out 72.5 hours earlier than the actual fault, and no false alarm occurs before the absolute value of the residual error exceeds the limit.

2) Failure F2: the 18 # unit is shut down due to a three-phase winding fault of the generator at 6: 10, 10 months, 2012. The absolute value of the residual generator temperature exceeds the alarm threshold value in 2012, 10, 8, 15:50, and the alarm is early-warned 51 hours earlier than the actual fault, as shown in fig. 6. The situation that the absolute value of the residual error is reduced appears in the second half of the curve, but the absolute value is still larger than the alarm threshold value, and the fault identification is not influenced.

3) Failure F3: the 18 th unit is stopped due to the failure of the water cooling circulation of the generator in 2016, 3, 11, 15: 30. According to the prediction of the GRA-LSTM model, the absolute value of the residual error of the generator temperature exceeds the alarm threshold value in 2016, 3, 16, 01:20, and the alarm is carried out 29.7 hours earlier than the actual fault occurrence as shown in FIG. 7. In the middle process, the situation that the absolute value of the residual error approaches the alarm threshold value occurs for many times, but the alarm is not triggered. The situation that the absolute value of the residual error is lower than the alarm threshold value appears in the second half section of the curve, but the trend of the increase of the absolute value of the residual error of the temperature of the generator is obvious, and the occurrence of faults can be judged.

FIG. 8 is a stacking diagnostic model correlation analysis. And extracting fault data of which the temperature residual absolute value exceeds an alarm threshold value in three generator fault cases to form a data set. The same state variables X as in the LSTM model described above are still selected as input variables for the diagnostic model. The output is 4 generator states, namely fault F1, fault F2, fault F3 and normal state Nr. Firstly, splitting an original data set into a training set and a testing set, respectively inputting the training set into five algorithms of RF, SVM, KNN, GBDT and XGboost for model training, then diagnosing the testing set, and finally performing the spearman correlation analysis on the obtained diagnosis result, wherein the result is shown in FIG. 8.

As can be seen from FIG. 8, the spearman correlation coefficients of the five algorithms are generally higher, wherein the correlation degrees of RF, KNN, GBDT and XGboost are the highest, which indicates that the fused model has the best diagnostic effect. Therefore, RF, KNN, GBDT and XGboost are selected as 4 base learners in the first layer of diagnostic model of the stacking fusion algorithm.

Inputting the training set into a constructed stacking fusion algorithm for training, and then diagnosing 4 generator states in the test set sample, wherein the stacking diagnosis result is shown as a confusion matrix in a table 2. The diagonal of the confusion matrix corresponds to the number of diagnostic errors and the row or column of the non-diagonal corresponds to the number of diagnostic errors.

TABLE 2Stacking fusion algorithm diagnostic results

As can be seen from table 2, the stacking fusion algorithm can fully diagnose the fault F3. For faults F1 and F2, there are only few diagnostic errors, because fault F2 occurs very close to fault F1, resulting in the two types of fault information overlapping and the stacking fusion algorithm failing to identify them effectively.

On the premise of using the same data set, the stacking fusion algorithm is compared with RF, SVM, KNN, GBDT, XGboost and stacking1, stacking2 and stacking3 diagnostic models which also use the fusion mode, and the diagnostic results of the models are shown in Table 3.

As can be seen from table 3, the diagnostic models (RF, GBDT, and XGBoost) using the integrated approach have higher diagnostic accuracy than the single diagnostic models (SVM and KNN). And 4 stacking models in a fusion mode are adopted, and the diagnosis precision is higher than that of a single algorithm model. Wherein, the diagnosis precision of the stacking model fused by the 4 base learners extracted by the spearman correlation analysis is the highest, which indicates the accuracy of the spearman correlation analysis. In addition, compared with other single algorithm models, the diagnosis precision of the stacking fusion algorithm is improved by 6.12% on average. Therefore, the effectiveness of the stacking fusion algorithm for generator fault diagnosis is verified.

TABLE 3 comparison of diagnostic model Performance

Note: stacking1 fuses KNN, SVM and XGboost algorithms; stacking2 fuses four algorithms with the lowest relevance degree, namely RF, SVM, GBDT and XGboost; the stacking3 fuses five algorithms of RF, KNN, SVM, GBDT and XGboost.

On the basis, the performance of the stacking fusion algorithm is evaluated by using the evaluation index of the diagnosis model. The diagnostic performance indicators for the various diagnostic models in three generator faults are shown in table 4.

TABLE 4 three different fault recognition effects of the generator

As can be seen from Table 4, for the fault F1, the stacking fusion algorithm has the highest correct rate and the lowest false alarm rate, and F is the same as the fault F1₁The score is also the highest, which indicates that the stacking fusion algorithm has the best comprehensive performance for diagnosing the fault F1; for the fault F2, the stacking fusion algorithm and the RF have the lowest false alarm rate (3%), but the former has the correct rate and F₁The score is the highest in all the diagnosis models, which indicates that the performance of the stacking fusion algorithm for diagnosing the fault F2 is the best; for the fault F3, the diagnostic models except for KNN can be identified by 100%, because the early warning time (29.7 hours) of the fault F3 is the shortest of the three faults, resulting in fewer training samples, so that the KNN model suitable for a large number of samples is misjudged. Therefore, the effectiveness of the stacking fusion algorithm in generator fault diagnosis is verified again.

While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and various equivalent modifications and substitutions can be easily made by those skilled in the art within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A fault diagnosis method for an offshore doubly-fed wind generator based on a GRA-LSTM-stacking model is characterized by comprising the following steps:

2. The method for diagnosing the fault of the offshore doubly-fed wind generator based on the GRA-LSTM-stacking model as claimed in claim 1, wherein the step 1 comprises the following sub-steps:

3. The method for diagnosing the fault of the offshore doubly-fed wind generator based on the GRA-LSTM-stacking model as claimed in claim 2, wherein the grey correlation coefficient in the step 12 is calculated by the following formula:

4. The method for diagnosing the fault of the offshore doubly-fed wind generator based on the GRA-LSTM-stacking model as claimed in claim 2, wherein the gray correlation degree in the step 13 is calculated by the following formula:

in the formula, r_jThe degree of association is a gray color,

the entropy weight theory is used for optimization, namely the value is 1.

5. The method for diagnosing the fault of the offshore doubly-fed wind generator based on the GRA-LSTM-stacking model as claimed in claim 1, wherein the step 2 specifically comprises: and taking the state variable extracted by the GRA model as input and the generator temperature as output, training the LSTM network, and iteratively updating the weight and the offset to minimize the error to obtain a generator temperature prediction model and output a generator temperature prediction value.

6. The method for diagnosing the fault of the offshore doubly-fed wind generator based on the GRA-LSTM-stacking model as claimed in claim 1, wherein the step 3 comprises the following sub-steps:

7. The method for diagnosing the fault of the offshore doubly-fed wind turbine based on the GRA-LSTM-stacking model as claimed in claim 6, wherein the alarm threshold value in the step 32 is calculated by the formula:

8. The method for diagnosing the fault of the offshore doubly-fed wind generator based on the GRA-LSTM-stacking model as claimed in claim 1, wherein the step 4 specifically comprises: and 3, carrying out data set division on the fault sample extracted in the step 3, training a first-layer base learner of a stacking fusion algorithm by adopting K-fold cross validation, outputting probability output serving as a diagnosis result, combining the probability output by the first-layer base learner and serving as a new data set, training the probability output by the first-layer base learner and serving as the input of a second-layer element learner, and outputting a final accurate diagnosis result.

9. The offshore doubly-fed wind turbine fault diagnosis method based on GRA-LSTM-stacking model as claimed in claim 8, wherein the related algorithm model selected in the first layer of base learner comprises: the method comprises the following steps of supporting a vector machine, a K nearest classification algorithm, a random forest integrated in a bagging mode, a gradient boosting decision tree and an extreme gradient boosting tree integrated in a boosting mode, wherein after each base learner in the first layer of base learners is independently diagnosed, the diagnosis result is further subjected to the sperman relevance analysis, and the calculation formula of the sperman relevance analysis is as follows:

and

10. The offshore doubly-fed wind turbine fault diagnosis method based on GRA-LSTM-stacking model as claimed in claim 8, wherein the related algorithm model selected in the second hierarchical learner includes GBDT.