A wind turbine operating state abnormality detection method based on state parameter reconstruction error

Abstract

The present invention belongs to the field of renewable energy power generation, and specifically relates to a method for detecting abnormal operation status of a wind turbine generator set based on state parameter reconstruction error. The method comprises: using data of a wind turbine generator set in normal operation to set a monitoring index threshold and a continuous over-limit time threshold; subjecting the data to be detected to sliding window processing to obtain an augmented state data matrix, and reconstructing the data to be detected; calculating the reconstruction error and the monitoring index; comparing the calculated reconstruction error and the monitoring index with the set monitoring index threshold and the continuous over-limit time threshold, respectively, to obtain a data detection result; the present invention reconstructs the operation data of the wind turbine generator set, and through the reconstruction error and the continuous over-limit number, the abnormal state of the operation data can be judged through an identification model, the calculation is convenient and quick, and a high abnormality identification capability can be obtained.

Description

Technical Field

[0001] The present invention belongs to the field of renewable energy power generation, and in particular relates to a method for detecting abnormal operation status of a wind turbine generator set based on state parameter reconstruction error. Background Art

[0002] The wind farm data acquisition and monitoring system provides a large amount of wind turbine operation data, but most of them are healthy operation data of the unit. Fault abnormality data is lacking or even difficult to obtain. When the wind turbine status parameter in the SCADA system exceeds the limit alarm, the unit's fault or abnormality has often developed to a certain extent. Therefore, how to effectively and comprehensively utilize the existing large amount of healthy operation data to realize wind turbine operation status monitoring is a difficult problem that all wind power researchers have to solve. If the abnormal operation status data of wind turbines can be accurately identified, it will have good practical application value for improving the safety, economy and health management level of wind turbine operation.

[0003] The current wind turbine operating status monitoring method based on SCADA data is to monitor and judge the operating status of unit components by analyzing a single or a small number of unit status parameters, which is difficult to fully reflect the operating status of the entire unit. Summary of the invention

[0004] Aiming at the problem that the existing wind turbine operation status monitoring method based on SCADA data has a low result in judging the operation status of the unit through a single or a small amount of data, the present invention proposes a wind turbine operation status abnormality detection method based on state parameter reconstruction error, the method comprising:

[0005] S1: Obtain the data to be detected;

[0006] S2: Use a sliding window to process the data to be detected to obtain an augmented state data matrix;

[0007] S3: Input the data to be detected into the trained SDAE network to obtain reconstructed data;

[0008] S4: Calculate the reconstruction error according to the augmented state data matrix and the reconstructed data, and calculate the monitoring index according to the reconstruction error;

[0009] S5: Set monitoring index thresholds and continuous crossing time thresholds;

[0010] S6: Compare the calculated monitoring index with the set monitoring index threshold. If the calculated monitoring index is smaller than the set monitoring index threshold, execute step S7; otherwise, return to step S1.

[0011] S7: Calculate the continuous crossing number of the reconstructed data, and compare the continuous crossing number with the set continuous crossing time threshold. If it is greater than the continuous crossing time threshold, the data to be detected is abnormal data, otherwise the data to be detected is noise data; the continuous crossing number is: the number of times in a monitored indicator that continuously exceeds the set monitoring indicator threshold;

[0012] S8: Obtain abnormal components of the wind turbine operating status according to the abnormal data.

[0013] Preferably, the process of processing the data to be detected using a sliding window includes:

[0014] S21: Set the width of the sliding window and the number of sliding windows;

[0015] S22: using a sliding window to collect data to be detected;

[0016] S23: Construct an augmented state data matrix based on the data collected by all sliding windows.

[0017] Preferably, the process of training the SDAE network includes: the SDAE network includes a first DAE model and a second DAE model;

[0018] Step 1: Obtain the original SCADA data, process the original SCADA data using a sliding window, and obtain an augmented matrix of the original data;

[0019] Step 2: Initialize the SDAE network parameters, including the initial noise ratio c0, the noise ratio threshold c r

[0020] Step 3: Add noise to the augmented matrix according to the noise ratio C0;

[0021] Step 4: Input the augmented matrix data after noise processing into the first DAE model for model training to obtain the first weight value and update the noise ratio;

[0022] Step 5: Compare the updated noise-addition ratio with the noise-addition ratio threshold. If they are not equal, return to step 3; if they are equal, execute step 6;

[0023] Step 6: Input the hidden layer output of the first DAE model and the initial noise ratio c0 into the second DAE model for training to obtain the second weight value, and update the noise ratio;

[0024] Step 7: Compare the updated noise-addition ratio with the noise-addition ratio threshold. If it is less than the noise-addition ratio threshold, the training ends and the training parameters are output. Otherwise, return to step 6.

[0025] Preferably, the formula for calculating the monitoring index based on the reconstruction error is:

[0026]

[0027] Among them, t k Indicates monitoring indicators. represents the reconstruction error of the kth normal sample, N represents the number of reconstructed data, u represents the mean vector, Cov -1 represents the inverse of the covariance matrix.

[0028] Preferably, the process of setting the monitoring indicator threshold and the continuous crossing time threshold includes:

[0029] S51: Acquire the original SCADA data of the normal operation of the system and preprocess the data; perform correlation analysis on the preprocessed data to obtain the short-term autocorrelation of the time series data of each state parameter of the unit;

[0030] S52: Processing the unit status data using a sliding window to obtain an augmented state data matrix of normal system operation;

[0031] S53: inputting the SCADA data into the trained SDAE network according to the short-term autocorrelation of the time series data of each state parameter of the unit to obtain the second reconstructed data;

[0032] S54: Reconstruction error and monitoring index of normal data of computer group according to the reconstructed data and the augmented state data matrix of normal operation;

[0033] S55: Determine the monitoring index threshold value according to the monitoring index structure of the normal data of the unit;

[0034] S56: Construct a continuous over-limit indicator threshold according to the sliding window size and the maximum value of the continuous over-limit of the monitoring indicator.

[0035] Furthermore, the correlation analysis of the preprocessed data includes: using a comprehensive correlation index to analyze the mutual correlation between the state parameters of the wind turbine set; using the autocorrelation coefficient and the partial autocorrelation coefficient to analyze the short-term autocorrelation of the time series data of each state parameter of the unit; based on the data collected by the SCADA system; calculating the autocorrelation coefficient of the time series delayed k order; and calculating the partial autocorrelation coefficient of the data based on the autocorrelation coefficient.

[0036] Furthermore, the partial autocorrelation coefficient is expressed as:

[0037]

[0038] in, represents the first partial autocorrelation coefficient, ρ j represents the jth autocorrelation coefficient, represents the k+1th partial autocorrelation coefficient, and k represents the length of the sequence.

[0039] Preferably, the process of determining the threshold of the monitoring indicator includes: constructing a probability density function of the reconstruction error Mahalanobis distance Set the confidence level α; according to the probability density function The probability value P(t <t d ); According to the set confidence α and probability value P(t <t d ) Reversely determine the monitoring indicator threshold t d .

[0040] Furthermore, the formula for the probability value of the computer group status monitoring index being less than the monitoring index threshold is:

[0041]

[0042]

[0043] Among them, P represents the probability value that the unit status is less than the monitoring index threshold, Represents the probability density function of the reconstruction error Mahalanobis distance, n represents the number of samples, t k is the monitoring index of the kth sample data point in the sample data, h is the bandwidth parameter, and K(·) is the kernel function.

[0044] Preferably, the expression of the continuous crossing limit index threshold is:

[0045] Q d =max(Q max ,β)

[0046] Among them, Q max is the maximum value of the continuous over-limit of the monitoring indicator of the training data, and β is the width of the sliding window.

[0047] Beneficial effects of the present invention:

[0048] The present invention uses a sliding window method to process the data to be detected to obtain an augmented state data matrix, and uses the data reconstructed by the SDAE network and the augmented state data matrix to obtain a reconstruction error, compares the reconstruction error with a set threshold value, and obtains abnormal operation state detection of the wind turbine generator set; the input data and reconstructed data are linked to make the final recognition result more accurate; the present invention designs a method for setting a monitoring index threshold and a continuous over-limit time threshold, through which the threshold is set more accurately and the result is more accurate; the present invention reconstructs the operation data of the wind turbine generator set to be detected, judges the abnormal state of the operation data by the reconstruction error and the continuous over-limit number, the calculation is convenient and fast, and the recognition result is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of a method for detecting abnormal operation state of a wind turbine generator set based on state parameter reconstruction error of the present invention;

[0050] Figure 2 This is a flowchart of training the SDAE network according to the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] A specific implementation method of a wind turbine operating state abnormality detection method based on state parameter reconstruction error, such as Figure 1 As shown, the method includes:

[0053] S1: Obtain the data to be detected;

[0054] S2: Use a sliding window to process the data to be detected to obtain an augmented state data matrix;

[0055] S3: Input the data to be detected into the trained SDAE network to obtain reconstructed data;

[0056] S4: Calculate the reconstruction error according to the augmented state data matrix and the reconstructed data, and calculate the monitoring index according to the reconstruction error;

[0057] S5: Set monitoring index thresholds and continuous crossing time thresholds;

[0058] S6: Compare the calculated monitoring index with the set monitoring index threshold. If the calculated monitoring index is smaller than the set monitoring index threshold, execute step S7; otherwise, return to step S1.

[0059] S7: Calculate the continuous over-limit number of the reconstructed data, and compare the continuous over-limit number with the set continuous over-limit time threshold. If it is greater than the continuous over-limit time threshold, the data to be detected is abnormal data, otherwise the data to be detected is noise data; the continuous over-limit number is: the number of times in the monitored index that exceeds the set monitoring index threshold for multiple times; the continuous over-limit number is at least Q d Second-rate.

[0060] S8: Obtain abnormal components of the wind turbine operating status according to the abnormal data.

[0061] Continuous over-limit number: The normal data obtained after the SCADA system is used to identify and correct the bad data of the collected original operating data may still contain some noise data. Such noise data will cause the monitoring indicators to exceed the limit, but the continuous over-limit time is short. Therefore, in order to accurately identify the operating status of the unit, the continuous over-limit number indicator Q of the monitoring indicator must also be considered. The threshold Qd of Q is obtained based on the statistical distribution of the monitoring indicators of normal sample data. After the data is processed through the sliding window, the augmented state data matrix is ​​obtained. The continuous β data points of the augmented state data matrix all contain part of the data information of the original data. Therefore, when a certain data in the original data is noise data, it will cause the monitoring indicators of the continuous data points of the augmented state data matrix to exceed the limit. Therefore, the minimum value of the continuous over-limit number of the monitoring indicator should be the width of the sliding window.

[0062] The process of using a sliding window to process the data to be detected includes:

[0063] Step (21), using the data matrix X of the normal state of the unit N , set the width of the sliding window to β, m samples have a total of m-β+1 sliding windows, record For the monitoring data of the i-th unit collected in the l-th sliding window, there is a formula:

[0064]

[0065] in, represents the i-th monitoring data collected by the (l+β-1)-th sliding window, and T represents transposition.

[0066] Step (22), the obtained l-th sliding window data S l for in, represents the monitoring data of the nth unit collected by the lth sliding window;

[0067] Step (23): Based on the data of the i-th unit monitoring parameter collected by the l-th sliding window obtained in step (21) and the data collected by the l-th sliding window obtained in step (22), the state parameter data is processed by the sliding window to obtain the corresponding augmented state data matrix Y=[S (1) ,S (2) ,…,S (m-β+1) ]; among them, S (m-β+1) Represents the monitoring data collected by the m-β+1th sliding window.

[0068] like Figure 2 As shown, the process of training the SDAE network includes: the SDAE network includes a first DAE model and a second DAE model;

[0069] Step 1: Obtain the original SCADA data, process the original SCADA data using a sliding window, and obtain an augmented matrix of the original data;

[0070] Step 2: Initialize the SDAE network parameters, including the initial noise ratio c0, the noise ratio threshold c T ;

[0071] Step 3: Add noise to the augmented matrix according to the noise ratio c0;

[0072] Specifically, after obtaining the augmented state data matrix Y N After that, according to the initial noise ratio c0, the matrix Y N Noise processing Get input data Among them, c0 represents the initial noise ratio, c T Indicates the final noise-addition ratio, c0>c1>…>c T , and c0-c1=…=c T-1 -c T =Δc, Δc>0; q() represents the process of noise addition, Δc represents the difference between the two noise addition ratios, which is a fine-tuning of the noise addition ratio.

[0073] Step 4: Input the augmented matrix data after noise processing into the first DAE model for training, obtain the first weight, and update the noise ratio; the noise ratio update formula is: c t =c t-1 -Δc.

[0074] Step 5: Compare the updated noise-addition ratio with the noise-addition ratio threshold. If they are not equal, return to step 3; if they are equal, execute step 6;

[0075] Step 6: Input the hidden layer output of the first DAE model and the initial noise ratio c0 into the second DAE model for training to obtain the second weight value, and update the noise ratio;

[0076] Step 7: Compare the updated noise-addition ratio with the noise-addition ratio threshold. If it is less than the noise-addition ratio threshold, the training ends and the training parameters are output. Otherwise, return to step 6.

[0077] After training, the optimal parameters output by the model include W1, b1, W2, b2 and h, where W1 represents the first weight, b1 represents the first bias, W2 represents the second weight, b2 represents the second bias, and h represents the hidden layer.

[0078] The process of setting the monitoring indicator threshold and the continuous crossing time threshold includes:

[0079] Step 1: Obtain the original SCADA data of the normal operation of the system and preprocess the data; perform correlation analysis on the preprocessed data to obtain the short-term autocorrelation of the time series data of each state parameter of the unit;

[0080] Step 2: Use a sliding window to process the unit state data to obtain an augmented state data matrix of the normal operation of the system; the autocorrelation coefficient of the wind turbine state parameter reflects the linear correlation between the previous moment value and the current moment value of the state parameter;

[0081] Step 3: Input the SCADA data into the trained SDAE network according to the short-term autocorrelation of the time series data of each state parameter of the unit to obtain the second reconstructed data;

[0082] Step 4: Reconstruction error and monitoring index of normal data of computer group according to the reconstructed data and the augmented state data matrix of normal operation;

[0083] Step 5: Determine the monitoring index threshold value based on the monitoring index structure of the normal data of the unit;

[0084] Step 6: Construct a continuous over-limit indicator threshold based on the sliding window size and the maximum value of the continuous over-limit of the monitoring indicator.

[0085] The process of preprocessing the raw SCADA data and performing correlation analysis on the preprocessed data includes:

[0086] Step (11) identifies and corrects the bad data of the original unit status data collected by the SCADA system, and obtains the data matrix of the normal state of the unit, which is in, represents the i-th sample data of the j-th state parameter, n is the number of sample data collected, and m is the number of unit state parameters.

[0087] Step (12), using a comprehensive correlation index to analyze the mutual correlation between the state parameters of the wind turbine generator set, wherein the calculation formula of the comprehensive correlation coefficient ICC is:

[0088]

[0089] in, is the Pearson correlation coefficient, is the Kendall correlation coefficient, is the Spearman correlation coefficient, x (k) represents the sample data of the kth state parameter, x (l) Represents the sample data of the lth state parameter.

[0090] Step (13), using autocorrelation coefficient and partial autocorrelation coefficient to analyze the short-term autocorrelation of the time series data of each state parameter of the unit; according to the data x1, x2, ..., x collected by the SCADA system n , define the time series {x t ,t∈T}, time series {x t The autocorrelation coefficient of the k-order delay is:

[0091]

[0092] Partial autocorrelation coefficient:

[0093]

[0094] Among them, ρ k represents the autocorrelation coefficient of the time series delayed k order, x t represents the time series, cov represents the covariance matrix, x t-k represents a time series with a delay of k, var() represents the second-order moment of the covariance probability density, The k-order delayed partial autocorrelation coefficient is the j-order autoregressive model for x t The coefficients of the linear minimum variance estimate are: The k+1th-order delayed partial autocorrelation coefficient is the jth-order autoregressive model for x. t The coefficients for linear minimum variance estimation.

[0095] Specifically, the reconstruction error E of the normal data of the computer group N The process of monitoring the normal data of the unit includes:

[0096] Step (41): first collect the original data X N After that, the sliding window is processed to obtain the augmented state data matrix Y N , using the SDAE model training with multiple noise ratios, we get the reconstructed data Z N , then the reconstruction error is E N =Y N -Z N ;

[0097] Step (42), because the input of the SMDAE model is multivariate state data, E N For multivariate data, in order to realize the abnormal state of the unit, it is necessary to N The condition monitoring indicators are extracted from the data; the Mahalanobis distance is used to measure the similarity between unknown data and known data;

[0098] Step (43), set E N The mean vector is μ=(μ1,μ2,…,μ n), the Mahalanobis distance of the reconstruction error is selected as the monitoring indicator, and the monitoring indicator formula is obtained as follows:

[0099]

[0100] in, represents the reconstruction error of the kth normal sample, u represents the mean vector, Cov -1 represents the inverse of the covariance matrix.

[0101] Determine the threshold t of the monitoring indicator d And the monitoring indicator continuous crossing time threshold Q d The process includes:

[0102] Step (51), determine the probability density function of the reconstruction error Mahalanobis distance of the normal operation data of the unit

[0103]

[0104] Among them, t k is the monitoring index of the kth sample data point in the sample data, n represents the number of sample data, h is the bandwidth parameter, K(·) is the kernel function, and satisfies the formula:

[0105]

[0106] Among them, K(u) represents the kernel function of u, and u2(K) represents the product of the constraint coefficient and the kernel function.

[0107] Step (52), determine the threshold value t of the monitoring indicator d The process includes: setting the confidence level α; according to the probability density function The probability value P(t <t d ); According to the set confidence α and probability value P(t <t d ) Reversely determine the monitoring indicator threshold t d ; The formula for the probability value of the computer group status monitoring index being less than the monitoring index threshold is:

[0108]

[0109]

[0110] Among them, P represents the probability value that the unit status is less than the monitoring index threshold, Represents the probability density function of the reconstruction error Mahalanobis distance, n represents the number of samples, t k is the monitoring index of the kth sample data point in the sample data, h is the bandwidth parameter, and K(·) is the kernel function.

[0111] Step (53), noise data will cause the monitoring index to exceed the limit. The continuous exceeding time of this type of noise is relatively short. In order to accurately identify it, it is necessary to further adopt the continuous exceeding number index Q; the threshold value Q of the continuous exceeding number index Q is d The formula satisfies:

[0112] Q d =max(Q max ,β)

[0113] Among them, Q max is the maximum value of the continuous over-limit of the monitoring indicator of the training data; β is the width of the sliding window.

[0114] A specific implementation of a method for detecting abnormal operation status of a wind turbine generator system based on state parameter reconstruction error, the method comprising:

[0115] The first step is to use the original normal operation SCADA data, denoted as X N The mutual correlation between state parameters is analyzed by comprehensive correlation index, and the short-term autocorrelation of the time series data of each state parameter of the unit is analyzed by autocorrelation coefficient and partial autocorrelation coefficient; the unit state data is processed by sliding window technology to obtain the augmented state data matrix Y for calculating the short-term autocorrelation of the parameters. N ; The SDAE model with deep network structure has high-intensity nonlinear projection ability; Here, the noise ratio is introduced, which is a very important tuning parameter of the DAE network. Therefore, multiple noise ratios are used to train the SDAE model and obtain the reconstructed output Z of normal data N ;

[0116] The second step is to use formula E N =Y N -Z N The reconstruction error E of the normal data of the computer group N , the monitoring index t of the normal data of the computer group is calculated according to the formula of the reconstruction error; the threshold value t of the monitoring index is determined according to the probability value that the monitoring index of the unit state is less than the monitoring index threshold d , according to the expression of the continuous over-limit number indicator threshold, determine the monitoring indicator continuous over-limit time threshold Q d ;

[0117] The third step is to input the data to be tested X M , the corresponding augmented state data matrix Y is obtained through sliding window processing M , the SDAE network trained with normal samples calculates the reconstructed output Z of the data to be detected M , according to E M =Y M -X M Calculate the reconstruction error E of the data to be detectedM ,according to Calculate the monitoring index t of the data to be detected M ; Judge t M >t d If it is wrong, the data to be tested is returned. If it is correct, the continuous crossing number Q is obtained by counting M , judge Q M >Q d If it is wrong, it is considered as noise data. If it is correct, it indicates that it is abnormal data. Then the contribution of each parameter is calculated to determine the status parameter or component that is most related to the abnormal status of the unit.

[0118] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting abnormal operation status of a wind turbine generator system based on state parameter reconstruction error, characterized in that: The method includes: S1: Obtain the data to be detected; S2: Use a sliding window to process the data to be detected to obtain an augmented state data matrix; S3: Input the data to be detected into the trained SDAE network to obtain reconstructed data; S4: Calculate the reconstruction error according to the augmented state data matrix and the reconstructed data, and calculate the monitoring index according to the reconstruction error; S5: Setting monitoring indicator thresholds and continuous crossing time thresholds; the specific steps include: S51: Acquire the original SCADA data of the normal operation of the system and preprocess the data; perform correlation analysis on the preprocessed data to obtain the short-term autocorrelation of the time series data of each state parameter of the unit; The correlation analysis of the preprocessed data includes: using comprehensive correlation indicators to analyze the mutual correlation between the state parameters of the wind turbine; using autocorrelation coefficients and partial autocorrelation coefficients to analyze the short-term autocorrelation of the time series data of each state parameter of the unit; according to the data collected by the SCADA system; calculating the autocorrelation coefficient of the time series delayed k order; calculating the partial autocorrelation coefficient of the data based on the autocorrelation coefficient; the partial autocorrelation coefficient is expressed as: in, represents the first partial autocorrelation coefficient, ρ j represents the jth autocorrelation coefficient, represents the k+1th partial autocorrelation coefficient, k represents the length of the sequence, The k-order delayed partial autocorrelation coefficient is the j-order autoregressive model for x t The coefficients of the linear minimum variance estimate are: The k+1th-order delayed partial autocorrelation coefficient is the jth-order autoregressive model for x. t Coefficients for linear minimum variance estimation; S52: Processing the unit status data using a sliding window to obtain an augmented state data matrix of normal system operation; S53: inputting the SCADA data into the trained SDAE network according to the short-term autocorrelation of the time series data of each state parameter of the unit to obtain the second reconstructed data; S54: Reconstruction error and monitoring index of normal data of computer group according to the reconstructed data and the augmented state data matrix of normal operation; S55: Determine the monitoring index threshold value according to the monitoring index structure of the normal data of the unit; S56: constructing a continuous over-limit indicator threshold according to the sliding window size and the maximum value of the continuous over-limit of the monitoring indicator; S6: Compare the calculated monitoring index with the set monitoring index threshold. If the calculated monitoring index is greater than the set monitoring index threshold, execute step S7, otherwise return to step S1; S7: Calculate the continuous crossing number of the reconstructed data, and compare the continuous crossing number with the set continuous crossing time threshold. If it is greater than the continuous crossing time threshold, the data to be detected is abnormal data, otherwise the data to be detected is noise data; the continuous crossing number is: the number of times that the monitored index continuously exceeds the set monitoring index threshold; S8: Obtain abnormal components of the wind turbine operating status according to the abnormal data.

2. The method for detecting abnormal operation status of a wind turbine generator system based on state parameter reconstruction error according to claim 1, characterized in that: The process of using a sliding window to process the data to be detected includes: S21: Set the width of the sliding window and the number of sliding windows; S22: using a sliding window to collect data to be detected; S23: Construct an augmented state data matrix based on the data collected by all sliding windows.

3. The method for detecting abnormal operation status of a wind turbine generator system based on state parameter reconstruction error according to claim 1, characterized in that: The process of training the SDAE network includes: the SDAE network includes a first DAE model and a second DAE model; Step 1: Obtain the original SCADA data, process the original SCADA data using a sliding window, and obtain an augmented matrix of the original data; Step 2: Initialize the SDAE network parameters, including the initial noise ratio c O , plus noise ratio threshold c T ; Step 3: Add noise to the augmented matrix according to the noise ratio c0; Step 4: Input the augmented matrix data after noise processing into the first DAE model for training, obtain the first weight value, and update the noise ratio; Step 5: Compare the updated noise-addition ratio with the noise-addition ratio threshold. If they are not equal, return to step 3; if they are equal, execute step 6; Step 6: Input the hidden layer output of the first DAE model and the initial noise ratio c0 into the second DAE model for training to obtain the second weight value, and update the noise ratio; Step 7: Compare the updated noise-addition ratio with the noise-addition ratio threshold. If it is less than the noise-addition ratio threshold, the training ends and the training parameters are output. Otherwise, return to step 6.

4. The method for detecting abnormal operation status of a wind turbine generator system based on state parameter reconstruction error according to claim 1, characterized in that: The formula for calculating the monitoring index based on the reconstruction error is: Among them, t k Indicates monitoring indicators. represents the reconstruction error of the kth normal sample, N represents the number of reconstructed data, u represents the mean vector, Cov -1 represents the inverse of the covariance matrix.

5. The method for detecting abnormal operation status of a wind turbine generator system based on state parameter reconstruction error according to claim 1, characterized in that: The process of determining the threshold of the monitoring indicator includes: constructing the probability density function of the reconstruction error Mahalanobis distance Set the confidence level α; according to the probability density function The probability value P(t <t d ); According to the set confidence α and probability value P(t <t d ) Determine the monitoring indicator threshold t d .

6. According to the method for detecting abnormal operation status of a wind turbine generator set based on state parameter reconstruction error of claim 5, the formula for calculating the probability value of the wind turbine generator set state monitoring index being less than the monitoring index threshold is: in, P represents the probability value that the unit status is less than the monitoring index threshold. Represents the probability density function of the reconstruction error Mahalanobis distance, n represents the number of samples, t k is the monitoring index of the kth sample data point in the sample data, h is the bandwidth parameter, and K(·) is the kernel function.

7. The method for detecting abnormal operation status of a wind turbine generator system based on state parameter reconstruction error according to claim 1, characterized in that: The expression of the continuous crossing limit indicator threshold is: Q d =max(Q max ,β) Among them, Q max is the maximum value of the continuous over-limit of the monitoring indicator of the training data, and β is the width of the sliding window.

Images