Wind turbine gearbox anomaly detection method considering multi-unit operating state similarity

Abstract

A wind turbine gearbox anomaly detection method considering the similarity of multi-unit operating states, based on the piecewise linear time series similarity evaluation method, evaluates a single state variable; considering the similarity of multiple state variables, the wind turbine operating state spatiotemporal similarity quantification method evaluates the similarity of each wind turbine and the operating state of the unit to be detected; a state estimation LSTM model is constructed, different LSTM models are trained; the accuracy and adaptability of different LSTM model performances are verified; select several LSTM models that perform well in performance verification, construct a combined state estimation model CPEM through weighted combination; estimate the target variable of the unit to be detected, calculate the residual error between the estimated value and the true value, and based on the residual effective value comparison and residual information entropy comparison, identify the anomaly. This method detects the operating state of the gearbox based on the historical operating data of the wind turbine, without the need for maintenance personnel to conduct on-site detection, reducing the operating and maintenance costs of the wind turbine.

Description

Technical Field

[0001] This invention relates to the field of gearbox operation anomaly detection technology for wind turbine generator sets, and specifically to a method for detecting gearbox anomalies in wind turbine generator sets that takes into account the similarity of operating states of multiple units. Background Technology

[0002] Large wind turbine generators (hereinafter referred to as "wind turbines") are complex electromechanical hybrid systems composed of multiple subsystems. They convert wind energy into electrical energy through the mutual coupling and synergistic effect of these subsystems. Because wind turbines operate in harsh environments such as high temperature, high humidity, salinity, and sandstorms, coupled with the randomness of wind speed and direction and the uncertainty of load, they are more prone to failure than traditional thermal / hydropower generators. Furthermore, due to differences in operating mechanisms, manufacturing materials, and the degree of external interference, the failure rates and the extent of their impact vary significantly among the subsystems. Among these, the gearbox, as the core of energy transmission in a wind turbine, can severely impact the overall power generation efficiency of the unit and even lead to prolonged shutdowns if it malfunctions. Therefore, taking effective measures to reduce the gearbox failure rate is crucial for improving the stable operation of the unit and increasing the economic benefits of wind farms.

[0003] Currently, most wind farms use periodic inspections to monitor the health status of various subsystems, including the gearbox. Since the gearbox of a wind turbine is a closed subsystem, on-site maintenance personnel often rely on personal experience or simple auxiliary equipment to assess its condition. For example, they might listen for unusual noises inside the gearbox to determine if the internal equipment is malfunctioning; or observe the oil pressure and connections for signs of oil leakage to assess oil pressure and airtightness. This method often only detects gearbox abnormalities or malfunctions with obvious symptoms, making it difficult to detect early abnormalities and minor faults. If minor gearbox faults are not detected and repaired in time, they can easily trigger a chain reaction of failures, resulting in significant losses. For example, early tooth root cracks in the gearbox, if not addressed promptly, can gradually deteriorate into severe tooth breakage, potentially triggering a chain reaction of failures within the gearbox and jeopardizing the normal operation of the entire transmission system; excessive rolling bearing temperatures in the gearbox, if not addressed in time, can lead to a serious fire and destroy the entire turbine. Therefore, it is necessary to find more reliable and sensitive gearbox condition assessment methods to detect abnormal gearbox conditions in a timely manner, so as to take effective measures to avoid failure.

[0004] With the widespread adoption of sensors and the rise of data mining technology, data-driven methods for diagnosing wind turbine gearbox faults have gradually attracted attention. These methods are based on data collected by the wind turbine's Supervisory Control and Data Acquisition (SCADA) or Condition Monitoring System (CMS), using locally or remotely deployed computers as analysis tools. Machine learning or other intelligent algorithms are employed to extract useful information from the data to assess the gearbox's operating status. This approach can detect abnormal gearbox operating conditions before the wind turbine's condition monitoring system triggers a fault alarm, providing maintenance personnel with more time to troubleshoot and eliminate faults in their early stages. Compared to currently used gearbox anomaly detection methods, data-driven methods offer higher accuracy, greater flexibility, and lower technical barriers, making them potentially widely applicable and widespread.

[0005] The accuracy of data-driven wind turbine gearbox anomaly detection methods is generally positively correlated with the amount of historical data used for model training. However, wind turbines operate in harsh environments for extended periods, and sensor failures, signal transmission anomalies, and periodic memory updates limit the amount of historical data that can be acquired, making it difficult to guarantee sufficient data for model training. Furthermore, many anomaly detection methods assume that the historical data used for model training is data monitored during normal wind turbine operation. However, actual wind turbines may have been operating abnormally for some time, meaning the monitored historical data may already contain anomalies. When there is insufficient historical data for model training or the historical data used to train the model itself contains anomalies, the reliability of the anomaly detection results will inevitably be affected. Therefore, addressing the issues of insufficient historical data and anomaly-containing training data in terms of the accuracy and reliability of gearbox anomaly detection results is of great significance. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for detecting gearbox anomalies in wind turbines that considers the similarity of operating states across multiple units. This method detects the gearbox's operating status based on historical operating data of the wind turbines, eliminating the need for maintenance personnel to conduct on-site inspections. This provides a basis for maintenance personnel to develop overhaul and maintenance plans, significantly reducing the operation and maintenance costs of wind turbines.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for detecting gearbox anomalies in wind turbine units considering the similarity of operating states of multiple units includes the following steps:

[0009] Step 1: Evaluate a single state variable based on a piecewise linearized time series similarity assessment method;

[0010] Step 2: Consider the spatiotemporal similarity quantification method of wind turbine operating state considering multiple state variables. Evaluate the similarity of the operating state of each wind turbine in the same wind farm with the turbine under test, and select data from multiple turbines with strong similarity in operating state with the turbine under test.

[0011] Step 3: Preprocess the historical data of the unit to be tested and the historical data of the selected unit to be tested in Step 2, and divide them into training data, test data and detection data according to time period;

[0012] Step 4: Construct a state estimation LSTM model. Using the training data of the selected similar units and the training data of the unit to be detected, train multiple different state estimation LSTM models.

[0013] Step 5: Set comprehensive evaluation metrics and use test data to verify the accuracy and adaptability of the LSTM model performance for different state estimation modes;

[0014] Step 6: Select several state estimation LSTM models that have performed well in the performance verification in Step 5, and construct a combined state estimation model (CPEM) by weighted combination.

[0015] Step 7: Based on the combined state estimation model in Step 6, estimate the target variables of the unit to be tested, and calculate the residual between the estimated value and the true value. Based on the comparison of the effective value of the residual and the comparison of the information entropy of the residual, perform anomaly identification.

[0016] Step 1 includes the following steps:

[0017] S1.1: Suppose there are two wind turbine state variable time series curves, L1 and L2, each containing L sampling points;

[0018] S1.2: A linear segmentation method is used to divide a long time series into different short time series. After linear segmentation, the state variable time series L1 and L2 can be approximately represented by h (h>2) segments and k (k>2) segments of broken lines, where the breakpoints are taken as the points of abrupt change in values ​​on the two curves. The broken lines after segmentation have only three trends: rising, falling, and remaining unchanged. These three trends are represented by 1, -1, and 0, respectively. Therefore, the state variable time series curves can be approximately represented by the datasets S1 and S2.

[0019]

[0020] In equation (1), t represents the time corresponding to each segmentation point; m represents the trend of the segmented broken line, and its value set is {1, -1, 0}; Δy is the difference between the normalized value at the end of each broken line and the normalized value at the beginning, which reflects the amplitude change of the broken line. The superscripts s1 and s2 are used to distinguish different broken lines, and the subscripts k and h identify the k-th and h-th broken line segments, respectively.

[0021] Elements in a set and Let represent the h-th linearized piecewise linear segment of curve L1 and the k-th linearized piecewise linear segment of curve L2, respectively. The sets of times corresponding to their linearization cutoff points are denoted as […]. and in: This represents the time corresponding to the h-th segment point of the time series piecewise linear line L1. This represents the time corresponding to the kth segment point of the time series piecewise linear curve L2.

[0022] S1.3: Based on the first linearization segmentation in S1.2, a second linearization segmentation is performed to ensure that the linearization segmentation points of curves L1 and L2 are exactly the same; the second segmentation point set T is obtained by... s1 and T s2 The union of the subsets, arranged chronologically according to the time of each partition point, is expressed as:

[0023] T = sort(T) s1 ∪T s2 )={t0,t1,t2,…t l-1 ,t l},(l≥h,l≥k) (2);

[0024] In equation (2): sort() is the calculation of sorting values ​​from smallest to largest; the symbol ∪ represents the union of sets; l represents the number of division points contained in each curve after the second division; h and k represent the number of division points contained in curves L1 and L2 after the first linearization division, respectively; t l Represents the time corresponding to the l-th segmentation point; the L1 and L2 linearized segmentation sets of the original time series curve after the second linearization segmentation. and Represented as:

[0025]

[0026] In equation (3): t0, t1, ..., t l-1 ,t l These represent the times corresponding to different segmentation points. This indicates that curve L1, after the second linearization segmentation, is determined by t. l-1 and t lThe transformation trend of a line segment is determined by two points. The amplitude trend is This indicates that curve L2, after the second linearization segmentation, is determined by t. l-1 and t l The transformation trend of a line segment is determined by two points. The amplitude trend is

[0027] At this point, the time series curve of the state variable is divided into multiple line segments of equal number and with completely consistent dividing points.

[0028] S1.4: The similarity S of the state variables according to equation (4) ST Perform the calculation:

[0029]

[0030] In equation (4): S ST The value of is in the range of [0,1], and the larger the value, the greater the similarity between the two state variables.

[0031] L represents the total number of linear segments, and i is the iteration count identifier. This represents the trend of change of the i-th line segment after L1 linearization of the curve. This represents the trend of the i-th line segment after L2 linearization of the curve. This represents the magnitude of the change in the i-th line segment after linearizing the curve L1. This represents the magnitude of the change in the i-th line segment after L2 linearization of the curve. This represents the trend distance of the corresponding line segment after linearization. This represents the magnitude distance of the corresponding line segment after linearization.

[0032] Step 2 includes the following steps:

[0033] S2.1: Select multiple state variables that can comprehensively reflect the similarity between the macroscopic and microscopic states of the wind turbine, and calculate the similarity of each state variable according to equation (5);

[0034]

[0035] In equation (5): S a It is the area of ​​the shaded region. Let S0 be the area of ​​the regular polygon representing the similarity of the i-th state variable, and n be the number of selected state variables. For the similarity of the first state variable, The similarity of the (i+1)th state variable, The similarity is for the (n+1)th state variable.

[0036] S2.2: Based on the radar image area comparison method, the spatiotemporal similarity of the wind turbine operating status is quantified;

[0037] S2.2.1: First, construct a radar chart with the same number of regular polygons as the state variables selected in S2.1. Define each axis of the radar chart as a state variable, and set the base length of the axis to 1.

[0038] S2.2.2: Next, starting from the center of the radar chart of the regular polygon, draw arrows outward along the axis. The length of the arrows represents the similarity of the state variables corresponding to the two wind turbines.

[0039] S2.2.3: Finally, connect all the arrow endpoints in sequence to form a closed figure, namely the shaded part of the radar chart. The proportion of the area of ​​this closed figure to the area of ​​the radar chart represents the spatiotemporal similarity of the operating states of the two wind turbine units.

[0040] Step 4 includes the following steps:

[0041] S4.1: Construct a state estimation model for the target variable using LSTM:

[0042] LSTM contains a forget gate f t Input gate i t and output gate O t Among them, the forgetting gate f t This determines which hidden state information passed from the previous unit needs to be forgotten in the new unit; input gate i t Used to determine what information the cell needs to store; Output gate O t This is used to determine which information needs to be output to the next unit from the current hidden state; the state update process is represented as:

[0043]

[0044] In equation (6): W f and b f These are the weights and biases of the forget gate, respectively; W i and b i These are the weights and bias terms of the input gate, respectively; W o and b o These are the weights and bias terms of the output gate, respectively; W c and b c These represent the weights and biases of the current state, respectively; g(,) denotes the gate function; ⊙ denotes the element-wise multiplication of the matrix; h t-1 This represents the hidden layer variable at time t-1, which has short-term memory capabilities. C t-1 and Ct X represents the cell layer variables at time t-1 and time t, respectively, and it has long-term memory function. t This represents the input variable at time t.

[0045] S4.2: A two-layer LSTM stacking mode is used to construct the state estimation model of the wind turbine. Multiple state estimation LSTM models are trained using SCADA data from the wind turbine under test and other wind turbines with similar operating states. The LSTM model with the best performance is then used to construct the combined state estimation model (CSEM). In step 5, the performance testing of different state estimation LSTM models is performed in two steps: first, the accuracy of the state estimation LSTM model's estimation results is tested based on the test data of each wind turbine; second, the adaptability of the corresponding state estimation LSTM model to the test data of the wind turbine under test is tested using the test data of the wind turbine under test.

[0046] In step 5, the performance evaluation criterion for all state estimation LSTM models is measured by a comprehensive score index β, which is calculated as follows:

[0047]

[0048] In equation (7): y i and represents the i-th true value and the estimated value of the target variable in the validation dataset, respectively; N is the total number of samples in the validation dataset; To verify the mean of the target variable in the dataset; E rmse and E mae Rm and Rab represent the root mean square and absolute mean of the residuals, respectively. The smaller their values, the more accurate the model estimate. 2 The decision coefficient reflects the goodness of fit of the model to the data. Its value ranges from [0,1]. The larger the value, the stronger the interpretability of the model to the data. According to equation (7), the comprehensive score index β is related to the above-mentioned E rmse E mae and R 2 The three indicators are closely related, and the larger the value, the better the overall performance of the model.

[0049] In step 6, the combined state estimation model CSEM is constructed using a weighted combination method, as shown in equation (8):

[0050]

[0051] In equation (8): M i This represents the i-th LSTM sub-model; m is the number of selected LSTM sub-models; M CSEMRepresents CSEM; γ i The similarity quantification value of the wind turbine is calculated based on equation (5). The combined weight is related to the spatiotemporal similarity of the wind turbine's operating state. That is, the more similar the operating state of the wind turbine in the same wind farm is to the turbine under test, the greater the contribution of the model trained based on the turbine's data to CSEM.

[0052] M0 represents the LSTM model trained based on the historical data of the unit under test, with γ0 = 1. When the training data of the unit under test is insufficient or contains many outliers, it is difficult to obtain a reliable M0. In this case, the value of γ0 is set to 0, indicating that M0 does not participate in the construction of CSEM, but only relies on the data of wind turbines with similar operating conditions to achieve the identification of anomalies in the unit under test.

[0053] In step 7, two indicators are used to determine whether the residuals are abnormal: the first indicator is the effective value R of the sample residuals. m R is used to measure the overall bias of sample residuals. m The higher the value, the higher the probability of detecting anomalies in the sample; another indicator is the residual information entropy E. n E is used to assess the degree of uncertainty regarding the occurrence of anomalies in a sample. n The smaller the value, the higher the probability that the sample is abnormal.

[0054] Step 7 includes the following steps:

[0055] S7.1: By using sliding window sampling, the long-term series of the target variable of the unit to be tested is divided into multiple short-term series for anomaly detection. The R-value of each short-term series is... m and E n The value is calculated according to equation (9):

[0056]

[0057] In equation (9): I represents the number of statistical intervals divided based on the residual distribution; d i p represents the number of residual samples falling within the i-th statistical interval; i This represents the proportion of samples within the corresponding statistical interval to the total tested samples; α determines the unit of measurement for information entropy, typically taken as the natural logarithm 'e'; T inv It is the sampling interval of the SCADA system; T r It refers to the window width used in the sliding window, which reflects the time scale of the detection.

[0058] S7.2: Set three statistical intervals: normal region, risk region, and high-risk region to calculate the information entropy for each detection period. The three statistical intervals are not continuous, and their boundary values ​​are based on the four important parameters Q of the box plot. low Q1, Q3 and Qup Determined, where Q1 and Q3 are the 0.25 and 0.75 quantiles of the box plot, respectively; Q up and Q low These are the upper and lower boundaries in the box plot, respectively, and their values ​​are calculated using equation (10).

[0059]

[0060] When the residual value is greater than Q1 but less than Q3, the residual falls within the normal range; when the residual value is less than Q... low or greater than Q up When the residual falls in the high-risk area, all other values ​​fall in the risk area, and the residuals in the high-risk area deviate significantly from the overall normal distribution range.

[0061] S7.3: Anomaly Detection Process: First, calculate the effective value and information entropy of the residual samples during the detection period; second, compare the effective value of the residuals during the detection period to ensure they do not exceed the threshold H. rm If the residuals do not exceed the threshold, it indicates that the residual distribution is within the normal range. If the effective value exceeds the range, it indicates that the sample residuals are too large. In this case, it is necessary to further determine whether this anomaly is caused by a few extremely large noises or by a large number of abnormal residuals by using the information entropy of the residuals. If the information entropy of the residuals is also less than the given threshold H, en If the effective value is too large, it is very likely due to abnormal operation of the object being detected; otherwise, it is due to noise interference.

[0062] This invention provides a method for detecting gearbox anomalies in wind turbine units that considers the similarity of operating states across multiple units. The technical advantages are as follows:

[0063] 1) This method improves the accuracy of gearbox anomaly detection for the target unit by making full use of SCADA data from other wind turbine units within the same wind farm.

[0064] 2) This invention utilizes historical SCADA data from wind turbine generators to detect abnormal conditions in the gearbox of the target unit, eliminating the need for on-site inspections by maintenance personnel. This method can identify anomalies in the gearbox oil temperature and internal bearing temperature, thus detecting potential gearbox malfunctions and providing a basis for maintenance personnel to develop repair and maintenance plans, significantly reducing the operation and maintenance costs of wind turbine generators. This method is particularly suitable for the abnormal inspection of wind turbine generators in remote areas and offshore wind turbine generators.

[0065] 3) This invention makes full use of data from all wind turbines in the same wind farm that are similar in operation to the target unit, thus expanding the scale of available data. It can effectively solve the problem in actual engineering where insufficient historical data of the unit to be tested or the unit itself has anomalies, which leads to the inability to accurately train the state estimation model.

[0066] 4) This invention proposes a state variable similarity comparison method based on piecewise linearization. By considering the trend changes and the average magnitude changes of all time series segments after piecewise linearization, this method can more accurately quantify the differences between the same variables of different wind turbines, overcoming the shortcomings of existing state variable similarity assessment methods that only consider numerical feature similarity and ignore time synchronization.

[0067] 5) This invention proposes a spatiotemporal similarity quantification method for wind turbine operating states that comprehensively considers the similarity of macroscopic and microscopic states. This method can objectively quantify the degree of similarity between the operating states of different turbines. 6) To identify abnormal states of the target variable, this invention proposes an abnormal data identification method based on the comparison of effective residual values ​​and residual information. This method can reliably distinguish between actual gearbox operating abnormalities and data anomalies caused by accidental factors such as noise, thereby improving the reliability of anomaly detection. Attached Figure Description

[0068] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0069] Figure 1 This is a flowchart of the anomaly detection process of the present invention;

[0070] Figure 2 This is a time series curve of the state variables of a wind turbine.

[0071] Figure 3 This is a schematic diagram illustrating the spatiotemporal similarity quantification of unit operating status based on the radar image area comparison method.

[0072] Figure 4(a) is a schematic diagram of the basic structure of RNN;

[0073] Figure 4(b) is a schematic diagram of the basic structure of LSTM.

[0074] Figure 5 This is a schematic diagram of the stacked LSTM model structure.

[0075] Figure 6 This is a flowchart for LSTM performance testing and CSEM construction.

[0076] Figure 7(a) shows the original route map of the statistical interval division;

[0077] Figure 7(b) shows the range of information entropy values;

[0078] Figure 7(c) is a flowchart of anomaly identification.

[0079] Figure 8 This is a statistical chart showing the importance coefficients of several key variables and the front bearing temperature of the gearbox, obtained based on the recursive feature elimination method.

[0080] Figure 9(a) shows the spatiotemporal similarity quantification results of the operating states of each unit in wind farm WF1 and WT0;

[0081] Figure 9(b) is a radar map showing the similarity quantification between the wind farm WT0 and some of its turbine units.

[0082] Figure 10 This is a comparison chart of CSEM estimation results based on different schemes in WF1.

[0083] Figure 11(a) shows the anomaly detection results based on different schemes in WF1 (Sch1);

[0084] Figure 11(b) shows the anomaly detection results based on different schemes in WF1 (Sch2);

[0085] Figure 11(c) shows the anomaly detection results based on different schemes in WF1 (Sch3).

[0086] Figure 12(a) shows the results of the spatiotemporal similarity quantification of the operating status of each unit in WF2 and WT0;

[0087] Figure 12(b) is a radar diagram showing the similarity quantification between each unit in WF2 and WT0.

[0088] Figure 13(a) shows the anomaly detection results based on different schemes in WF2 (Sch1);

[0089] Figure 13(b) shows the anomaly detection results based on different schemes in WF2 (Sch2);

[0090] Figure 13(c) shows the anomaly detection results based on different schemes in WF2 (Sch3).

[0091] Figure 14(a) shows the performance test results based on the WF0 validation dataset in WF1;

[0092] Figure 14(b) shows the performance test results based on the WF0 validation dataset in WF2.

[0093] Figure 15(a) shows the anomaly detection results based on the WF0 detection dataset in WF1;

[0094] Figure 15(b) shows the anomaly detection results based on the WF0 detection dataset in WF2.

[0095] Figure 16(a) shows the anomaly detection results of different methods under the condition of abnormal training data (comparison of estimated value and true value); Figure 16(b) shows the anomaly detection results of different methods under the condition of abnormal training data (identified anomalies). Detailed Implementation

[0096] This paper proposes a method for detecting gearbox anomalies in wind turbines considering the similarity of operating states across multiple units. First, a piecewise linearized time-series similarity quantification method is proposed. Then, a spatiotemporal similarity quantification method for wind turbine operating states, comprehensively considering the similarity of macroscopic and microscopic variables across different wind turbines, is designed. This method can quantify the similarity of the operating states of the target turbine with other turbines within the same wind farm. Next, historical data from multiple turbines with strong similarity to the target turbine are selected to train several different Long Short-Term Memory (LSTM) network models. After verifying the accuracy and adaptability of each model to the target turbine data using test data, the LTM network model with good accuracy and adaptability is selected to construct a Combined State Estimation Model (CPEM). This model can reliably estimate the target variables of the turbine under test. Finally, to verify whether the target variable data is abnormal, an anomaly data identification method based on a combination of residual effective values ​​and residual information entropy is proposed.

[0097] Figure 1 The flowchart of the anomaly detection method of this invention includes three main modules: variable selection and evaluation of the similarity of wind turbine operating states, training and performance evaluation of different LSTM models and construction of combined models, and anomaly detection. Specifically, it can be implemented in 9 steps:

[0098] Step 1: Based on professional knowledge, select multiple state variables for spatiotemporal similarity assessment of wind turbine operating status. On this basis, combine the feature selection algorithm: recursive feature elimination method, select the variable that best reflects the gearbox operating status as the target variable (the output variable of the model), and select the variables closely related to it as the input variables of the model.

[0099] Step 2: Considering the characteristics of the state variables monitored by the SCADA system of wind turbine generators and the shortcomings of current commonly used methods, a time series similarity evaluation method based on piecewise linearization is proposed for the evaluation of single state variables, as follows:

[0100] The foundation of wind turbine operating state similarity assessment lies in the similarity of key state variables. Therefore, it is essential to first objectively and accurately assess the similarity of individual state variables. Based on the data characteristics collected by the SCADA system, it is known that the state variables of wind turbines are stored in time series form. Thus, the collected data possesses both numerical and temporal attributes. However, most existing methods for assessing the similarity of individual state variables utilize the statistical or distributional characteristics of the numerical values ​​for similarity comparison. While these methods can reflect the spatial similarity of corresponding state variables using numerical characteristics, they neglect the temporal synchronicity of these state variables. Therefore, this invention proposes a wind turbine state variable similarity assessment method based on piecewise linearized comparison.

[0101] Suppose we have two wind turbine state variable time series curves, L1 and L2, each containing L sampling points, as shown in the figure. Figure 2 The blue and green dashed lines in the diagram illustrate this. To quantify their similarity, a linear segmentation method is first used to divide the long-term time series into different short-term time series. After linear segmentation, L1 and L2 can be approximated by h (h>2) segments and k (k>2) segments of broken lines, where the breakpoints can be taken as the points of abrupt change in value on the two curves, such as... Figure 2 As shown by the red and yellow solid lines in the image. From Figure 2 As can be seen from this, the broken lines after these segments exhibit only three trends: rising, falling, and remaining unchanged. If we use 1, -1, and 0 to represent these three trends respectively, then the original curve can be approximated by the datasets S1 and S2:

[0102]

[0103] In equation (1): t represents the time corresponding to each segmentation point; m represents the trend of the segmented broken line, with a value set of {1, -1, 0}; Δy is the difference between the normalized value at the end and the normalized value at the beginning of each broken line, reflecting the amplitude change of the broken line; superscripts s1 and s2 are used to distinguish different broken lines; subscripts k and h identify the k-th and h-th broken line segments, respectively. The elements in the set... and Let represent the h-th linearized piecewise linear segment of curve L1 and the k-th linearized piecewise linear segment of curve L2, respectively. The sets of times corresponding to their linearization cutoff points can be denoted as […]. and

[0104] Since the first linearization segmentation is based on the variation characteristics of each curve, the segmentation points in the two linearized datasets S1 and S2 cannot be guaranteed to be completely identical. This is not conducive to the comparison of time series. Therefore, a second linearization segmentation is needed based on the first linearization segmentation to ensure that the linearized segmentation points of curves L1 and L2 are completely identical. The second segmentation point set T is obtained by analyzing T... s1 and T s2 The union of the subsets, arranged chronologically according to the time of each partition point, is expressed as:

[0105] T = sort(T) s1 ∪T s2 )={t0,t1,t2,…t l-1 ,t l},(l≥h,l≥k) (2);

[0106] In equation (2): sort() is used to sort the values ​​in ascending order, and the symbol ∪ represents the union of the sets. The linearized L1 and L2 sets of the original curves after the second segmentation. and It can be represented as:

[0107]

[0108] At this point, the original time series is divided into multiple line segments of equal number with identical dividing points. Then, the similarity S of the variables can be further calculated using equation (4). ST Perform the calculation:

[0109]

[0110] In equation (4): S ST The value of is in the range of [0,1], and the larger the value, the greater the similarity between the two variables. It can be observed that the state variable similarity assessment method proposed in this invention uses piecewise linearization to differentiate the comparison of the original long-term series, and assesses the overall similarity of the wind turbine state variables by the average magnitude of the trend changes and amplitude changes of all linearized line segments. Based on this method, assessing the similarity of state variables avoids the shortcomings of statistical features and spatial distance comparisons, which cannot take into account temporal synchronization. Therefore, the similarity comparison results of wind turbine state variables obtained using this method are more reliable.

[0111] Step 3: Based on the method in Step 2 above, a spatiotemporal similarity quantification method for wind turbine operating states that comprehensively considers the similarity of multiple state variables is proposed. This method evaluates the similarity of the operating states of each turbine in the same wind farm with the turbine under test, and selects data from multiple turbines with strong similarity to the turbine under test to expand the data scale for anomaly detection, as detailed below:

[0112] To accurately measure the spatiotemporal similarity of wind turbine operating states, it is necessary to comprehensively consider the similarity of multiple state variables, rather than judging solely by the similarity of a single variable. Therefore, this invention further proposes a method for quantifying the spatiotemporal similarity of wind turbine operating states that comprehensively considers both macroscopic and microscopic state similarities. Macroscopic state similarity refers to the similarity of the wind turbine's operating environment and energy conversion efficiency; microscopic variables refer to the similarity of the operating states of specific monitored objects or subsystems, such as gearboxes and generators. The selection of variables reflecting macroscopic state similarity largely relies on professional knowledge, such as common parameters like wind speed, main bearing speed, and output power. However, when selecting variables reflecting microscopic state similarity, professional knowledge and data correlation analysis results can be combined to choose variables that directly or indirectly reflect the operating state of the monitored object. Typically, the required microscopic variables differ depending on the monitored object.

[0113] Multiple variables that can comprehensively reflect the similarity of the macroscopic and microscopic states of wind turbines are selected. The similarity of each state variable can be calculated according to equation (5). Based on this, a method based on radar chart area comparison is designed to quantify the spatiotemporal similarity of the wind turbine's operating state. The specific quantification principle is as follows: Figure 3 As shown. First, construct a regular polygon radar chart with the same number of selected state variables. Define each axis of the radar chart as a state variable, and set the baseline length of the axis to 1. Second, draw blue arrows outward from the center of the polygon along the axes. The length of the arrow represents the similarity of the state variables corresponding to the two wind turbines. Finally, connect the endpoints of all the blue arrows sequentially to form a closed figure, i.e. Figure 3 The shaded area on the radar chart, representing the proportion of the shaded region to the total radar chart area, indicates the spatiotemporal similarity of the operating states of the two wind turbines. The corresponding similarity quantification formula is:

[0114]

[0115] In equation (5): S a It is the area of ​​the shaded region. Let S0 be the area of ​​the regular polygon representing the similarity of the i-th state variable, and n be the number of selected state variables. This method comprehensively considers the temporal and spatial similarity of multiple state variables of the wind turbine, and can obtain a relatively reliable quantification value of the spatiotemporal similarity of the operating state.

[0116] Step 4: Perform data preprocessing on the historical data of the unit to be tested and the historical data of the selected unit to maintain the synchronization of the data of the above units with the data of the unit to be tested in terms of sampling time, and divide them into training data, test data and detection data according to time period.

[0117] Step 5: Based on the Long Short Term Memory (LSTM) network algorithm, train multiple different LSTM models using the training data of the selected similar units and the training data of the unit to be tested. If the historical data of the unit to be tested is insufficient, the historical data of the unit to be tested can be omitted from the model training.

[0118] Step 6: Set comprehensive evaluation indicators and verify the performance of different LSTM models in two steps using test data: the first step is to verify the accuracy of the LSTM model obtained by training the unit using its own test data, and the second step is to verify the adaptability of the model to the test unit data using the test data of the unit to be tested.

[0119] Step 7: Select several LSTM models that perform well in both performance tests, and construct a Combination State Estimation Model (CPEM) by weighted combination.

[0120] The specific technical solutions for steps five through seven are as follows:

[0121] 1) State estimation LSTM model:

[0122] Many state variables of wind turbine generators exhibit short-term dependencies. To better uncover the short-term dependencies of variables themselves and the conditional dependencies between variables, this invention employs LSTM to construct a state estimation model for the target variable. LSTM is a deep learning algorithm derived from Recurrent Neural Networks (RNNs). It not only possesses the memory characteristics and parameter sharing features of RNNs, but also effectively solves the gradient vanishing and gradient exploding problems that may occur during RNN training by adding internal control gate structures. Therefore, it has a greater advantage when learning features with time attributes.

[0123] Figures 4(a) and 4(b) illustrate the basic structural units of RNN and LSTM. In these figures, C represents the unit state, h represents the hidden state, X represents the input information, and the subscripts t and t-1 are used to distinguish between the current unit and the previous unit. As can be seen from Figures 4(a) and 4(b), the basic unit of an RNN contains only a single neural layer, and only one state flow C is transmitted between units. However, in an LSTM, both state flows C and h are transmitted between units, and the LSTM contains four interconnected neural network layers. The three additional layers are essentially three special control gate structures: the forget gate f... t Input gate i t and output gate O t The forget gate determines which hidden state information passed from the previous unit needs to be forgotten in the new unit; the input gate determines which information the current cell needs to store; and the output gate determines which information the current hidden state needs to be output to the next unit. The state update process within the control gates can be represented as follows:

[0124]

[0125] In equation (6): W f and b f These are the weights and biases of the forget gate, W. i and b i These are the weights and biases of the input gate, W. o and b o These are the weights and biases of the output gate, W. c and b c These are the weights and biases of the current state, respectively. g(,) represents the gate function, and ⊙ represents the element-wise multiplication of the corresponding matrix elements.

[0126] In addition to short-term dependence, the state variables of wind turbine generators also exhibit significant nonlinear characteristics. To better learn the nonlinear time series features, this invention employs a stacked LSTM layer approach to construct a state estimation model for the target variable. Stacking LSTMs increases the model's depth, thereby effectively improving its nonlinear representation capability; however, this also increases computational and memory consumption. Therefore, this invention uses a two-layer stacking mode to construct the wind turbine generator state estimation model, thus balancing learning capability and efficiency. The structure of the stacked model is as follows: Figure 5 As shown, multiple LSTM models are trained using SCADA data from the wind turbine to be tested and other turbines with similar operating conditions. The LSTM model with the best performance is then used to construct the final CSEM.

[0127] 2) Performance evaluation of LSTM and construction of CSEM:

[0128] Only LSTMs that simultaneously meet the requirements of accuracy and adaptability can be used to build the final CSEM, and the process is as follows: Figure 6 As shown, the performance testing of these different LSTM models is divided into two steps: first, the accuracy of the corresponding LSTM model estimation results is tested based on the validation dataset of each unit; then, the adaptability of these models to the data of the unit under test is tested using the validation dataset of the unit under test. The performance evaluation criterion for all models is measured by a comprehensive score index β, the calculation formula of which is:

[0129]

[0130] In equation (7): y i and Let be the true value and the estimated value of the target variable in the validation dataset, respectively, and N be the total number of samples in the validation dataset. To verify the mean of the target variable in the dataset, E rmse and E mae Rm and Rab represent the root mean square and absolute mean of the residuals, respectively. The smaller their values, the more accurate the model estimate. 2 is the decision coefficient, which reflects the goodness of fit of the model to the data. Its value ranges from [0,1]. The larger the value, the stronger the interpretability of the model to the data. According to equation (7), the comprehensive score index β is closely related to the above three indices. The larger its value, the better the comprehensive performance of the model. Therefore, the LSTM with relatively better performance can be selected to construct the final CSEM based on the scores of each LSTM model in the accuracy test and the adaptability test.

[0131] CSEM is constructed using a weighted combination method, as shown in equation (8):

[0132]

[0133] In equation (8): M i This represents the i-th LSTM sub-model, where m is the number of sub-models selected, and M... CSEM Represents CSEM, γ i The similarity quantification value of the wind turbine is calculated based on equation (5). It can be found that the combined weight of the proposed method is related to the spatiotemporal similarity of the wind turbine's operating state. That is, the more similar the operating states of the turbines in the same wind farm are to the turbine under test, the greater the contribution of the model trained based on the turbine data to CSEM.

[0134] It should be noted that M0 represents the LSTM model trained based on the historical data of the unit under test, with γ0 = 1. When the training data of the unit under test is insufficient or contains many outliers, it is difficult to obtain a reliable M0. In this case, the value of γ0 can be set to 0, indicating that M0 does not participate in the construction of CSEM, but only relies on the data of wind turbines with similar operating conditions to achieve the identification of anomalies in the unit under test.

[0135] Step 8: Estimate the target variable of the unit to be tested based on CPEM, and calculate the residual between the estimated value and the true value.

[0136] Step 9: In order to eliminate the interference of noise and random factors and accurately identify abnormal residuals, an abnormal data detection method based on the comparison of residual effective values ​​and residual information entropy is further proposed. Based on this method, abnormal residuals caused by abnormal equipment operation can be effectively identified.

[0137] Anomalies in the target variable can cause significant changes in the residual values ​​between its true and estimated values, altering the information entropy of the residual sequence. Therefore, this invention proposes an anomaly detection method that combines comparison of residual effective values ​​and residual information entropy. This method uses two indicators to determine whether the residuals are anomalies; the first indicator is the effective value R of the sample residuals. m R is used to measure the overall bias of sample residuals. m The higher the value, the higher the probability of detecting anomalies in the sample. Another indicator is the information entropy E of the residual. n It can be used to assess the degree of uncertainty regarding the occurrence of anomalies in a sample. n The smaller the value, the higher the probability that the detected sample contains anomalies. To improve detection accuracy, this invention uses sliding window sampling to divide the long-term series of the target variable into multiple short time-segment series for anomaly detection. The R-value of each short series is... m and E n The value can be calculated according to equation (9):

[0138] In equation (9): I is the number of statistical intervals divided based on the residual distribution, d i p represents the number of residual samples falling within the i-th statistical interval. i To determine the proportion of samples within the corresponding statistical interval to the total tested samples, α determines the unit of measurement for information entropy. Typically, α is taken as the natural logarithm 'e', ​​where T is the value of the sample. inv T is the sampling interval of the SCADA system. r It refers to the window width used in the sliding window, which reflects the time scale of the detection.

[0139] This invention sets three statistical intervals to calculate the information entropy for each detection period. These statistical intervals are determined by a box plot of the residuals of all target variables obtained from the validation dataset. Figure 7(a) marks these three statistical intervals with different colors, representing the normal region, the risk region, and the high-risk region, respectively. It can be observed that these three statistical intervals are not continuous, and their boundary values ​​are based on four important parameters of the box plot, namely Q. low Q1, Q3 and Q up It is confirmed that Q1 and Q3 are the 0.25 and 0.75 quantiles of the box plot, respectively. up and Q low These are the upper and lower boundaries in the box plot, respectively, and their values ​​can be calculated using equation (10). When the residual value is greater than Q1 but less than Q3, the residual falls within the normal range; when the residual value is less than Q... low or greater than Q up When the residuals fall within the high-risk region, all other values ​​fall within the risk region, with the residuals in the high-risk region significantly deviating from the normal distribution range of the population. When a large number of samples fall within the normal and risk regions, their effective values ​​will not deviate significantly from the population. However, when a high proportion of samples fall within the high-risk region or there are a few extremely large outliers, the effective values ​​of the samples will deviate significantly from the population.

[0140] Figure 7(b) shows the proportion of the test sample residuals falling within these three statistical regions. i The relationship between (i=1,2,3) and the information entropy of the corresponding detection sample residuals is shown in Figure 7(b). It is easy to see that when a large proportion of samples fall in the same interval, the information entropy of the sample residuals is very small, which indicates that the probability of random events occurring is very high.

[0141] Based on the above inferences, the anomaly detection process shown in Figure 7(c) is proposed. First, the effective value and information entropy of the residual samples during the detection period are calculated respectively. Second, the effective value of the residuals during the detection period is compared to the threshold H. rm If the residuals do not exceed the threshold, it indicates that the residual distribution is within the normal range. If the effective value exceeds the range, it indicates that the sample residuals are too large. In this case, it is necessary to further determine whether this anomaly is caused by a few extremely large noises or by a large number of abnormal residuals by using the information entropy of the residuals. If the information entropy of the residuals is also less than the given threshold H, en If the effective value is too large, it is very likely due to abnormal operation of the object being detected; otherwise, it is due to noise interference.

[0142] The method presented in this section employs a simple and effective approach to set H. rm and H enTo determine the value of the target variable, firstly, the residual during normal operation is calculated using the validation dataset of the unit under test; then, the residual sequence is divided into different detection segments using sliding window sampling; finally, the R-value of each segment is calculated. m and E n Value, take all detected segments R m and E n The maximum values ​​are respectively used as H rm and H en The value of .

[0143] Step 10: Verification by Example:

[0144] The data in this case study comes from two onshore wind farms in City a of Province A and central Province B. SCADA data from 33 identical wind turbines commissioned at the same time were selected from each wind farm for analysis. The wind turbine to be monitored was labeled as turbine number 0, and the other turbines were labeled sequentially from 1 to 32. Table 1 shows the corresponding data and fault information for the two wind farms. It can be seen that there is a significant difference in the amount of effective data provided by each turbine in the two wind farms. During the study, the data was rigorously screened to ensure synchronization of all data. Specifically, if the data of one wind turbine in the same wind farm was deleted, the data of all other wind turbines at that moment were also deleted. The deleted data included data with wind speeds lower than the cut-in wind speed and higher than the cut-out wind speed, as well as data from wind turbine downtime and maintenance periods. Furthermore, the anomaly detection timescale in this paper was set to 1 hour.

[0145] Table 1. SCADA Data Information and Dataset Division of Wind Farms

[0146]

[0147] (1) Detection of abnormal temperature in the front bearing of the gearbox based on WF1 wind farm data:

[0148] The total training data for each wind turbine unit from November 1, 2018 to April 19, 2019 is 246,240 records, of which 224,273 records are suitable for similarity assessment and model training after filtering. Furthermore, eight state variables are selected in the example to evaluate the spatiotemporal similarity of the operating states of different wind turbine units, including three macroscopic variables and five microscopic variables. The macroscopic variables include wind speed, main shaft speed, and output power, which reflect the energy conversion efficiency of the wind turbine unit from wind energy to mechanical energy and from mechanical energy to electrical energy. The five microscopic variables include the gearbox front bearing temperature as the target variable and four closely related variables. A recursive feature elimination algorithm is used in the example to evaluate the importance coefficients of each variable to the gearbox front bearing temperature. Figure 8 This demonstrates the importance coefficients between the main variables and the front bearing temperature of the gearbox obtained based on the recursive feature elimination method. Figure 8The gearbox oil temperature, gearbox rear bearing temperature, generator front bearing temperature, and engine compartment temperature were selected as the remaining four micro variables.

[0149] Having determined the state variables used for comparing the spatiotemporal similarity of wind turbine operating states, the proposed method can be used to quantify the spatiotemporal similarity of the operating states of each turbine in wind farm WF1 with WT0. The quantification results are shown in Figure 9(a). In the example, the five turbines WT6, WT11, WT23, WT17, and WT32, which rank among the top five in terms of spatiotemporal similarity with WT0, were selected as similar turbines. Their similarity quantification values ​​γ are shown in Table 2.

[0150] Table 2 Performance test of LSTM models trained based on data from different units in WF1

[0151]

[0152] Figure 9(b) shows the similarity quantification radar chart between WT0 and these five units. It can be seen that the similarity between each variable of these selected units and the corresponding variable of WT0 exceeds 0.8.

[0153] Once generator units with similar operating conditions are identified, different LSTM models can be trained using their SCADA data. The input variables for training the models include preprocessed gearbox oil temperature, gearbox rear bearing temperature, generator front bearing temperature, wind speed, and nacelle temperature for each unit; the output variable is the estimated gearbox front bearing temperature for that unit. After training, the accuracy and adaptability of each LSTM model to the WT0 data need to be evaluated using both the unit's own test dataset and the WT0 validation dataset. The results of the two performance evaluations for each model are shown in Table 2. When using their own validation data for accuracy testing, all LSTM models trained on different generator unit data showed good accuracy. However, when using the WT0 validation data for adaptability testing, the overall score β of all models decreased significantly. This indicates that different models have different adaptability to the WT0 SCADA data. Furthermore, Table 2 also shows that the LSTM models corresponding to units with strong similarity to WT0 also exhibit relatively better adaptability.

[0154] Furthermore, based on the results shown in Table 2, the LSTM model that performed well in both performance tests was selected for the next step of constructing the Combined State Estimation Model (CSEM). These sub-models were trained using SCADA on WT0, WT6, WT11, and WT23, respectively. To verify whether utilizing units with similar operating states is beneficial to improving the reliability of anomaly detection results, three different CSEM construction schemes were designed as shown in Table 3, and the performance of these three different CSEMs was tested using WT0 validation data.

[0155] Table 3 Different Combination Strategies of CSEM

[0156]

[0157] Figure 10 The results show the comparison between the gearbox front bearing temperature estimated by different schemes using CSEM in WF1 and the actual temperature. Table 4 gives the performance indicators of different CSEMs and the corresponding maximum effective value of residuals and information entropy.

[0158] Table 4. CSEM performance test results based on different schemes in WF1

[0159] <![CDATA[E mae ]]> <![CDATA[E rmse ]]> <![CDATA[R 2 ]]> β <![CDATA[H rm ]]> <![CDATA[H en ]]> Sch1 0.319 0.377 0.967 1.389 0.865 0.328 Sch2 0.331 0.408 0.958 1.296 0.937 0.351 Sch3 0.291 0.323 0.978 1.593 0.543 0.296

[0160] from Figure 10 It can be seen that the temperature curve estimated by CSEM based on Sch3 is closest to the actual temperature curve, and its performance evaluation index is also better than the other two schemes. This means that using SCADA data from multiple similar wind turbines to build a combined estimation model is beneficial to improving the accuracy of CSEM estimation results. Furthermore, from... Figure 10 It can also be found that the performance metrics of CSEM in Sch1 and Sch2 are very similar, which indicates that an effective CSEM can be constructed to estimate the state variables of WT0 using data from multiple similar units.

[0161] Furthermore, based on these three methods, the abnormal condition of the front bearing of the gearbox was detected, and the detection results are as follows: Figures 11(a) to 11(c) As shown. Figures 11(a) to 11(c) The three sub-graphs in the image represent the detection results of three different schemes. On the left side of each sub-graph, red and blue curves show the true and estimated values ​​of the gearbox front bearing temperature, respectively. The red dashed line represents the abnormal alarm time recorded by the SCADA system. On the right side of each sub-graph, the final abnormal detection result is shown based on the comparison of the residual effective value and residual information. The blue and red curves represent the effective value Rm and information entropy En of the residual, respectively. The magenta and cyan horizontal lines represent their corresponding thresholds Hrm and Hen. The green shaded area represents the identified abnormal bearing temperature period.

[0162] from Figures 11(a) to 11(c)As can be seen from each sub-figure, the period on the right where the estimated bearing temperature differs significantly from the actual value is basically the same as the period on the left where the residual data anomaly was detected. This verifies the effectiveness of the proposed anomaly detection method. Comparing the detection results of the three schemes reveals that Sch3 detected the bearing temperature anomaly earliest at 6:00 on May 29, 2019, 14 days earlier than the system fault alarm. This indicates that comprehensively utilizing SCADA data sequence models from WT0 and other similar units can improve CSEM sensitivity, thus facilitating early anomaly detection. Furthermore, comparing the anomaly detection results of Sch1 and Sch2 shows that the earliest detection times of the gearbox front bearing temperature anomaly are almost identical for both schemes, 10 days earlier than the system fault alarm. This demonstrates that using SCADA data from units with similar operating conditions to the unit under test for anomaly detection is feasible.

[0163] (2) Gearbox oil temperature anomaly detection based on WF2 wind farm data:

[0164] As shown in Table 1, the available data for the WF2 wind farm is relatively limited. Therefore, utilizing data from turbines with similar operating conditions within the same wind farm is more practically meaningful for improving the accuracy of anomaly detection. Data from February 1, 2018 to March 23, 2018 was selected to quantitatively assess the spatiotemporal similarity of the operating conditions of each turbine in WF2 with that of WTO. The total number of assessment data for all turbines was 14,688, of which 12,727 were valid after cleaning. Wind speed, main shaft speed, and output power were selected as macroscopic variables, while gearbox oil temperature, gearbox rear bearing temperature, gearbox high-speed shaft speed, gearbox inlet oil temperature, and gearbox front bearing temperature were selected as microscopic variables.

[0165] Figure 12(a) shows the spatiotemporal similarity quantification results of the operating status of each unit in WF2 and WT0. Based on the five units WT08, WT05, WT01, WT15 and WT03, which were selected as similar units, Figure 12(b) shows the similarity radar chart of these five units and WT0.

[0166] Table 5 Performance test of LSTM models trained based on data from different units in WF2

[0167]

[0168] Table 6. CSEM performance test results based on different schemes in WF2

[0169]

[0170] Table 5 presents the performance test results of the LSTM trained on WT0 and these five wind turbine units in two tests. It can be seen that because WF2 provides less data than WF1, the overall performance of the LSTM model trained on each turbine's data is slightly worse than the models shown in Table 2. However, the conclusions reflected in Table 5 and Table 2 are basically consistent: the accuracy test results for all turbine units are good, but the adaptability test results are significantly worse. Similarly, based on the two performance test results, the relatively better-performing LSTM model was selected as a sub-model to construct the CSEM. These selected sub-models were trained on SCADA data from WT0, WT01, WT05, and WT08, respectively. Table 6 presents the performance indicators, corresponding maximum effective residual values, and information entropy of the different CSEMs constructed under three different schemes. It can be found that the CSEM constructed by fully utilizing similar wind turbine data shows improvements in all indicators.

[0171] (3) Comparative verification:

[0172] To verify the advantages of the proposed method, four different anomaly detection methods were further used as comparative models. These methods respectively utilize multiple linear regression (MLR), support vector machine (SVM), deep neural network (DNN), and multi-model combined estimation (MMC) to construct state estimation models for the target variable. All four comparative validation models were trained using only historical data from WT0, and the input and output variables remained consistent with the previous examples. It should be noted that MMC differs from the constructed CSEM. Although both are combined estimation models based on linear combinations, MMC constructs sub-models using different algorithms. All sub-models are trained using WT0 data, and the combined weights in MMC are obtained by setting the objective function to minimize the sum of squared estimated residuals of all sub-models, requiring an optimization algorithm to solve for them.

[0173] Based on the data provided by WF1 and WF2, the performance of four comparative models and the CSEM constructed in Sch3 was compared. Figures 14(a) and 14(b) show the performance evaluation results of each model obtained using the WTO validation datasets from different wind farms. To facilitate the comparison of the comprehensive score β, the β index of all models in Figures 14(a) and 14(b) are normalized values, with the normalization cardinality being the largest β value among the five models. It can be found that the combined models MMC and CSEM outperform the other models, and the performance differences between different models may be related to the total amount of training data. When the WTO training data in WF1 is sufficient, the performance evaluation index differences between different models are small, as shown in Figure 14(a); when the WTO training data in WF2 is insufficient, the performance of all models except CSEM decreases significantly, as shown in Figure 14(b). The reason why CSEM can maintain good performance when the WTO training data is insufficient is that multiple wind turbines with similar operating states were used in the construction process of CSEM to alleviate the impact of insufficient data on model performance.

[0174] Next, the anomaly detection performance of different methods is compared using the WTO detection datasets from two wind farms. For ease of comparison, all methods employ the anomaly identification method proposed in this paper. Figures 15(a) and 15(b) show the anomaly detection results of different methods. It can be observed that when performing anomaly detection on WTO in WF1, due to sufficient training data, the detection results of CSEM and MMC in Figure 15(a) are quite close, both detecting abnormal front bearing temperature of the gearbox 14 days earlier than the system fault alarm. Moreover, these two methods based on combined model estimation detect anomalies earlier than other methods based on single model estimation. However, when the training data for WTO in WF2 is limited, the anomaly detection results of different models shown in Figure 15(b) show significant differences, but CSEM still detects the abnormal oil temperature of the gearbox earlier than other methods. This comparative experiment demonstrates that the proposed method is more sensitive to anomalies and can ensure the accuracy of anomaly detection results even when the training data for wind turbines is insufficient.

[0175] (4) Validation of detection performance under abnormal training data conditions:

[0176] To verify the robustness of the proposed anomaly detection method, the reliability of anomaly detection by different methods was further compared when anomaly data existed in the training data. As shown in Figures 15(a) and 15(b), before the SCADA system alarm, anomalies were detected in the target variables of both faulty units across multiple time periods. Therefore, this comparative experiment will use a portion of the detection dataset containing anomaly data to train the estimation model. Due to the small amount of detection data in WF2, this comparative experiment only uses the detection data of WT0 in WF1. A total of 12,774 data records from May 25, 2019 to June 5, 2019 were used as the new training set, and data from June 6, 2019 to June 15, 2019 were used as the new detection data. Based on the previous anomaly detection results for WT0 in WF1, the proportion of anomaly data in the new training data is 26.4%. All four comparison methods used only historical data from WT0 between May 25, 2019 and June 5, 2019 to train their models. CSEM, however, used data from both WT0 and three other similar aircraft during the same period. Since there was no validation dataset for this comparison experiment, the anomaly detection threshold H was set accordingly. rm and H en It was determined using a new training dataset.

[0177] Figures 16(a) and 16(b) show the anomaly detection results of different methods. Figure 16(a) compares the estimated values ​​with the true values, while Figure 16(b) shows the identified anomalies. It can be seen that the presence of anomalies in the training data has a significant negative impact on anomaly detection. However, compared to other methods, CSEM still maintains good accuracy under these circumstances, detecting anomalies earliest and detecting the most anomalies before the system fault alarm. Comparing the anomaly detection results of different methods shown in Figures 16(a) and 16(b), it can be found that the anomaly detection results of CSEM constructed using normal and anomaly-containing training data are roughly the same after June 6, 2019. However, the anomaly detection results of MLR, SVM, DNN, and MMC show significant differences after June 6, 2019. MLR, SVM, and MMC identify normal temperature data as anomalies, while DNN does not even detect the abnormal temperature at the time of the system fault alarm. The above results demonstrate that the proposed method can effectively mitigate the impact of anomaly data in the target unit's training data on anomaly detection results.

[0178] This invention details the principle of using SCADA data from other units in the same wind farm to detect abnormalities in the gearbox of a target unit. It also illustrates the specific implementation steps of the detection using practical examples. The detection results show that the method provided by this invention can detect abnormal operating conditions of the gearbox earlier than SCADA system fault alarms, even when historical data of the target wind turbine is insufficient or missing. This provides a reference for operators to formulate maintenance and repair plans.

Claims

1. A method for detecting gearbox anomalies in wind turbine units considering the similarity of operating states of multiple units, characterized in that... Includes the following steps: Step 1: Evaluate a single state variable based on a piecewise linearized time series similarity assessment method; Step 2: Consider the spatiotemporal similarity quantification method of wind turbine operating state considering multiple state variables. Evaluate the similarity of the operating state of each wind turbine in the same wind farm with the turbine under test, and select data from multiple turbines with strong similarity in operating state with the turbine under test. Step 3: Preprocess the historical data of the unit to be tested and the historical data of the selected unit to be tested in Step 2, and divide them into training data, test data and detection data according to time period; Step 4: Construct a state estimation LSTM model. Using the training data of the selected similar units and the training data of the unit to be detected, train multiple different state estimation LSTM models. Step 5: Set comprehensive evaluation metrics and use test data to verify the accuracy and adaptability of the LSTM model performance for different state estimation modes; Step 6: Select several state estimation LSTM models that have performed well in the performance verification in Step 5, and construct a combined state estimation model by weighted combination. Step 7: Based on the combined state estimation model in Step 6, estimate the target variables of the unit to be tested, and calculate the residual between the estimated value and the true value. Based on the comparison of the effective value of the residual and the comparison of the information entropy of the residual, perform anomaly identification. Step 1 includes the following steps: S1.1: Suppose there are two wind turbine state variable time series curves, L1 and L2, each containing L sampling points; S1.2: A linearization segmentation method is used to divide a long-term time series into different short-term time series. The state variable time series L1 and L2 after linearization segmentation can be approximated by... h Duan He k The broken line segment represents, h >2, k >2; the breakpoints are taken as the points of abrupt change in value on the two curves; the broken lines after the division have only three trends: rising, falling, and remaining unchanged. These three trends are represented by 1, -1, and 0 respectively. Then, the time series curve of the state variable can be approximated by the dataset. S 1 and S 2 means: （1）； In equation (1), t Indicates the time corresponding to each segmentation point; m The broken line represents the trend of the segmentation, and its value set is {1, -1, 0}; The sum of the normalized values ​​at the ends and the normalized values ​​at the beginning of each polyline represents the difference between the end and beginning values. It reflects the amplitude variation of the polyline. The superscripts s1 and s2 are used to distinguish different polylines, and the subscripts are... k and h Then, the k-th and h-th polyline segments are identified respectively; Elements in a set and These represent the first and second lines of curve L1, respectively. h The first segment of the linearized piecewise linear curve and the first segment of curve L2 k The linearized polylines; the sets of times corresponding to their linearized cutoff points are respectively denoted as... and ;in: This indicates that the time series piecewise linear line L1 is at the 1st... h The time corresponding to each segment point The time series piecewise linear curve L2 represents the time series line at the 1st... k The time corresponding to each segment point; S1.3: Based on the first linearization segmentation in S1.2, a second linearization segmentation is performed to ensure that the linearization segmentation points of curves L1 and L2 are exactly the same; the set of the second segmentation points. T Through the and The union of the subsets, arranged chronologically according to the time of each partition point, is expressed as: (2)； In formula (2): sort () indicates that the values ​​are sorted in ascending order; the symbol This means finding the union of sets. l This indicates the number of division points contained in each curve after the second division. h and k These represent the number of segmentation points contained in curves L1 and L2 after the first linearization segmentation, respectively. Indicates the first l The time corresponding to each segmentation point; ; The L1 and L2 linearized segmentation sets of the original time series curve after the second linearization segmentation. and Represented as: (3)； In formula (3): These represent the time points corresponding to different segmentation points; This indicates that curve L1, after the second linearization segmentation, is... and The transformation trend of a line segment is determined by two points. The amplitude trend is ; This indicates that curve L2, after the second linearization segmentation, is... and The transformation trend of a line segment is determined by two points. The amplitude trend is ; At this point, the time series curve of the state variable is divided into multiple line segments of equal number and with completely consistent dividing points; S1.4: Similarity of state variables according to equation (4) Perform the calculation: (4)； In equation (4): The value of is in the range of [0,1], and the larger the value, the greater the similarity between the two state variables; L represents the total number of linear segments, and i is the iteration count identifier. m i s1 This represents the trend of change of the i-th line segment after L1 linearization of the curve. m i s2 △ represents the trend of the i-th line segment after L2 linearization of the curve. y i s1 Δ represents the magnitude of the change in the i-th line segment after linearizing the curve L1. y i s2 This represents the magnitude of change of the i-th line segment after L2 linearization of the curve; m i s1 - m i s2 This represents the trend distance of the corresponding line segment after linearization; △ y i s1 -△ y i s2 This represents the magnitude distance of the corresponding line segment after linearization. Step 2 includes the following steps: S2.1: Select multiple state variables that can comprehensively reflect the similarity between the macroscopic and microscopic states of the wind turbine, and calculate the similarity of each state variable according to equation (5); （5）； In equation (5): S a It is the area of ​​the shaded region. For the similarity of the i-th state variable, S 0 represents the area of ​​the regular polygon, and n is the number of selected state variables. For the similarity of the first state variable, The similarity of the (i+1)th state variable, The similarity of the (n+1)th state variable; This is a quantified value for similarity. S2.2: Based on the radar image area comparison method, the spatiotemporal similarity of the wind turbine operating status is quantified; S2.2.1: First, construct a radar chart with the same number of regular polygons as the state variables selected in S2.

1. Define each axis of the radar chart as a state variable, and set the base length of the axis to 1. S2.2.2: Next, starting from the center of the radar chart of the regular polygon, draw arrows outward along the axis. The length of the arrows represents the similarity of the state variables corresponding to the two wind turbines. S2.2.3: Finally, connect all the arrow endpoints in sequence to form a closed figure, that is, the shaded part of the radar chart. The proportion of the area of ​​the closed figure to the area of ​​the radar chart represents the spatiotemporal similarity of the operating status of the two wind turbines. In step 5, the performance evaluation criterion for the state estimation LSTM model is determined by a comprehensive evaluation index. β To measure this, the calculation formula is as follows: （7）； In equation (7): and These are the first and second target variables in the validation dataset. i One true value and one estimated value; N To verify the total number of samples in the dataset; To verify the mean of the target variable in the dataset; and These represent the root mean square and absolute mean of the residuals, respectively. The smaller their values, the more accurate the model estimate. R 2 The decision coefficient reflects the goodness of fit of the model to the data. Its value ranges from [0,1]. The larger the value, the stronger the interpretability of the model to the data. According to equation (7), the comprehensive score index is... β With the above , and R 2 The three indicators are closely related, and the larger the value, the better the overall performance of the model.

2. The wind turbine gearbox anomaly detection method considering the similarity of operating states of multiple units as described in claim 1, characterized in that: Step 4 includes the following steps: S4.1: Construct a state estimation model for the target variable using LSTM: LSTM contains a forget gate. Input gate and output gate Among them, the Gate of Oblivion This determines which hidden state information passed down from the previous unit needs to be forgotten in the new unit; input gate Used to determine what information the current cell needs to store; output gate This is used to determine which information needs to be output to the next unit from the current hidden state; the state update process is represented as: (6)； In formula (6): W f and b f These are the weights and biases of the forget gate, respectively; W i and b i These are the weights and bias terms of the input gate, respectively; W o and b o These are the weights and bias terms of the output gate, respectively; W c and b c These are the weights and biases of the current state, respectively; g (,) denotes a gate function; ⊙ denotes element-wise multiplication of matrices. h t-1 This represents the hidden layer variable at time t-1, which has short-term memory capabilities. C t-1 and C t These represent the cell layer variables at time t-1 and time t, respectively, and they possess long-term memory capabilities. X t Represents the input variable at time t; S4.2: A two-layer LSTM stacking mode is adopted to construct the state estimation model of the wind turbine. After training multiple state estimation LSTM models using SCADA data of the turbine under test and other turbines with similar operating states, the state estimation LSTM model with relatively better performance will be used to construct the combined state estimation model CSEM.

3. The wind turbine gearbox anomaly detection method considering the similarity of operating states of multiple units as described in claim 1, characterized in that: In step 6, the combined state estimation model CSEM is constructed using a weighted combination method, as shown in equation (8): (8) ； In equation (8): M i This represents the i-th LSTM sub-model; m is the number of LSTM sub-models selected. M CSEM Represents CSEM; The similarity quantification value of the wind turbine is calculated based on equation (5); the combined weight is related to the spatiotemporal similarity of the wind turbine's operating state, that is, the more similar the operating state of the wind turbine in the same wind farm is to the unit under test, the greater the contribution of the model trained based on the unit's data to CSEM. M 0 indicates that the LSTM model is trained based on the historical data of the unit under test, and the corresponding... =1, when the training data for the unit under test is insufficient or contains many outliers, it is difficult to obtain reliable results. M At 0 o'clock, The value of is set to 0, indicating that M 0 does not participate in the construction of CSEM, but relies solely on wind turbine data with similar operating conditions to identify anomalies in the unit under test.

4. The wind turbine gearbox anomaly detection method considering the similarity of operating states of multiple units as described in claim 1, characterized in that: Step 7 includes the following steps: S7.1: By using sliding window sampling, the long-term series of the target variable of the unit under test is divided into multiple short-term series for anomaly detection. Each short-term series... R m and E n The value is calculated according to equation (9): （9）； In equation (9): I This represents the number of statistical intervals divided based on the residual distribution. d i For falling in the first i Number of residual samples within each statistical interval; This represents the proportion of samples within the corresponding statistical interval to the total number of samples tested. α The unit of measurement for information entropy is determined, and it is usually taken as... α The value is the natural logarithm. e ; T inv It is the sampling interval of the SCADA system; T r It refers to the window width used in the sliding window, which reflects the time scale of the detection; S7.2: Set three statistical intervals: normal region, risk region, and high-risk region to calculate the information entropy for each detection period. The three statistical intervals are not continuous, and their boundary values ​​are based on four important parameters of the box plot. Q low , Q 1, Q 3 and Q up Certainly, among which: Q 1 and Q 3 represents the 0.25 and 0.75 quantiles of the box plot, respectively; Q up and Q low These are the upper and lower boundaries in the box plot, respectively, and their values ​​are calculated using equation (10). (10) ； When the residual value is greater than Q 1 and less than Q At 3, the residual falls within the normal range; when the residual value is less than 3, the residual falls within the normal range. Q low or greater than Q up At that time, the residuals fell into the high-risk area; The remaining values ​​all fall within the risk area, and the residuals in the high-risk area deviate significantly from the overall normal distribution range; S7.3: Anomaly Detection Process: First, calculate the effective value and information entropy of the residual samples during the detection period; second, compare the effective value of the residuals during the detection period to see if it exceeds the threshold. H rm If the residuals do not exceed the threshold, it indicates that the residual distribution is within the normal range. If the effective value exceeds the range, it indicates that the sample residuals are too large. In this case, it is necessary to further determine whether this anomaly is caused by a few extremely large noises or by a large number of abnormal residuals by using the information entropy of the residuals. If the information entropy of the residuals is also less than the given threshold... H en If the effective value is too large, it is very likely due to abnormal operation of the object being detected; otherwise, it is due to noise interference.

Images