Wind turbine generator system whole system anomaly detection method and system based on auto-encoder

Abstract

The invention discloses a wind turbine generator system whole system anomaly detection method and system based on an auto-encoder. The method comprises the steps that state parameters of a wind turbine generator system are collected and preprocessed; state parameter correlation analysis and selection; constructing and pre-training a parameter reconstruction model based on an auto-encoder; evaluating wind turbine generators in similar states to train a parameter reconstruction model based on an auto-encoder; setting an anomaly detection dynamic threshold curve; and identifying abnormal parameters of the wind turbine generator. According to the method, related state parameters are analyzed to construct a wind turbine generator system whole-system parameter reconstruction model based on an auto-encoder, a dynamic detection threshold curve is constructed for each parameter by evaluating rich learning data of wind turbine generators in similar states, an abnormal parameter identification strategy is designed, and the accuracy of the abnormal parameter identification strategy is improved. And analyzing and identifying parameters which are most likely to cause abnormity according to the over-limit error. According to the invention, the accuracy of wind turbine generator anomaly detection is improved, and support is provided for operation and maintenance planning of the wind turbine generator.

Description

Technical Field

[0001] The present invention relates to the technical field of business process mining, and in particular to a method and system for detecting anomalies in the entire wind turbine system based on an autoencoder. Background Art

[0002] As non-renewable energy sources become increasingly scarce, global low-carbon policies are stimulating the development of new energy sources. As one of the renewable clean energy sources, wind energy resources have abundant total reserves on the earth. As the installed capacity of wind turbines continues to increase, efficient and intelligent operation and maintenance services have attracted widespread attention in the wind power industry. According to a 2017 survey by Accenture, the operation and maintenance costs of onshore wind power account for about 20% of the levelized cost of energy (LCOE) of wind power. In order to reduce operation and maintenance costs and avoid serious losses caused by major accidents, advanced condition monitoring and anomaly detection methods are urgently needed.

[0003] Condition monitoring is to monitor the operating status of each component by collecting wind turbine operating data in real time. According to different data sources, the research on wind turbine condition monitoring is divided into condition monitoring based on condition monitoring system (CMS) and condition monitoring based on supervisory control and data acquisition system (SCADA). Wind turbines are equipped with SCADA system by default, which remotely monitors and records the operating parameters, environmental parameters, fault shutdown and maintenance information of the core components of wind turbines in real time. There is no need to install additional measuring equipment, which reduces the operation and maintenance costs. Therefore, the use of SCADA condition monitoring data for wind turbine condition monitoring has aroused the research interest of practitioners. Existing research on wind turbine condition monitoring mainly focuses on fault diagnosis and anomaly detection. The basic idea of ​​fault diagnosis is to mine existing fault information from historical data and establish an identification model of related faults to help operation and maintenance personnel formulate reasonable maintenance plans in a timely manner. If anomalies can be found before the failure, maintenance can be arranged in advance to avoid the occurrence of failures. Abnormal detection of wind turbines usually establishes a fault warning model based on normal state information. Its purpose is to detect unit abnormalities before a fault occurs so as to arrange maintenance plans in advance to prevent the fault from occurring.

[0004] Existing anomaly detection methods mostly focus on monitoring a single component. In order to cover the entire wind turbine, a detection model needs to be equipped for each component. This will greatly reduce the efficiency of system analysis, bring additional computing costs, and may ignore the coupling relationship between components. In addition, due to the complex structure of wind turbines, it is often difficult to determine the type of fault in the early stage of a fault, which increases the workload of operation and maintenance personnel. Summary of the invention

[0005] The first purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and provide a method for detecting abnormalities in the entire system of a wind turbine based on an autoencoder. The method provides a full-system abnormality detection and abnormality identification analysis strategy for the wind turbine, analyzes relevant state parameters to construct a full-system parameter reconstruction model of the wind turbine based on the autoencoder, and enriches the learning data by evaluating wind turbines in similar states, constructs a dynamic threshold curve for abnormality detection for each parameter, designs an abnormal parameter identification strategy, analyzes and identifies the parameters most likely to cause abnormalities based on over-limit errors, and provides support for wind turbine operation and maintenance planning.

[0006] The second object of the present invention is to provide a wind turbine system-wide abnormality detection system based on an autoencoder.

[0007] The first object of the present invention is achieved by the following technical solution: a method for detecting abnormalities of the entire wind turbine system based on an autoencoder, comprising the following steps: 1) Wind turbine status parameter collection and preprocessing: Understand the types, operating data and sampling frequency of wind turbines in the wind farm, delete redundant parameters, expand time information, collect state parameter data of wind turbines during normal operation, and perform normalization processing; 2) State parameter correlation analysis and selection: Analyze the correlation between state parameters. Use the LightGBM-based parameter self-selection strategy to evaluate the correlation between all state parameters using the state parameter data preprocessed in step 1. Generate a correlation vector for each state parameter, cluster all state parameters, and select state parameters with obvious correlation based on the clustering results. 3) Build and pre-train a parameter reconstruction model based on an autoencoder: The state parameters selected in step 2) are used as input and output to construct a parameter reconstruction model based on an autoencoder, and the data of each wind turbine obtained in step 1) are used to train a parameter reconstruction model based on an autoencoder to obtain a pre-trained parameter reconstruction model; 4) Evaluate wind turbines in similar states to train a parameter reconstruction model based on an autoencoder: The pre-trained parameter reconstruction model obtained in step 3) is used to test the parameter reconstruction error of different wind turbine data, and wind turbines in similar states are judged according to the reconstruction error. Then, the data of all wind turbines in similar states are used to train the parameter reconstruction model based on the autoencoder. 5) Anomaly detection dynamic threshold curve setting: The absolute reconstruction error calculated by the parameter reconstruction model based on the autoencoder trained in step 4) is used as input, and the anomaly detection threshold is determined for each state parameter in different value ranges according to the confidence level. All thresholds are linearly interpolated to obtain the anomaly detection dynamic threshold curve of the state parameter; 6) Identification of abnormal parameters of wind turbines: The parameter reconstruction model based on the autoencoder and the anomaly detection dynamic threshold curve obtained in step 4) and step 5) are used to monitor the wind turbine operating data in real time. When an abnormality of the wind turbine is detected, the parameters most relevant to the abnormality are identified through state parameter reconstruction error and parameter correlation analysis.

[0008] Further, the step 1) includes the following steps: 1.1) Understand the wind turbine models, operating data and sampling frequency in the wind farm, and export the operating data in the wind power SCADA system according to the same model; 1.2) Perform a preliminary screening of the operating data exported in step 1.1), delete irrelevant or redundant parameters, expand the unit continuous operation time parameters according to the original data information, retain only the data when the unit is in the grid-connected operation state, delete the data at zero power and negative power, delete the data containing timestamps of missing values ​​and abnormal values, delete the first six timestamp data when the unit is initially started, that is, half an hour of operation information. At the initial start, the unit needs to adjust the cabin and blade angles and the internal temperature of the cabin due to its own state and environmental factors. After half an hour, the changes in all state parameters tend to be stable. Delete these data to avoid capturing transient information when the unit is initially started, and use the maximum-minimum normalization method to unify the order of magnitude of all state parameters.

[0009] Further, the step 2) includes the following steps: 2.1) Analyze the correlation between state parameters, and select the relevant parameter combination SS for each state parameter i using the LightGBM-based parameter self-selection strategy for the state parameter data preprocessed in step 1 i , and generate a correlation vector according to the degree of correlation; The LightGBM-based parameter self-selection strategy is a wrapper parameter selection method, which predicts the evaluated parameters by adjusting different combinations of other parameters in sequence, and selects the most relevant parameter combination for the evaluation parameters according to the prediction accuracy; the LightGBM is an efficient gradient boosting decision tree algorithm framework, which adopts a gradient-based one-sided sampling and mutually exclusive feature bundling method, which effectively improves the training speed of the algorithm while ensuring the prediction accuracy of the algorithm; 2.2) Based on the correlation vectors of all state parameters obtained in step 2.1), a density-based clustering algorithm is used to cluster all state parameters, and a combination of state parameters with obvious correlation is selected according to the clustering results; The density-based clustering algorithm divides data areas with high density into clusters, and can find clusters of arbitrary shapes in a noisy spatial database.

[0010] Further, in step 3), a parameter reconstruction model based on an autoencoder is constructed and pre-trained, the state parameters with a correlation selected in step 2) are used as input and output, an incomplete autoencoder is used to construct a parameter reconstruction model, and the data of each wind turbine obtained in step 1) is used to train a parameter reconstruction model based on an autoencoder respectively; The autoencoder is a type of neural network model used for unsupervised learning, which takes input information as a learning target and can be used for data dimension reduction, feature extraction and data reconstruction tasks; the incomplete autoencoder is a special autoencoder that requires the encoding dimension to be smaller than the input dimension and reduces noise in the data by limiting the encoding dimension.

[0011] Further, the step 4) comprises the following steps: 4.1) Evaluate wind turbines in similar states. Use the pre-trained parameter reconstruction model obtained in step 3) to test the parameter reconstruction error of other wind turbine data. According to a specified threshold, the wind turbines with a parameter reconstruction error less than the specified threshold are judged as wind turbines in similar states. 4.2) Fuse the state parameter data of wind turbines in similar states in step 4.1) and train a parameter reconstruction model based on an autoencoder.

[0012] Further, in step 5), setting a dynamic threshold curve for abnormality detection includes the following steps: 5.1) Each state parameter is initially divided into m uniform intervals according to its value range, and the intervals with small sample size are merged with the adjacent intervals; 5.2) Calculate an anomaly detection threshold for each interval of all state parameters divided in step 5.1) according to the probability density function and confidence level. Calculate the absolute reconstruction error of each state parameter according to the parameter reconstruction model based on the autoencoder trained in step 4). To determine the probability distribution of the absolute reconstruction error, use the kernel density estimation method to estimate the probability density function. , and its calculation formula is as follows: ; In the formula, n represents the number of samples, h represents the bandwidth, represents the kth sample value, K(·) represents the kernel function, and the Gaussian kernel function is used for calculation. The calculation formula is as follows: ; Combined with the probability density function, the anomaly detection threshold for each interval of all parameters is determined by a 99.99% confidence level; 5.3) Linearly interpolate the anomaly detection thresholds of all m intervals of each state parameter obtained in step 5.2), and the anomaly detection thresholds d between the j-1th interval and the jth interval adjacent to state parameter i j-1 and d j The linear interpolation function g i,j (a) The calculation formula is as follows: ; In the formula, a represents the state parameter in the interval [a j-1 , a j ], and all intervals are linearly interpolated to obtain the abnormal detection dynamic threshold curve of the state parameter, denoted as g i (a), i=1,2...49, i represents the state parameter number.

[0013] Further, in step 6), the wind turbine operation data is monitored in real time to detect abnormalities according to the autoencoder-based parameter reconstruction model obtained in step 4) and the abnormality detection dynamic threshold curve calculated in step 5), and the parameters most relevant to each abnormality are identified through parameter reconstruction error and parameter correlation analysis, including the following steps: 6.1) Detect abnormal state parameters. Reconstruct the state parameters of the wind turbine in real time according to the parameter reconstruction model based on the autoencoder obtained in step 4), and compare the parameter reconstruction error with the abnormal detection dynamic threshold curve calculated in step 5). When the average absolute reconstruction error value of the state parameter i in l consecutive time stamps is greater than the average value of l corresponding abnormal detection thresholds, it is recorded as a parameter over-limit warning. The ratio q of the average absolute reconstruction error to the average threshold is used to calculate the parameter over-limit warning. i,t Quantify the degree of overrun within l timestamps before time t. The calculation formula is as follows: ; In the formula, e i, t represents the reconstruction error of state parameter i at time t, b t represents the value of parameter i at time t, The anomaly detection threshold corresponding to the value of state parameter i at time t is represented. Adjacent over-limit records with a time interval of no more than 1 hour, i.e., 12 consecutive timestamps, are merged into one anomaly. At this time, state parameter i is recorded as an over-limit parameter. The sum of the ratio of the reconstruction error of each over-limit parameter in the adjacent over-limit records to the anomaly detection threshold is recorded as Q. i,T , according to Q i,T Sort the out-of-limit parameters in descending order to obtain the abnormal out-of-limit parameter set DS; 6.2) According to step 2.1) parameter correlation analysis, the relevant parameter set SS of each state parameter i is obtained i , related parameter set SS i Sort by descending order of relevance measure; 6.3) Calculate the abnormal over-limit parameter set DS obtained in step 6.1) and the relevant parameter set SS obtained in step 6.2) i The similarity of , first delete SS i The tail parameter makes len(SS i )≤len(DS), i∈[1, 49], where len(·) represents the number of parameters in the input set. Let the abnormal over-limit parameter set DS={p1, p2,..., p s , ..., p z}, where p z Represents the zth abnormal out-of-limit parameter. DS and SS are calculated based on the parameter ranking and the distance between the two sets. i Similarity score Score i , the calculation formula is as follows: ; ; ; In the formula, Returns a Boolean value representing the parameter Is it SS? i , z-s+1 means The higher the ranking of the score in DS, the stronger the relevance. express The distance score between the two sets, and Respectively In SS i and the ranking in DS, the state parameters are arranged in descending order according to the similarity score Score to obtain the similarity set RS; 6.4) Based on the overlap and ranking of the state parameters in the DS obtained in step 6.1) and the RS obtained in step 6.2), the top three state parameters most relevant to the abnormality are selected to provide support for maintenance personnel in formulating the operation and maintenance plan of the wind turbine.

[0014] The second object of the present invention is achieved by the following technical solution: a wind turbine system-wide abnormality detection system based on an autoencoder, used to implement the above-mentioned wind turbine system-wide abnormality detection method based on an autoencoder, which includes: The wind turbine state parameter acquisition and preprocessing module is used to understand the wind turbine model, operation data and sampling frequency in the wind farm, delete irrelevant or redundant parameters, expand the continuous operation time parameters of the unit according to the original data information, retain only the data of the unit in the grid-connected operation state, delete the data at zero power and negative power, delete the data containing the timestamps of missing values ​​and abnormal values, delete the first six timestamp data when the unit is initially started, that is, half an hour of operation information. When the unit is initially started, it needs to adjust the cabin and blade angle and adjust the internal temperature of the cabin due to its own state and environmental factors. After half an hour, the changes of all state parameters tend to be stable. These data are deleted to avoid capturing the transient information when the unit is initially started, and the maximum-minimum normalization method is used to unify the order of magnitude of all state parameters; The state parameter correlation analysis and selection module is used to analyze the correlation between state parameters. The preprocessed state parameter data is used to select relevant parameter combinations for each state parameter using the LightGBM-based parameter self-selection strategy, and a correlation vector is generated according to the degree of correlation. The density-based clustering algorithm is used to cluster all state parameter correlation vectors, and the state parameter combination with obvious correlation is selected according to the clustering results. Constructing and pre-training a parameter reconstruction model module, which is used to take the selected state parameters as input and output, use an incomplete autoencoder to construct a parameter reconstruction model, and use the obtained data of each wind turbine to train a parameter reconstruction model based on the autoencoder; The module for evaluating wind turbines in similar states and training parameter reconstruction models is used to evaluate wind turbines in similar states, use the pre-trained parameter reconstruction model to test the parameter reconstruction error of other wind turbine data, and judge the wind turbines with reconstruction errors less than the specified threshold as wind turbines in similar states according to the specified threshold. The state parameter data of wind turbines in similar states are integrated to train the parameter reconstruction model based on the autoencoder. The module for setting the dynamic threshold curve for abnormal detection is used to initially divide each state parameter into m uniform intervals according to the value range, merge the intervals with smaller sample sizes with adjacent intervals, calculate an abnormal detection threshold for each interval of all state parameters according to the probability density function and the confidence level, calculate the absolute reconstruction error of each state parameter by using the trained autoencoder-based parameter reconstruction model, estimate the probability density function by the kernel density estimation method to determine the probability distribution of the absolute reconstruction error, and determine the abnormal detection threshold of each interval of all parameters at a confidence level of 99.99% in combination with the probability density function, and perform linear interpolation on the abnormal detection threshold of all intervals of each state parameter; The abnormal parameter identification module of the wind turbine reconstructs the state parameters of the wind turbine in real time according to the parameter reconstruction model based on the autoencoder, and compares the parameter reconstruction error with the abnormal dynamic detection threshold. When it is detected that the average absolute reconstruction error value of the state parameter i in l consecutive time stamps is greater than the average threshold, it is recorded as a parameter over-limit warning, and the adjacent over-limit records within no more than 1 hour are merged into one abnormality. At this time, the state parameter i is recorded as the over-limit parameter, and the sum of the ratios of the reconstruction error of each over-limit parameter in the adjacent over-limit records to the abnormal detection threshold is recorded as Q i,T , according to Q i,T Sort the out-of-limit parameters in descending order to obtain the abnormal out-of-limit parameter set DS, and obtain the relevant parameter set SS of each state parameter i based on parameter correlation analysis i , calculate DS and SS i Similarity score Score i ,According to the similarity score, the state parameters are arranged in descending order to obtain the similarity set RS.,Combining the overlap and ranking of the state parameters in DS and RS, the top three state parameters most relevant to the anomaly are selected to provide support for maintenance personnel to formulate the operation and maintenance plan of the wind turbine.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. The present invention realizes for the first time the whole system abnormality detection concept of wind turbine in the early fault stage.

[0016] 2. The present invention realizes for the first time a state parameter selection method based on clustering of wind turbine state parameter correlation vectors.

[0017] 3. The present invention realizes for the first time an abnormal parameter identification strategy for abnormal detection of the entire wind turbine system.

[0018] 4. The present invention uses an incomplete autoencoder to construct a parameter reconstruction model, which has a certain noise reduction capability for noisy data.

[0019] 5. The present invention realizes the establishment of a dynamic threshold curve for anomaly detection by dividing the intervals using reconstruction errors, thereby improving the accuracy of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The figure is a logical flow diagram of the method of the present invention.

[0021] Figure 2 Schematic diagram of the system of the present invention.

[0022] Figure 3 This is a correlation diagram of wind turbine state parameters.

[0023] Figure 4 Structural diagram of the model for parameter reconstruction.

[0024] Figure 5 It is a dynamic threshold curve diagram of the cabin angle.

[0025] Figure 6 Anomaly detection visualization for failures. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0027] Example 1 like Figure 1 As shown, this embodiment discloses a method for detecting abnormalities in the entire wind turbine system based on an autoencoder, comprising the following steps: 1) Wind turbine status parameter collection and preprocessing: 1.1) Understand the wind turbine models, operating data and sampling frequency in the wind farm, and export the operating data in the wind power SCADA system according to the same model; The SCADA data collected in this step comes from a wind farm in Hebei, China. The rated power of the wind farm is 100MW. 41 2.0 MW wind turbines of the same model are selected as the research objects, numbered WT1 to WT41. The data sampling frequency is 5 minutes, including 56 parameter variables including the operation data, environmental data and operation status of the wind turbines.

[0028] 1.2) Perform a preliminary screening of the operating data exported in step 1.1), delete irrelevant or redundant parameters, expand the "unit continuous operation time" parameter according to the original data information, retain only the data in the "grid-connected operation" state of the unit, delete the data at zero power and negative power, delete the data containing timestamps of missing values ​​and abnormal values, delete the first six timestamp data at the initial start of the unit, that is, half an hour of operating information. At the initial start, the unit needs to adjust the cabin and blade angles and the internal temperature of the cabin due to its own state and environmental factors. After half an hour, the changes in all state parameters tend to be stable. Delete these data to avoid capturing transient information at the initial start of the unit, and use the maximum-minimum normalization method to unify the order of magnitude of all state parameters.

[0029] In this step, 49 relevant parameters are finally obtained, as shown in Table 1.

[0030] Table 1 Wind turbine parameters 2) State parameter correlation analysis and selection: 2.1) Analyze the correlation between state parameters, and select the relevant parameter combination SS for each state parameter i using the LightGBM-based parameter self-selection strategy for the state parameter data preprocessed in step 1i , and generate a correlation vector according to the degree of correlation; The parameter self-selection strategy based on LightGBM is a wrapper-type parameter selection method, which predicts the evaluated parameters by adjusting different combinations of other parameters in sequence, and selects the most relevant parameter combination for the evaluated parameters according to the prediction accuracy; the LightGBM is an efficient gradient boosting decision tree algorithm framework, which adopts gradient-based one-sided sampling and mutually exclusive feature bundling methods, which effectively improves the training speed of the algorithm while ensuring the prediction accuracy of the algorithm.

[0031] In this step, 49 state parameters are visualized according to the correlation vectors, such as Figure 3 As shown, it can be seen that the correlation between these 49 parameters is relatively close.

[0032] 2.2) Based on the correlation vectors of all state parameters obtained in step 2.1), a density-based clustering algorithm is used to cluster all state parameters, and a combination of state parameters with obvious correlation is selected based on the clustering results.

[0033] The density-based clustering algorithm divides data regions with sufficiently high density into clusters, and can find clusters of arbitrary shapes in a noisy spatial database.

[0034] In this step, three density clustering algorithms, Mean Shift, OPTICS (Ordering Points To Identify the Clustering Structure), and DBSCAN (Density-Based Spatial Clustering of Applications with Noise), which do not require the specification of the number of clusters, are used for clustering. The results all cluster 49 parameters into one category. The analysis results show that the parameters are coupled with each other and are difficult to simply decompose.

[0035] 3) The state parameters with correlation selected in step 2) are used as input and output, and an undercomplete autoencoder is used to construct a parameter reconstruction model. The data of each wind turbine obtained in step 1) is used to train a parameter reconstruction model based on an autoencoder.

[0036] The autoencoder is a type of neural network model used for unsupervised learning, which uses input information as a learning target and can be used for tasks such as data dimensionality reduction, feature extraction, and data reconstruction; the incomplete autoencoder is a special autoencoder that requires the encoding dimension to be smaller than the input dimension, and reduces noise in the data by limiting the encoding dimension.

[0037] In this step, an incomplete autoencoder is used to construct a parameter reconstruction model structure as follows Figure 4 shown.

[0038] 4) Evaluate wind turbines in similar states to train a parameter reconstruction model based on an autoencoder: 4.1) Evaluate wind turbines in similar states. Use the pre-trained parameter reconstruction model obtained in step 3) to test the parameter reconstruction error of other wind turbine data. According to a specified reconstruction error threshold, the wind turbines with reconstruction errors less than the specified threshold are judged as wind turbines in similar states. In this step, the normal operation data of WT17, which has not experienced any failure for one year, is used to train the reconstruction model to test the reconstruction error of 41 wind turbines. It is observed that the reconstruction mean absolute error (MAE) of 35 wind turbines is concentrated between 0.17 and 0.45, while the MAE values ​​of the remaining wind turbines are significantly larger, so it is determined that the status of these 35 wind turbines is similar.

[0039] 4.2) Fuse the state parameter data of wind turbines in similar states in step 4.1) and train a parameter reconstruction model based on an autoencoder.

[0040] In this step, the WT17 data, the data of the six wind turbines with the highest similarity, and the data of all similar wind turbines are used to train three models, denoted as AE_WT17, AE_WT17_6, and AE_WT17_35. Table 2 lists the test results of the three models on the six wind turbines, among which the AE_WT17 model has the largest reconstruction error and the AE_WT17_35 model has the smallest reconstruction error, indicating that more data contains richer working conditions.

[0041] Table 2 Comparison of MAE of reconstruction models of different wind turbine data In this step, Table 3 shows the comparison results between the proposed method and the baseline method. ED represents the anomaly detection method that uses the Euclidean distance to calculate the error between the input vector and the reconstructed vector, and MD represents the anomaly detection method that uses the Mahalanobis distance to calculate the error between the input vector and the reconstructed vector. MD_EWMA represents the method of detecting anomalies by calculating the exponentially weighted moving average (EWMA) based on the Mahalanobis distance. Single represents the anomaly detection method that monitors each parameter separately, and Single_Dynamic refers to the anomaly detection method that builds a dynamic threshold curve based on Single. According to the experimental results, compared with the existing anomaly detection methods, the two methods of calculating the error of each parameter separately, Single and Single_Dynamic, can detect anomalies more comprehensively and reduce false alarms as much as possible. The Single_Dynamic method achieved the highest Recall value, which shows that the setting of the dynamic threshold curve can further improve the accuracy of detecting anomalies.

[0042] Table 3 Comparison results of anomaly detection accuracy 5) Set the dynamic threshold curve for anomaly detection: 5.1) Each state parameter is initially divided into m uniform intervals according to its value range, and the intervals with smaller sample sizes are merged with adjacent intervals; 5.2) Calculate an anomaly detection threshold for each interval of all state parameters divided in step 5.1) according to the probability density function and confidence level. Calculate the absolute reconstruction error of each state parameter according to the parameter reconstruction model based on the autoencoder trained in step 4). To determine the probability distribution of the absolute reconstruction error, use the kernel density estimation (KDE) method to estimate the probability density function. , and its calculation formula is as follows: ; In the formula, n represents the number of samples, h represents the bandwidth, represents the kth sample value, K(·) represents the kernel function, and the Gaussian kernel function is used for calculation. The calculation formula is as follows: ; Combined with the probability density function, the anomaly detection threshold for each interval of all parameters is determined by a 99.99% confidence level; 5.3) Linearly interpolate the anomaly detection thresholds of all m intervals of each state parameter obtained in step 5.2), and the anomaly detection thresholds d between the j-1th interval and the jth interval adjacent to state parameter i j-1 and d j The linear interpolation function g i,j (a) The calculation formula is as follows: ; In the formula, a represents the state parameter in the interval [a j-1 , a j ], and all intervals are linearly interpolated to obtain the abnormal detection dynamic threshold curve, denoted as g i (a), i=1,2...49, i represents the state parameter number.

[0043] In this step, according to the experimental debugging, the parameter value range is initially divided into m = 10 parts, and the detection threshold of each part is calculated at a confidence level of 99.99%. Taking the state parameter "cabin angle" as an example, its abnormal detection dynamic threshold curve is as follows: Figure 5 shown.

[0044] 6) Based on the autoencoder-based parameter reconstruction model obtained in step 4) and the abnormality detection dynamic threshold curve calculated in step 5), the wind turbine operating data is monitored in real time to detect abnormalities, and the parameters most relevant to each abnormality are identified through parameter reconstruction error and parameter correlation analysis, including the following steps: 6.1) Detect abnormal state parameters. Reconstruct the state parameters of the wind turbine in real time according to the parameter reconstruction model based on the autoencoder obtained in step 4), and compare the parameter reconstruction error with the abnormal detection dynamic threshold curve calculated in step 5). When the average absolute reconstruction error value of the state parameter i in l consecutive time stamps is greater than the average value of l corresponding abnormal detection thresholds, it is recorded as a parameter over-limit warning. The ratio q of the average absolute reconstruction error to the average threshold is used to calculate the parameter over-limit warning. i,t Quantify the degree of overrun within l timestamps before time t. The calculation formula is as follows: ; In the formula, e i, t represents the reconstruction error of state parameter i at time t, b t represents the value of parameter i at time t, The anomaly detection threshold corresponding to the value of state parameter i at time t is represented. Adjacent over-limit records with a time interval of no more than 1 hour, i.e., 12 consecutive timestamps, are merged into one anomaly. At this time, state parameter i is recorded as an over-limit parameter. The sum of the ratio of the reconstruction error of each over-limit parameter in the adjacent over-limit records to the anomaly detection threshold is recorded as Q. i,T , according to Q i,TSort the out-of-limit parameters in descending order to obtain the abnormal out-of-limit parameter set DS; 6.2) According to step 2.1) parameter correlation analysis, the relevant parameter set SS of each state parameter i is obtained i , related parameter set SS i Sort by descending order of relevance measure; 6.3) Calculate the abnormal over-limit parameter set DS obtained in step 6.1) and the relevant parameter set SS obtained in step 6.2) i The similarity of , first delete SS i The tail parameter makes len(SS i )≤len(DS), i∈[1, 49], where len(·) represents the number of parameters in the input set. Let the abnormal over-limit parameter set DS={p1, p2,..., p s ,..., p z}, where p z Represents the zth abnormal out-of-limit parameter, and calculates DS and SS based on the parameter ranking and the distance between the two sets. i Similarity score Score i , the calculation formula is as follows: ; ; ; In the formula, Returns a Boolean value representing the parameter Is it SS? i , z-s+1 means The higher the ranking of the score in DS, the stronger the relevance. express The distance score between the two sets, and Respectively In SS i and the ranking in DS, the state parameters are arranged in descending order according to the similarity score Score to obtain the similarity set RS; 6.4) Based on the overlap and ranking of the state parameters in the DS obtained in step 6.1) and the RS obtained in step 6.2), the top three state parameters most relevant to the abnormality are selected to provide support for maintenance personnel in formulating the operation and maintenance plan of the wind turbine.

[0045] In this step, a pitch system failure is taken as an example to show the abnormal parameter identification results. Figure 6As shown in Table 4. It can be seen from Table 4 that the unit first detected an abnormality about 2000 data points (at least 10 days) before the 12076 fault occurred. The variable pitch communication failure was caused by the damage of the communication line of the central control box, which made it impossible to correctly transmit key information to the blade angle adjustment system. An abnormal temperature in the central control box was detected at the 2477th data point, which means that the control module may have a fault and cannot adjust the fan status in time, causing the operating load to increase and cause abnormal temperature of the electronic control equipment. An abnormal temperature of the blade 2 motor was detected at the 3065th data point, which means that the fault occurred in the blade 2 motor or the blade system. Figure 6 It can be seen that Single_Dynamic improves the ability to capture minor anomalies by establishing dynamic detection thresholds on each parameter, and reduces the false alarms caused by data imbalance to a certain extent.

[0046] Table 4 Results of abnormal parameter identification strategy Example 2 This embodiment discloses a wind turbine system-wide abnormality detection system based on an autoencoder, which is used to implement the wind turbine system-wide abnormality detection method based on an autoencoder described in Example 1. Figure 2 As shown, the system includes the following functional modules: The wind turbine state parameter acquisition and preprocessing module is used to understand the wind turbine model, operation data and sampling frequency in the wind farm, delete irrelevant or redundant parameters, expand the continuous operation time parameters of the unit according to the original data information, retain only the data of the unit in the grid-connected operation state, delete the data at zero power and negative power, delete the data containing the timestamps of missing values ​​and abnormal values, delete the first six timestamp data when the unit is initially started, that is, half an hour of operation information. When the unit is initially started, it needs to adjust the cabin and blade angle and adjust the internal temperature of the cabin due to its own state and environmental factors. After half an hour, the changes of all state parameters tend to be stable. These data are deleted to avoid capturing the transient information when the unit is initially started, and the maximum-minimum normalization method is used to unify the order of magnitude of all state parameters; The state parameter correlation analysis and selection module is used to analyze the correlation between state parameters. The preprocessed state parameter data is used to select relevant parameter combinations for each state parameter using the LightGBM-based parameter self-selection strategy, and a correlation vector is generated according to the degree of correlation. The density-based clustering algorithm is used to cluster all state parameter correlation vectors, and the state parameter combination with obvious correlation is selected according to the clustering results. Constructing and pre-training a parameter reconstruction model module, which is used to take the selected state parameters as input and output, use an incomplete autoencoder to construct a parameter reconstruction model, and use the obtained data of each wind turbine to train a parameter reconstruction model based on the autoencoder; The module for evaluating wind turbines in similar states and training parameter reconstruction models is used to evaluate wind turbines in similar states, use the pre-trained parameter reconstruction model to test the parameter reconstruction error of other wind turbine data, and judge the wind turbines with reconstruction errors less than the specified threshold as wind turbines in similar states according to the specified threshold. The state parameter data of wind turbines in similar states are integrated to train the parameter reconstruction model based on the autoencoder. The module for setting the dynamic threshold curve for abnormal detection is used to initially divide each state parameter into m uniform intervals according to the range of its reconstruction error, merge the intervals with smaller sample sizes with adjacent intervals, calculate an abnormal detection threshold for each interval of all state parameters according to the probability density function and the confidence level, calculate the absolute reconstruction error of each state parameter by using the trained autoencoder-based parameter reconstruction model, estimate the probability density function by the kernel density estimation method to determine the probability distribution of the absolute reconstruction error, and determine the abnormal detection threshold of each interval of all parameters at a confidence level of 99.99% in combination with the probability density function, and perform linear interpolation on the abnormal detection threshold of all intervals of each state parameter; The abnormal parameter identification module of the wind turbine reconstructs the state parameters of the wind turbine in real time according to the parameter reconstruction model based on the autoencoder, and compares the parameter reconstruction error with the abnormal dynamic detection threshold. When it is detected that the average absolute reconstruction error value of the state parameter i in l consecutive time stamps is greater than the average threshold, it is recorded as a parameter over-limit warning, and the adjacent over-limit records within no more than 1 hour are merged into one abnormality. At this time, the state parameter i is recorded as the over-limit parameter, and the sum of the ratios of the reconstruction error of each over-limit parameter in the adjacent over-limit records to the abnormal detection threshold is recorded as Q i,T , according to Q i,T Sort the out-of-limit parameters in descending order to obtain the abnormal out-of-limit parameter set DS, and obtain the relevant parameter set SS of each state parameter i based on parameter correlation analysis i , calculate DS and SS i Similarity score Score i ,According to the similarity score, the state parameters are arranged in descending order to obtain the similarity set RS.,Combining the overlap and ranking of the state parameters in DS and RS, the top three state parameters most relevant to the anomaly are selected to provide support for maintenance personnel to formulate the operation and maintenance plan of the wind turbine; The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A wind turbine system-wide anomaly detection method based on an autoencoder, characterized in that: The following steps are involved: 1) Wind turbine status parameter collection and preprocessing: Understand the types, operating data and sampling frequency of wind turbines in the wind farm, delete redundant parameters, expand time information, collect state parameter data of wind turbines during normal operation, and perform normalization processing; 2) State parameter correlation analysis and selection: Analyze the correlation between state parameters. Use the LightGBM-based parameter self-selection strategy to evaluate the correlation between all state parameters using the state parameter data preprocessed in step 1. Generate a correlation vector for each state parameter, cluster all state parameters, and select state parameters with obvious correlation based on the clustering results. 3) Build and pre-train a parameter reconstruction model based on an autoencoder: The state parameters selected in step 2) are used as input and output to construct a parameter reconstruction model based on an autoencoder, and the data of each wind turbine obtained in step 1) are used to train a parameter reconstruction model based on an autoencoder to obtain a pre-trained parameter reconstruction model; 4) Evaluate wind turbines in similar states to train a parameter reconstruction model based on an autoencoder: The pre-trained parameter reconstruction model obtained in step 3) is used to test the parameter reconstruction error of different wind turbine data, and wind turbines in similar states are judged according to the reconstruction error. Then, the data of all wind turbines in similar states are used to train the parameter reconstruction model based on the autoencoder. 5) Anomaly detection dynamic threshold curve setting: The absolute reconstruction error calculated by the parameter reconstruction model based on the autoencoder trained in step 4) is used as input, and the anomaly detection threshold is determined for each state parameter in different value ranges according to the confidence level. All thresholds are linearly interpolated to obtain the anomaly detection dynamic threshold curve of the state parameter; 6) Identification of abnormal parameters of wind turbines: The parameter reconstruction model based on the autoencoder and the anomaly detection dynamic threshold curve obtained in step 4) and step 5) are used to monitor the wind turbine operating data in real time. When an abnormality of the wind turbine is detected, the parameters most relevant to the abnormality are identified through state parameter reconstruction error and parameter correlation analysis.

2. The method for detecting abnormality of the entire wind turbine system based on an autoencoder according to claim 1, characterized in that: The step 1) comprises the following steps: 1.1) Understand the wind turbine models, operating data and sampling frequency in the wind farm, and export the operating data in the wind power SCADA system according to the same model; 1.2) Perform a preliminary screening of the operating data exported in step 1.1), delete irrelevant or redundant parameters, expand the unit continuous operation time parameters according to the original data information, retain only the data when the unit is in the grid-connected operation state, delete the data at zero power and negative power, delete the data containing timestamps of missing values ​​and abnormal values, delete the first six timestamp data when the unit is initially started, that is, half an hour of operation information. At the initial start, the unit needs to adjust the cabin and blade angles and the internal temperature of the cabin due to its own state and environmental factors. After half an hour, the changes in all state parameters tend to be stable. Delete these data to avoid capturing transient information when the unit is initially started, and use the maximum-minimum normalization method to unify the order of magnitude of all state parameters.

3. The method for detecting abnormality of the entire wind turbine system based on an autoencoder according to claim 2 is characterized in that: The step 2) comprises the following steps: 2.1) Analyze the correlation between state parameters, and select the relevant parameter combination SS for each state parameter i using the parameter self-selection strategy based on LightGBM for the state parameter data preprocessed in step 1 i , and generate a correlation vector according to the degree of correlation; The LightGBM-based parameter self-selection strategy is a wrapper parameter selection method, which predicts the evaluated parameters by adjusting different combinations of other parameters in sequence, and selects the most relevant parameter combination for the evaluation parameters according to the prediction accuracy; the LightGBM is an efficient gradient boosting decision tree algorithm framework, which adopts a gradient-based one-sided sampling and mutually exclusive feature bundling method, which effectively improves the training speed of the algorithm while ensuring the prediction accuracy of the algorithm; 2.2) Based on the correlation vectors of all state parameters obtained in step 2.1), a density-based clustering algorithm is used to cluster all state parameters, and a combination of state parameters with obvious correlation is selected according to the clustering results; The density-based clustering algorithm divides data areas with high density into clusters, and can find clusters of arbitrary shapes in a noisy spatial database.

4. The method for detecting abnormality of the entire wind turbine system based on an autoencoder according to claim 3 is characterized in that: In step 3), a parameter reconstruction model based on an autoencoder is constructed and pre-trained, the state parameters with correlation selected in step 2) are used as input and output, an incomplete autoencoder is used to construct the parameter reconstruction model, and the data of each wind turbine obtained in step 1) is used to train a parameter reconstruction model based on an autoencoder respectively; The autoencoder is a type of neural network model used for unsupervised learning, which takes input information as a learning target and can be used for data dimension reduction, feature extraction and data reconstruction tasks; the incomplete autoencoder is a special autoencoder that requires the encoding dimension to be smaller than the input dimension and reduces noise in the data by limiting the encoding dimension.

5. The method for detecting abnormality of the entire wind turbine system based on an autoencoder according to claim 4 is characterized in that: The step 4) comprises the following steps: 4.1) Evaluate wind turbines in similar states. Use the pre-trained parameter reconstruction model obtained in step 3) to test the parameter reconstruction error of other wind turbine data. According to a specified threshold, the wind turbines with a parameter reconstruction error less than the specified threshold are judged as wind turbines in similar states. 4.2) Fuse the state parameter data of wind turbines in similar states in step 4.1) and train a parameter reconstruction model based on an autoencoder.

6. The method for detecting abnormality of the entire wind turbine system based on an autoencoder according to claim 5, characterized in that: In step 5), setting the anomaly detection dynamic threshold curve includes the following steps: 5.1) Each state parameter is initially divided into m uniform intervals according to its value range, and the intervals with small sample size are merged with the adjacent intervals; 5.2) Calculate an anomaly detection threshold for each interval of all state parameters divided in step 5.1) according to the probability density function and confidence level. Calculate the absolute reconstruction error of each state parameter according to the parameter reconstruction model based on the autoencoder trained in step 4). To determine the probability distribution of the absolute reconstruction error, use the kernel density estimation method to estimate the probability density function. , and its calculation formula is as follows: ； In the formula, n represents the number of samples, h represents the bandwidth, represents the kth sample value, K(·) represents the kernel function, and the Gaussian kernel function is used for calculation. The calculation formula is as follows: ； Combined with the probability density function, the anomaly detection threshold for each interval of all parameters is determined by a 99.99% confidence level; 5.3) Linearly interpolate the anomaly detection thresholds of all m intervals of each state parameter obtained in step 5.2), and the anomaly detection thresholds d between the j-1th interval and the jth interval adjacent to state parameter i j-1 and d j The linear interpolation function g i,j (a) The calculation formula is as follows: ； In the formula, a represents the state parameter in the interval [a j-1 , a j ], and all intervals are linearly interpolated to obtain the abnormal detection dynamic threshold curve of the state parameter, denoted as g i (a), i=1,2...49, i represents the state parameter number.

7. The method for detecting abnormality of the entire wind turbine system based on an autoencoder according to claim 6, characterized in that: In step 6), the wind turbine operating data is monitored in real time to detect anomalies according to the autoencoder-based parameter reconstruction model obtained in step 4) and the abnormality detection dynamic threshold curve calculated in step 5), and the parameters most relevant to each anomaly are identified through parameter reconstruction error and parameter correlation analysis, including the following steps: 6.1) Detect abnormal state parameters. Reconstruct the state parameters of the wind turbine in real time according to the parameter reconstruction model based on the autoencoder obtained in step 4), and compare the parameter reconstruction error with the abnormal detection dynamic threshold curve calculated in step 5). When the average absolute reconstruction error value of the state parameter i in l consecutive time stamps is greater than the average value of l corresponding abnormal detection thresholds, it is recorded as a parameter over-limit warning. The ratio q of the average absolute reconstruction error to the average threshold is used to calculate the parameter over-limit warning. i,t Quantify the degree of overrun within l timestamps before time t. The calculation formula is as follows: ； In the formula, e i,t represents the reconstruction error of state parameter i at time t, b t represents the value of parameter i at time t, The anomaly detection threshold corresponding to the value of state parameter i at time t is represented. Adjacent over-limit records with a time interval of no more than 1 hour, i.e., 12 consecutive timestamps, are merged into one anomaly. At this time, state parameter i is recorded as an over-limit parameter. The sum of the ratio of the reconstruction error of each over-limit parameter in the adjacent over-limit records to the anomaly detection threshold is recorded as Q. i,T , according to Q i,T Sort the out-of-limit parameters in descending order to obtain the abnormal out-of-limit parameter set DS; 6.2) According to step 2.1) parameter correlation analysis, the relevant parameter set SS of each state parameter i is obtained i , related parameter set SS i Sort by descending order of relevance measure; 6.3) Calculate the abnormal over-limit parameter set DS obtained in step 6.1) and the relevant parameter set SS obtained in step 6.2) i The similarity of , first delete SS i The tail parameter makes len(SS i )≤len(DS), i∈[1, 49], where len(·) represents the number of parameters in the input set. Let the abnormal over-limit parameter set DS={p1, p2,..., p s ,..., p z }, where p z Represents the zth abnormal out-of-limit parameter. DS and SS are calculated based on the parameter ranking and the distance between the two sets. i Similarity score Score i , the calculation formula is as follows: ； ； ； In the formula, Returns a Boolean value representing the parameter Is it SS? i , z-s+1 means The higher the ranking of the score in DS, the stronger the relevance. express The distance score between the two sets, and Respectively In SS i and the ranking in DS, the state parameters are arranged in descending order according to the similarity score Score to obtain the similarity set RS; 6.4) Based on the overlap and ranking of the state parameters in the DS obtained in step 6.1) and the RS obtained in step 6.2), the top three state parameters most relevant to the abnormality are selected to provide support for maintenance personnel in formulating the operation and maintenance plan of the wind turbine.

8. The wind turbine system-wide abnormality detection system based on autoencoder is characterized by: The method for detecting abnormalities of the entire wind turbine system based on an autoencoder according to any one of claims 1 to 7 comprises: The wind turbine state parameter acquisition and preprocessing module is used to understand the wind turbine model, operation data and sampling frequency in the wind farm, delete irrelevant or redundant parameters, expand the continuous operation time parameters of the unit according to the original data information, retain only the data of the unit in the grid-connected operation state, delete the data at zero power and negative power, delete the data containing the timestamps of missing values ​​and abnormal values, delete the first six timestamp data when the unit is initially started, that is, half an hour of operation information. When the unit is initially started, it needs to adjust the cabin and blade angle and adjust the internal temperature of the cabin due to its own state and environmental factors. After half an hour, the changes of all state parameters tend to be stable. These data are deleted to avoid capturing the transient information when the unit is initially started, and the maximum-minimum normalization method is used to unify the order of magnitude of all state parameters; The state parameter correlation analysis and selection module is used to analyze the correlation between state parameters. The preprocessed state parameter data is used to select relevant parameter combinations for each state parameter using the LightGBM-based parameter self-selection strategy, and a correlation vector is generated according to the degree of correlation. The density-based clustering algorithm is used to cluster all state parameter correlation vectors, and the state parameter combination with obvious correlation is selected according to the clustering results. Constructing and pre-training a parameter reconstruction model module, which is used to take the selected state parameters as input and output, use an incomplete autoencoder to construct a parameter reconstruction model, and use the obtained data of each wind turbine to train a parameter reconstruction model based on the autoencoder; The module for evaluating wind turbines in similar states and training parameter reconstruction models is used to evaluate wind turbines in similar states, use the pre-trained parameter reconstruction model to test the parameter reconstruction error of other wind turbine data, and judge the wind turbines with reconstruction errors less than the specified threshold as wind turbines in similar states according to the specified threshold. The state parameter data of wind turbines in similar states are integrated to train the parameter reconstruction model based on the autoencoder. The module for setting the dynamic threshold curve for abnormal detection is used to initially divide each state parameter into m uniform intervals according to the value range, merge the intervals with smaller sample sizes with adjacent intervals, calculate an abnormal detection threshold for each interval of all state parameters according to the probability density function and the confidence level, calculate the absolute reconstruction error of each state parameter by using the trained autoencoder-based parameter reconstruction model, estimate the probability density function by the kernel density estimation method to determine the probability distribution of the absolute reconstruction error, and determine the abnormal detection threshold of each interval of all parameters at a confidence level of 99.99% in combination with the probability density function, and perform linear interpolation on the abnormal detection threshold of all intervals of each state parameter; The abnormal parameter identification module of the wind turbine reconstructs the state parameters of the wind turbine in real time according to the parameter reconstruction model based on the autoencoder, and compares the parameter reconstruction error with the abnormal dynamic detection threshold. When it is detected that the average absolute reconstruction error value of the state parameter i in l consecutive time stamps is greater than the average threshold, it is recorded as a parameter over-limit warning, and the adjacent over-limit records within no more than 1 hour are merged into one abnormality. At this time, the state parameter i is recorded as the over-limit parameter, and the sum of the ratios of the reconstruction error of each over-limit parameter in the adjacent over-limit records to the abnormal detection threshold is recorded as Q i,T , according to Q i,T Sort the out-of-limit parameters in descending order to obtain the abnormal out-of-limit parameter set DS, and obtain the relevant parameter set SS of each state parameter i based on parameter correlation analysis i , calculate DS and SS i Similarity score Score i ,According to the similarity score, the state parameters are arranged in descending order to obtain the similarity set RS.,Combining the overlap and ranking of the state parameters in DS and RS, the top three state parameters most relevant to the anomaly are selected to provide support for maintenance personnel to formulate the operation and maintenance plan of the wind turbine.

Images