Fan fault diagnosis method based on CapBLSTM-MOHHO

Abstract

The present invention discloses a wind turbine fault diagnosis method based on CapBLSTM-MOHHO, including: performing feature selection through SCADA historical data, obtaining non-redundant feature data related to wind power, and preprocessing the non-redundant feature data; using extreme value theory to analyze the residual between the predicted value and the actual value, and determining the alarm threshold; constructing a prediction model, inputting the SCADA online data to be diagnosed into the prediction model, outputting the predicted value, and determining the operating state of the wind turbine based on the predicted value and the alarm threshold, wherein the prediction model includes a plurality of prediction units, the prediction unit is constructed using a capsule-bidirectional long short-term memory network, and the prediction unit is obtained by training the preprocessed non-redundant feature data. The method of the present invention can effectively process sequence data and accurately predict wind power, realize fault warning of the wind turbine, and is suitable for multiple wind farms under the jurisdiction of a wind power centralized control center.

Description

Technical Field

[0001] The present invention relates to the technical field of fan fault diagnosis, and in particular to a fan fault diagnosis method based on CapBLSTM-MOHHO. Background Art

[0002] With the continuous development of the wind power industry, the scale of wind farms continues to grow, and wind power companies often own multiple wind farms distributed in different regions. However, due to the relatively scattered locations of wind farms, management is difficult and the operation and maintenance costs are high. In addition, most wind farms are located in remote mountains and Gobi deserts, and they also face problems such as shortage of operation and maintenance talents and difficulty in recruiting personnel. This internal and external environment has put forward higher requirements on the management and management level of wind power companies. In order to achieve the goal of "no or few people on duty, remote monitoring", effectively integrate company resources, reduce operation and maintenance costs, improve the company's economic benefits, and enhance the core competitiveness of wind power companies, the "centralized control center" has emerged as a new operating model for new energy companies.

[0003] Due to the remote location of the wind farm, the wind turbine operating environment is harsh, and the impact of sandstorms can easily cause wind turbine failures. The warranty period of wind turbines is generally only 3-5 years. The wind turbines built earlier have long exceeded the warranty period, and the aging of wind turbines is prone to frequent failures. Frequent failures of wind turbines will, on the one hand, cause insufficient output performance of wind turbines and affect the power generation efficiency of wind turbines; on the other hand, long-term shutdown and maintenance of wind turbines due to failures will cause significant economic losses to the wind farm.

[0004] An efficient wind turbine fault diagnosis method can analyze the online data of the wind turbine, monitor the operating status of the wind turbine in real time, and issue an alarm message when the wind turbine is in an abnormal operating state. It can identify the fault in time at the early stage of the fault to avoid the degradation of the fault. Therefore, an efficient wind turbine fault diagnosis method can improve the operation and maintenance strategy from post-maintenance and planned maintenance to condition-based maintenance and predictive maintenance, which helps to reduce the operation and maintenance cost and ensure the long-term healthy and stable development of the wind power industry. Since the remote control center established by the wind power company needs to monitor and manage the wind turbines in different wind farms, and the environment of each wind farm is different, it is of great significance to study a wind turbine fault diagnosis method suitable for multiple wind farms. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a wind turbine fault diagnosis method based on CapBLSTM (capsule-bidirectional long short-term memory network)-MOHHO (multi-objective Harris Hawk optimization method), and adopts a capsule-bidirectional long short-term memory network prediction model to predict the wind power value, and calculates the residual between the predicted value and the actual value.

[0006] To achieve the above object, the present invention provides a fan fault diagnosis method based on CapBLSTM-MOHHO, comprising:

[0007] Performing feature selection through SCADA historical data to obtain non-redundant feature data related to wind power, and preprocessing the non-redundant feature data;

[0008] Use extreme value theory to analyze the residual between the predicted value and the actual value to determine the alarm threshold;

[0009] A prediction model is constructed, the SCADA online data to be diagnosed is input into the prediction model, a prediction value is output, and the operating state of the fan is determined based on the prediction value and the alarm threshold, wherein the prediction model includes a plurality of prediction units, the prediction units are constructed by using a capsule-bidirectional long short-term memory network, and the prediction units are obtained by training the preprocessed non-redundant feature data.

[0010] Preferably, obtaining the non-redundant characteristic data related to wind power includes:

[0011] Based on the RRelief feature selection model, nonlinear feature selection is performed on the historical data collected by the SCADA system to obtain the non-redundant feature data related to wind power, wherein the non-redundant feature data related to wind power includes wind speed, wind direction, air pressure, temperature, rotor speed, gearbox oil temperature and impeller speed.

[0012] Preferably, the method for performing nonlinear feature selection on the data collected by the SCADA system based on the RRelief feature selection model is:

[0013]

[0014] Where P diffF is the difference between samples in feature F in neighboring samples, P diffC is the difference between sample prediction values ​​in adjacent samples, P diffC|diffF is the degree of difference in sample prediction values ​​among neighboring samples that differ in feature F, is the feature weight set.

[0015] Preferably, the non-redundant feature data is preprocessed, including abnormal data processing, missing value processing and data normalization processing;

[0016] Wherein, the abnormal data processing is to eliminate the abnormal data through the box plot method;

[0017] The missing value processing is to fill the missing values ​​using the mean value interpolation method and the linear interpolation method according to the degree of data missing;

[0018] The data normalization process is as follows:

[0019]

[0020] In the formula, A max and A min Respectively represent the maximum and minimum values ​​of feature A, p i is the eigenvalue, p′ i is the normalized eigenvalue.

[0021] Preferably, the capsule-bidirectional long short-term memory network is used to extract the spatial feature relationship between the wind power and the optimal feature set at the same time point, and the capsule-bidirectional long short-term memory network includes:

[0022] Convolutional layer: used to extract spatial features of input multidimensional time series;

[0023] Main capsule layer: used to obtain the low-level features extracted by the convolutional layer;

[0024] Digital capsule layer: used to extract high-level features of input multidimensional data;

[0025] Bidirectional long short-term memory network: used to obtain the time feature relationship between the historical time points of the time series, where in the forward direction, the features from the past values ​​of wind power and the best feature set factors are stored through memory unit 1, and in the backward direction, the features from the future values ​​are stored through memory unit 2. The forward and backward hidden states are connected to obtain the final hidden output.

[0026] Preferably, the SCADA online data to be diagnosed is input into the prediction model, and the prediction value is output, including:

[0027] Selecting non-redundant feature data related to wind power in a number of typical wind farms, and using the non-redundant feature data to train an initial prediction unit to obtain a number of the prediction units;

[0028] The prediction unit is applied to the wind power centralized control center by adopting the transfer learning method, and the multivariate data of any wind farm of the wind power centralized control center is respectively input into the prediction unit, and several prediction results are output. The weights of the prediction results are determined by using the multi-objective Harris Eagle optimization method to obtain the prediction value.

[0029] Preferably, the method of applying the prediction unit to the wind power centralized control center by using the transfer learning method is:

[0030] Y s =f s (X s ,δ s )

[0031] Yt =f t (X t ,δ t )

[0032] In the formula, f s Indicates that from X s To the source domain predicted value Y s Function mapping, f t Indicates that from X t To the target domain predicted value Y t Function mapping, δ s and δ t are the model parameters of the source domain and the target domain respectively, X s is the source domain data, X t is the target domain data;

[0033] Among them, transfer learning is used to obtain the mapping f from the source domain s Migrate to the target domain, from the target domain data X t The mapping f of the target task is learned in t .

[0034] Preferably, the mathematical expression of the multi-objective Harris Hawk optimization method is:

[0035]

[0036] In the formula, Indicates w i The estimated value of w i Represents the weight of each basic model; e represents the error value; represents the predicted value; y t represents the actual value, i is the number of typical wind farms selected, t is the number of iterations, is the predicted value of the i-th pre-trained model.

[0037] Preferably, determining the operating state of the fan includes:

[0038] If the residual between the predicted value and the actual value is within the alarm threshold range, there is no abnormality; if the residual between the predicted value and the actual value exceeds the alarm threshold, it is determined that an abnormality or fault occurs and an alarm signal is issued.

[0039] Preferably, the alarm threshold is determined based on an adaptive threshold method of extreme values, wherein the adaptive threshold is:

[0040]

[0041] In the formula, R th is the obtained threshold, μ(t) is the mean, σ(t) is the standard deviation, λ is the confidence limit, and p is the probability of the cumulative distribution function.

[0042] Compared with the prior art, the present invention has the following advantages and technical effects:

[0043] The method of the present invention can effectively process sequence data and accurately predict wind power, realize fault warning of wind turbines, and is applicable to multiple wind farms under the jurisdiction of wind power centralized control centers. In addition, the adaptive threshold method based on extreme values ​​can be adjusted according to actual conditions, thereby improving the accuracy and stability of the threshold and thus improving the accuracy of wind turbine fault warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0045] Figure 1 A schematic diagram of a fan fault diagnosis process according to an embodiment of the present invention;

[0046] Figure 2 It is a schematic diagram of the overall structure of the prediction method according to an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of monitoring variables of a SCADA system according to an embodiment of the present invention;

[0048] Figure 4 A bidirectional long short-term memory network structure diagram of an embodiment of the present invention;

[0049] Figure 5 This is a flow chart of the multi-objective Harris Hawk optimization algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0052] The data-driven wind turbine fault diagnosis method mainly uses the data in the existing SCADA system, and calculates the difference between the actual value and the predicted value by establishing a model to obtain the unit status. Wind power is an important performance indicator of wind turbines. Therefore, a wind turbine fault diagnosis method based on CapBLSTM-MOHHO suitable for multiple wind farms is proposed. The capsule-bidirectional long short-term memory network prediction model is used to predict the wind power value and calculate the residual between the predicted value and the actual value.

[0053] This fan fault diagnosis method requires processing the residuals to set the alarm threshold. Due to the uncertainty of the wind profile, the operating conditions of the fan during operation vary greatly, resulting in the standard of "normal" performance of the fan changing every moment. Using a fixed threshold for state detection is prone to false warnings. Therefore, the proposed fan fault diagnosis method uses an adaptive threshold estimation method to identify abnormal data.

[0054] In this embodiment, the fan fault diagnosis method based on CapBLSTM-MOHHO is as follows: Figure 1 First, four typical wind farms are selected from many wind farms and the input data of the four typical wind farms are trained through the capsule-bidirectional long short-term memory network to obtain four pre-trained models. Then, the four pre-trained models are applied to the wind power control center by using the transfer learning method. The four pre-trained models have learned the wind power knowledge of the four typical wind farms and saved it in the network parameters. The wind power prediction result of any wind farm in the control center is trained by the four pre-trained models, and MOHHO determines the weight combination of the four pre-trained models. Finally, the residual is calculated between the predicted value and the actual value, and the state monitoring index is determined by the method based on adaptive threshold. The operating status of the wind turbine is judged by analyzing the relationship between the residual and the state monitoring index. The flow chart is as follows Figure 2 .

[0055] Step 1: Feature Selection

[0056] The data in the SCADA system is the data monitored by each component of the fan and uploaded to the centralized control center. It is the recorded value of each fan sensor monitoring point. The collected fan data is complex and varied, such as Figure 3 In order to find non-redundant features related to wind power, avoid irrelevant and uninformative features that reduce computational efficiency and improve the accuracy of wind power prediction, the RReliefF (Regression ReliefF) algorithm is used to perform nonlinear feature selection on the information collected by the SCADA system. As a result, wind speed, wind direction, air pressure, temperature, rotor speed, gearbox oil temperature and impeller speed are selected as the best feature set.

[0057] The RReliefF feature selection model is improved on the basis of ReliefF, inheriting the basic idea of ​​ReliefF to evaluate feature importance, and is suitable for regression tasks in machine learning. It evaluates feature weights based on the correlation between features and the predicted output of nonlinear time series data, and sorts features by weight. Features with weights less than a certain threshold are subsequently discarded. Given the feature weight set corresponding to the original feature attribute F The calculation formula is as follows:

[0058]

[0059] Where P diffF is the difference between samples in feature F in neighboring samples, P diffC is the difference between sample prediction values ​​in adjacent samples, P diffC|diffF P is the degree of difference in sample prediction values ​​among neighboring samples that have differences in feature F. diffF , P diffC and P diffC|diffF The definition is as follows:

[0060]

[0061] Step 2: Determine model inputs.

[0062] Step 3: Data Preprocessing

[0063] SCADA data comes from wind farm wind speed sensors, weather stations and other devices. If the collected data is directly input into the prediction model, it may be affected by missing values ​​and outliers, which will reduce the accuracy of the prediction model. Therefore, in order to ensure the integrity, accuracy and effectiveness of the collected data, the optimal feature set and wind power SCADA data selected in step 1 are preprocessed before modeling the wind farm wind power prediction sequence. The steps are as follows:

[0064] 1) Abnormal data processing

[0065] For abnormal data, the box plot method is used to remove them. This method divides the data set into quartiles and defines values ​​less than QL-1.5IQR or greater than QU+1.5IQR as abnormal values, where QL represents the lower quartile, i.e. one quarter of the values ​​in the data are less than it; QU is called the upper quartile, i.e. one quarter of the values ​​in the data are greater than it; IQR is called the interquartile range, i.e. the difference between the upper quartile QU and the lower quartile QL.

[0066] When using the box plot method to identify outliers, you can directly use the data set to draw, without the need for the data set to follow a specific form of distribution. Therefore, the box plot method can truly and intuitively show the original appearance of the distribution of characteristic data in the data set. At the same time, because the quartile itself has inherent robustness, the outlier judgment method of the box plot is based on the quartile and the range between quartiles, which can fully guarantee the accuracy of the judgment.

[0067] 2) Missing value processing

[0068] According to the degree of data missing, the average interpolation method and linear interpolation method are used to fill the missing values ​​to ensure the continuity of data in time and provide guarantee for subsequent predictions:

[0069] For non-continuous missing data, the average value of adjacent data is used for interpolation, and the average value of the two data before and after the missing data is calculated to replace it;

[0070] For continuous missing data, linear interpolation is used for interpolation, and the formula is as follows:

[0071]

[0072] Where P n , P n+x and P n+y They respectively represent the nth, n+xth, and n+yth moment values ​​of the same dimensional data.

[0073] 3) Data Normalization

[0074] Since the feature data collected by the SCADA system is diverse and the order of magnitude of the data is quite different, the feature values ​​with a large numerical range will cover the feature values ​​with a small numerical range during the prediction process, thus affecting the prediction accuracy. Therefore, the data normalization method is used to scale or convert the features to a common range to achieve the numerical uniformity of the data. The data features are normalized to [0,1] through the formula for prediction, and then the prediction results are denormalized to obtain the actual predicted value. The calculation formula is as follows:

[0075]

[0076] In the formula, A max and A min Respectively represent the maximum and minimum values ​​of feature A, p i is the eigenvalue, p′ i is the normalized eigenvalue.

[0077] Step 4: Base model prediction

[0078] In order to obtain the deep features of the input multivariate input data, a capsule-bidirectional long short-term memory network prediction model is proposed. The model first extracts the spatial features of the input multidimensional time series through the convolution layer of the capsule network. The main capsule layer obtains the low-level features extracted by the convolution layer, and the digital capsule layer finally extracts the high-level features of the input multidimensional data. Finally, the bidirectional long short-term memory network is used to obtain the temporal feature relationship between the historical time points of the time series, and the prediction results are obtained based on the spatiotemporal correlation analysis.

[0079] The wind power and optimal feature set data of any wind turbine in any wind farm in the centralized control are used as the input of the prediction model, and the spatial feature relationship between the wind power and the optimal feature set at the same time point is extracted using a capsule network-based method. The capsule network structure constructed in the prediction model has a convolution kernel size of 3×3, 128 channels, 4 steps, and 8 capsules in the main capsule layer; the number of primary caps in the digital capsule layer is 30, and the number of capsules is 16.

[0080] The output of the capsule network is used as the input of the bidirectional long short-term memory network to further capture the time correlation characteristics in the time series. Considering that the change of wind power value is not only affected by the best feature set factors at the previous moment, but also the best feature set information in the future is closely related to wind power, a bidirectional long short-term memory network is used for bidirectional time feature extraction. Its structure is as follows: Figure 4 In the forward direction, the memory cell stores features from past values ​​of wind power and the best feature set factors, and in the backward direction, it stores features from future values. The hidden states in these two directions are concatenated to produce the final hidden output, calculated as follows:

[0081]

[0082] In the formula, and yes and The weight parameter, b y is the offset vector, is the hidden output sequence of the forward layer, is the hidden output of the backward layer, and the final output result is given by h t express.

[0083] Step 5: Transfer Learning

[0084] The basic model prediction method is used to train the data of the four sites to obtain four pre-trained models, named Pretrained-1, Pretrained-2, Pretrained-3, and Pretrained-4. The four pre-trained models save the data knowledge learned from the four sites in the model parameters. The four pre-trained models are migrated to the control center through the transfer learning method. The control center has pre-trained models for the four sites, and the wind power prediction results of other wind farms under its jurisdiction can be obtained through these four pre-trained models. The transfer learning method reduces the time for retraining models in other wind farms in the control center and improves the prediction efficiency.

[0085] Transfer learning aims to use the knowledge from the source domain to improve the learning performance of the target domain, avoiding the disadvantage of starting a new model from scratch, achieving feature migration and parameter sharing, and thus improving the learning efficiency of the model. The data of the selected typical location wind farm is set as the source domain data X s , the data of other wind farms in the centralized control area are set as the target domain data X t , then the source domain task and the target domain task are expressed as follows:

[0086] Y s =f s (X s ,δ s )

[0087] Y t =f t (X t ,δ t )

[0088] In the formula, f s Indicates that from X s To the source domain predicted value Y s Function mapping, f t Indicates that from X t To the target domain predicted value Y t Function mapping, δ s and δ t are the model parameters of the source domain and the target domain respectively. Transfer learning is used to map the source domain to s Migrate to the target domain, from the target domain data X t The mapping f of the target task is learned in t .

[0089] Step 6: Combine model predictions

[0090] All four pre-trained models are migrated to the centralized control center. Because different wind farms have different characteristics, the characteristics of the four pre-trained model sites should be fully utilized. Therefore, after the multivariate data of any wind farm in the centralized control center is input into the four pre-trained models, four sets of prediction results are obtained, and the weights of the four sets of prediction results are determined using the multi-objective Harris Hawk optimization algorithm. By combining the weights, the advantages of each pre-trained model can be better utilized, thereby obtaining better prediction results.

[0091] Multi-objective optimization problems involve multiple problem solutions, which may cause conflicts between different problem objectives. To resolve these conflicts, the Pareto optimal solution method can be used to determine the best trade-offs between the problems. Multi-objective optimization is usually expressed as a minimization problem.

[0092] Minimize:

[0093] Subject to:

[0094] Where k represents the number of objective functions; n1 represents the number of inequality constraints; m represents the number of equality constraints; n2 represents the number of variables; [L t ,U t ] is the boundary of the tth variable.

[0095] The objective function of the multi-objective Harris Hawk optimization algorithm is:

[0096] Minimize:fitness(y)={f1,f2}

[0097] Subject to:

[0098] In the formula, N represents the dimension of the test set; represents the predicted value; y t Represents actual value.

[0099] The mathematical expression of the combined model is as follows:

[0100]

[0101] In the formula, Indicates w i The estimated value of w i represents the weight of each basic model; e represents the error value.

[0102] The optimization problem is expressed mathematically as:

[0103] Minimize:

[0104] Subject to:

[0105] The multi-objective Harris Hawk optimization algorithm is an algorithm that extends the Harris Hawk optimization algorithm to multi-objective optimization problems. In the multi-objective Harris Hawk optimization algorithm, not only the trade-off relationship between objective functions is considered between individuals, but also the non-dominated relationship between individuals is considered, and the optimal solution on the Pareto frontier can be found. The flowchart of the multi-objective Harris Hawk optimization algorithm is as follows Figure 5 shown.

[0106] Step 7: Determine the alarm threshold

[0107] The alarm threshold of condition monitoring is determined by an extreme value-based adaptive threshold method.

[0108] The operating conditions of the fan are very complex. Under the influence of some uncontrollable objective factors, the fan SCADA parameters may deviate from the normal value within the alarm range, which is manifested as an extreme point on the residual value. If a fixed threshold is used to set the alarm threshold of the fault alarm system, when the extreme point of the residual error is higher than the fixed threshold, a false alarm may occur. In order to solve the false alarm problem caused by the fixed alarm threshold, the present invention applies the extreme value theory to analyze the residual between the predicted value and the actual value, and designs an adaptive threshold for fault diagnosis of the fan.

[0109] Assume that X1, X2, …, X n is a vector of n samples of independently randomly distributed variables, whose distribution function is F(x). i Each sample vector of M contains a certain amount of values ​​over a period of time. n =max(X1,X2,...,X n ) represents the maximum value of n sample vectors. n , its probability distribution function can be expressed as:

[0110] P r (M n ≤x)={F(z)} n

[0111] F n (x) is usually unknown, so we need to use the extreme distribution function of the maximum or minimum value to replace F. n (x). When n→∞, F n (x)→0. Normalize the extreme distribution function to avoid M n Degenerates to a point. Assuming there are two normalized parameters μ and σ, when n→∞:

[0112]

[0113] Where μ>0, -∞<σ<∞, -∞<λ<∞, μ is the location parameter, σ is the scale parameter, and G(z) is the generalized extreme value distribution function.

[0114] The prediction model is trained by a large amount of SCADA normal data to learn the temporal and spatial correlation of the input SCADA multi-dimensional parameters. Because the model is trained under normal conditions, most of the data are normal values, the residuals will not change much, and the average value can be approximated as a constant.

[0115] μ(t)=β0

[0116] σ(t)=β0+β t g(t)

[0117] In the formula, β0, β t is a constant coefficient and g(t) is a function that describes the variable operating conditions.

[0118] The adaptive threshold is obtained by the following formula:

[0119]

[0120] In the formula, R th is the obtained threshold, μ(t) is the mean, σ(t) is the standard deviation, λ is the confidence limit, and p is the probability of the cumulative distribution function, which is calculated by the cumulative distribution function (CDF).

[0121] Step 8: Determine the fan status

[0122] Combining the predicted value and actual value of wind power of the combined model, the residual is obtained to analyze the operating status of the wind turbine. If the residual is within the alarm threshold, there is no abnormality. If the residual exceeds the alarm threshold, it is judged that an abnormality or failure occurs and an alarm signal is issued.

[0123] The present invention can effectively process sequence data and accurately predict wind power, realize fault warning of wind turbines, and is applicable to multiple wind farms under the jurisdiction of wind power centralized control centers. In addition, the adaptive threshold method based on extreme values ​​can be adjusted according to actual conditions, thereby improving the accuracy and stability of the threshold and thus improving the accuracy of wind turbine fault warning.

[0124] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A fan fault diagnosis method based on CapBLSTM-MOHHO is characterized in that: include: Performing feature selection through SCADA historical data to obtain non-redundant feature data related to wind power, and preprocessing the non-redundant feature data; Use extreme value theory to analyze the residual between the predicted value and the actual value to determine the alarm threshold; Constructing a prediction model, inputting the SCADA online data to be diagnosed into the prediction model, outputting a prediction value, and determining the operating state of the fan based on the prediction value and the alarm threshold, wherein the prediction model includes a plurality of prediction units, the prediction units are constructed by using a capsule-bidirectional long short-term memory network, and the prediction units are obtained by training the preprocessed non-redundant feature data; Wherein, inputting the SCADA online data to be diagnosed into the prediction model and outputting the prediction value comprises: Selecting non-redundant feature data related to wind power in a number of typical wind farms, and using the non-redundant feature data to train an initial prediction unit to obtain a number of the prediction units; The prediction unit is applied to the wind power centralized control center by adopting the transfer learning method, and the multivariate data of any wind farm of the wind power centralized control center is respectively input into the prediction unit, and a number of prediction results are output, and the weights of the prediction results are determined by using the multi-objective Harris Eagle optimization method to obtain the prediction value; The method of applying the prediction unit to the wind power centralized control center by using the transfer learning method is as follows: Y s =f s (X s ,d s ) In the formula, f s Indicates that from X s To the source domain predicted value Y s Function mapping, f t Indicates that from X t To the target domain predicted value Y t Function mapping, δ s and δ t are the model parameters of the source domain and the target domain respectively, X s is the source domain data, X t is the target domain data; Among them, transfer learning is used to obtain the mapping f from the source domain s Migrate to the target domain, from the target domain data X t The mapping f of the target task is learned in t ; The mathematical expression of the multi-objective Harris Hawk optimization method is: In the formula, Indicates w i The estimated value of w i represents the weight of each basic model; e it Indicates the error value; represents the predicted value; y t represents the actual value, i is the number of typical wind farms selected, t is the number of iterations, is the predicted value of the i-th pre-trained model.

2. The fan fault diagnosis method based on CapBLSTM-MOHHO according to claim 1 is characterized in that: Acquiring the non-redundant characteristic data related to wind power, including: Based on the RRelief feature selection model, nonlinear feature selection is performed on the historical data collected by the SCADA system to obtain the non-redundant feature data related to wind power, wherein the non-redundant feature data related to wind power includes wind speed, wind direction, air pressure, temperature, rotor speed, gearbox oil temperature and impeller speed.

3. The fan fault diagnosis method based on CapBLSTM-MOHHO according to claim 2 is characterized in that: The method of nonlinear feature selection for data collected by the SCADA system based on the RRelief feature selection model is as follows: Where P diffF is the difference between samples in feature F in neighboring samples, P diffC is the difference between sample prediction values ​​in adjacent samples, P diffC|diffF is the degree of difference in sample prediction values ​​among neighboring samples that differ in feature F, is the feature weight set.

4. The fan fault diagnosis method based on CapBLSTM-MOHHO according to claim 1 is characterized in that: Preprocessing the non-redundant feature data, including abnormal data processing, missing value processing and data normalization processing; Wherein, the abnormal data processing is to eliminate the abnormal data through the box plot method; The missing value processing is to fill the missing values ​​using the mean value interpolation method and the linear interpolation method according to the degree of data missing; The data normalization process is as follows: In the formula, A max and A min Respectively represent the maximum and minimum values ​​of feature A, p i is the eigenvalue, p i ' is the normalized eigenvalue.

5. The fan fault diagnosis method based on CapBLSTM-MOHHO according to claim 1, characterized in that: The capsule-bidirectional long short-term memory network is used to extract the spatial feature relationship between the wind power and the optimal feature set at the same time point, and the capsule-bidirectional long short-term memory network includes: Convolutional layer: used to extract spatial features of input multidimensional time series; Main capsule layer: used to obtain the low-level features extracted by the convolutional layer; Digital capsule layer: used to extract high-level features of input multidimensional data; Bidirectional long short-term memory network: used to obtain the time feature relationship between the historical time points of the time series, where in the forward direction, the features from the past values ​​of wind power and the best feature set factors are stored through memory unit 1, and in the backward direction, the features from the future values ​​are stored through memory unit 2. The forward and backward hidden states are connected to obtain the final hidden output.

6. The fan fault diagnosis method based on CapBLSTM-MOHHO according to claim 1, characterized in that: Determining the operating state of the fan includes: If the residual between the predicted value and the actual value is within the alarm threshold range, there is no abnormality; if the residual between the predicted value and the actual value exceeds the alarm threshold, it is determined that an abnormality or fault occurs and an alarm signal is issued.

7. The fan fault diagnosis method based on CapBLSTM-MOHHO according to claim 6 is characterized in that: The alarm threshold is determined by an adaptive threshold method based on extreme values, wherein the adaptive threshold is: In the formula, R th is the obtained threshold, μ(t) is the mean, σ(t) is the standard deviation, λ is the confidence limit, and p is the probability of the cumulative distribution function.

Images