Wind turbine generator set anomaly detection method considering state similarity of multiple generator sets

Abstract

The invention relates to a wind turbine generator set anomaly detection method considering multi-unit state similarity. Firstly, an evaluation index system is constructed, and design parameters and historical operation data are analyzed. Then data acquisition and preprocessing are carried out, an SCADA system is used for acquiring data, and denoising is carried out in combination with a local outlier factor algorithm and wavelet transform; and then qualitative evaluation is carried out, a cloud model is constructed to describe data uncertainty, a membership degree is calculated, and a cluster is dynamically adjusted. And then comprehensive evaluation is carried out, index weights are determined through an analytic hierarchy process, and scores are calculated in combination with membership degrees. And finally, performing anomaly detection, setting a threshold value, combining a dynamic clustering method, and taking a quantile of a sliding time window and a similar unit prediction residual error as a real-time early warning threshold value. According to the method, the accuracy and reliability of wind turbine generator anomaly detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine operation monitoring, and in particular to a wind turbine abnormality detection method considering the similarity of multiple turbine states. Background Art

[0002] With the continuous development of wind power generation technology, wind turbines, as an important clean energy power generation equipment, have attracted extensive attention for their operating status monitoring and anomaly detection. At present, there are many wind turbine anomaly detection methods, which can be roughly divided into model-based methods and data-driven methods.

[0003] Model-based methods usually rely on the physical model of the wind turbine. They describe the operating characteristics of the wind turbine by establishing an accurate mathematical model, and detect anomalies based on the difference between the model output and the actual measured value. For example, aerodynamic models, mechanical dynamics models, etc. are used to predict key parameters such as the power output and speed of the wind turbine, and the predicted values ​​are compared with the actual values. However, this type of method has some limitations: the model is highly complex. The wind turbine is a complex electromechanical system affected by a variety of physical factors. The establishment of an accurate physical model requires consideration of a large number of parameters and complex physical processes, which makes the model construction and solution process very complex and the computational cost high; the model accuracy is highly dependent on the model accuracy, and the accuracy of the model directly affects the results of anomaly detection. Due to the complexity and variability of the actual operating environment, such as the randomness of wind speed, the uncertainty of wind direction, and the change of air density, it is difficult to ensure that the model can accurately describe the actual operating status of the wind turbine, which leads to misjudgment of the detection results.

[0004] The data-driven method uses the historical operation data of wind turbines to discover abnormal patterns in the data through data analysis and mining techniques. Common methods include statistical analysis methods, machine learning methods, etc. For example, statistical indicators such as mean and standard deviation are used to determine whether the data is abnormal, or machine learning algorithms such as neural networks and support vector machines are used to classify and predict the operating status of wind turbines. However, this type of method also has some shortcomings: lack of consideration of the similarity of unit status. In actual wind farms, there are usually multiple wind turbines running at the same time. These units may have similarities in design parameters, operating environment and operating status. However, most of the existing data-driven methods analyze a single unit and do not make full use of the similarity information between units. As a result, some abnormal features related to unit similarity may be ignored during anomaly detection, reducing the accuracy and reliability of detection; data quality issues have a great impact, and data-driven methods have high requirements for data quality, such as data integrity, accuracy and consistency. However, the data collected by wind turbines during operation may be interfered by various factors, such as sensor failure, data transmission errors, environmental noise, etc. These data quality issues may cause deviations in data analysis results, thereby affecting the effectiveness of anomaly detection.

[0005] To this end, we propose a wind turbine anomaly detection method that considers the similarity of multiple unit states. Summary of the invention

[0006] In view of the above deficiencies in the prior art, an object of the present invention is to provide a method for detecting abnormalities of wind turbines taking into account the similarity of states of multiple turbines, so as to improve the accuracy and reliability of abnormality detection of wind turbines.

[0007] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0008] The method for detecting abnormality of wind turbines considering the similarity of multiple turbine states comprises the following steps:

[0009] S1. Build an evaluation index system, analyze the design parameters and historical operation data of wind turbines, and identify the performance indicators and environmental factors of wind turbines;

[0010] S2. Data collection and preprocessing: The SCADA system is used to automatically and in real time collect the operating data of the wind turbines, and the local outlier factor (LOF) algorithm is used to preprocess the operating data of the wind turbines, select the wavelet basis function and the number of decomposition layers, and perform denoising on different types of data;

[0011] S3. Qualitative evaluation: Use the processed data in S2 to build a cloud model to describe the uncertainty and fuzziness of real-time operation data, calculate the membership of data points at different evaluation levels, and convert quantitative data into qualitative evaluation; dynamically adjust the cluster based on the calculation results of the membership to reflect the changes in the real-time operation status of the wind turbine;

[0012] S4. Comprehensive evaluation: using the real-time operating status of the wind turbines determined in S3, constructing a judgment matrix and performing consistency tests through the analytic hierarchy process (AHP), determining the weights of each evaluation index, and calculating the comprehensive evaluation score by combining the membership degree and weight;

[0013] S5. Anomaly detection: The anomaly detection threshold is set based on the results of comprehensive evaluation. Combined with the dynamic clustering method, the quantile of the sliding window and the prediction residual of similar units is used as the real-time warning threshold to detect the abnormal state of the wind turbine in real time.

[0014] Furthermore, in S1, the design parameters include blade length, hub height, rated power, generator type, etc. The design parameters directly affect the performance and operating characteristics of the wind turbine.

[0015] Furthermore, in S1, the historical operation data includes parameters such as power output, rotation speed, temperature, vibration, etc. The historical operation data reflects the operation status and performance of the wind turbine under different environmental conditions.

[0016] Furthermore, in S1, the method for identifying the performance indicators and environmental factors of the wind turbine generator set is:

[0017] Assume that the performance index of the wind turbine is Y, and the design parameters and environmental factors are X1, X2, ..., X n , then the multiple regression model can be expressed as:

[0018] Y=β0+β1X1+β2X2+…+β n X n +ε;

[0019] Among them, β0, β1, β2, β n is the regression coefficient, ε is the error term;

[0020] Through regression analysis of historical operation data, the influence of various factors on performance indicators is determined. The size and sign of the regression coefficient reflect the importance and influence direction of the factor.

[0021] If the regression coefficient β i is positive and large, it means that factor X i It has a positive impact on the performance indicator Y, and the impact is large; conversely, if the regression coefficient is negative or small, it means that the impact of the factor is small or negative.

[0022] Furthermore, in S2, the SCADA system transmits the operating data of the wind turbine to the central control system through sensors and data acquisition equipment for real-time monitoring and analysis.

[0023] Furthermore, in S2, the local outlier factor (LOF) algorithm is an algorithm for detecting outliers in data; if there are some outliers in the operating data of the wind turbine, these outliers will affect the subsequent analysis and evaluation results; the local outlier factor (LOF) algorithm performs a preprocessing method on the operating data of the wind turbine:

[0024] First, calculate its local reachability density (LRD);

[0025] Suppose the distance between point p and point o is d(p,o), and the k-distance neighborhood of point p is N k (p), which is the set of k points closest to point p; the k-distance neighborhood of point o is N k (o), that is, the set of k points closest to point o;

[0026] The reachable distance rd of point p k (p,o) reflects the reachability from point p to point o, and its expression is:

[0027] rd k (p,o)=max{k-dt(o),d(p,o)};

[0028] The reachable distance rd of point o k (o,p) reflects the reachability from point o to point p, and its expression is:

[0029] rd k (o,p)=max{k-dt(p),d(o,p)};

[0030] Where k-dt(o) is the distance from point o to its kth nearest neighbor; the calculation of the reachable distance comprehensively considers the k distance of point o and the actual distance between point p and point o; k-dt(p) is the distance from point p to its kth nearest neighbor; the calculation of the reachable distance comprehensively considers the k distance of point o and the actual distance between point p and point o

[0031] The local reachability density lrd of point p k (p) is:

[0032]

[0033] N k (p) is the k-distance neighborhood of point p; the local reachability density lrd of point p k(p) reflects the density of data points around point p. The larger the local reachable density, the denser the data points around point p. The smaller the local reachable density, the sparser the data points around point p.

[0034] The local reachability density lrd of point o k (o) is:

[0035]

[0036] The local reachability density lrd of point o k (o) reflects the density of data points around point o. The larger the local reachable density, the denser the data points around point o. The smaller the local reachable density, the sparser the data points around point o.

[0037] Then calculate the local outlier factor of point p:

[0038]

[0039] The local outlier factor of point p measures the degree of outlier of point p relative to its surrounding data points; if LOF k (p) is much greater than 1, then point p is an outlier; the local outlier factor determines whether point p is an outlier by comparing the ratio of the local reachable density of point p to the local reachable density of the data points around it; if the local reachable density of point p is significantly lower than the local reachable density of the data points around it, then it is an outlier, which may be caused by sensor failure, data transmission error, etc.

[0040] Furthermore, in S2, the method of selecting the wavelet basis function and the number of decomposition layers includes:

[0041] S21. Wavelet basis function selection, select basis functions such as Morlet wavelet, Morlet wavelet has good localization characteristics in the time-frequency domain, and can better capture the mutation and high-frequency components in the vibration signal;

[0042] S22. Determine the number of decomposition layers. The number of decomposition layers is determined based on the sampling frequency and the highest frequency component of the signal. Suppose the sampling frequency of the signal is f s , the highest frequency component is f max , then according to the formula To estimate an initial decomposition level J;

[0043] S23. Denoising, applying wavelet transform to decompose the data, decomposing the signal into components of different frequencies;

[0044] S231. Wavelet decomposition, assuming the original signal is s(t), the wavelet coefficients d at different scales can be obtained through wavelet transform j,k and the scale factor c J,k; j represents the decomposition layer, and the value range of j is from J to 1; k represents the time index, and the wavelet transform formula is:

[0045]

[0046]

[0047] Among them, ψ j,k (t) represents the wavelet function ψ j,k (t); represents the scaling function;

[0048]

[0049]

[0050] S232.Threshold processing:

[0051] The noise component is removed by threshold processing method; assuming that the wavelet coefficient fixed threshold is λ, the processed wavelet coefficient for:

[0052]

[0053] The noise component is removed by threshold processing method; assuming that the fixed threshold of scale factor is λ′, the scale factor after processing is for:

[0054]

[0055] Among them, sign(d j,k ) means taking d j,k The symbol, sign(c J,k ) means taking c J,k Symbols;

[0056] S233. Reconstruct the signal, reconstruct the denoised signal by inverse wavelet transform; the reconstruction formula is:

[0057]

[0058] It means reconstructing the denoised signal through inverse wavelet transform; by combining the processed scaling coefficients and wavelet coefficients with the corresponding scaling function and wavelet function, the signal is reconstructed to obtain the denoised signal.

[0059] Further, the S3 includes:

[0060] S31. Build a cloud model. Use the data processed in S2 to build a cloud model. Use three digital features: expected Ex, entropy En, and super entropy He to describe the uncertainty and fuzziness of real-time operation data. Expected Ex represents the central value of the concept, entropy En reflects the degree of uncertainty of the concept, and super entropy He further describes the uncertainty of entropy.

[0061]

[0062]

[0063]

[0064] In the formula, Ex stands for expectation, which is the central tendency or average level of the data; En stands for entropy, which reflects the degree of uncertainty of the concept; He stands for hyperentropy, which further describes the uncertainty of entropy En; x i represents a data point; n represents the number of data points; En i Represents the entropy value corresponding to each data point; Entropy En i The average value of

[0065] S32. Calculate the membership and qualitative evaluation, and use the normal cloud model to calculate the membership of each data point at different evaluation levels; by calculating the membership, convert the quantitative data into qualitative evaluation, that is, determine the membership of different evaluation levels;

[0066] The degree of membership indicates the degree to which a data point belongs to a certain evaluation level; for example, the operating status of a wind turbine can be divided into several levels such as "good", "average", and "poor", and then the degree of membership of each data point in these levels can be calculated;

[0067] The membership function calculation formula of the normal cloud model is:

[0068]

[0069]

[0070] In the formula, U(x) represents the membership function value; x represents the value of the data point, and x is the operating data of the wind turbine obtained by sensors and data acquisition equipment; En′ is used to adjust the shape or width of the function, and represents the probability or uncertainty of the center of the width distribution;

[0071] S33. Dynamically adjust the division of clusters, using the membership of different evaluation levels determined in S32 to divide data points with similar membership into the same cluster, so that the cluster can better reflect the actual operating status of the wind turbine;

[0072] Further, the S33 includes:

[0073] S331. Assume that the membership threshold range is [u min ,u max ,], and classify the data points with membership within this range into a preliminary cluster C = {C1, C1, C3, …C n};

[0074] S332. Calculate the similarity between clusters. Use cosine similarity to calculate the similarity between clusters. The cosine similarity formula is:

[0075]

[0076] In the formula, C i and C j Represent two clusters respectively; and Represent cluster C i and cluster C j vector; |C i | and |C j | respectively represent cluster C i and cluster C j The number of data points in ; θ represents the angle between the two vectors; where,

[0077]

[0078]

[0079] in, and are the vector representations of data points x and y in the preliminary cluster C respectively;

[0080] S333. Merge similar clusters. If two clusters C i and C j If the cosine similarity of exceeds the preset threshold T, that is, cos(θ)>T, cluster C i and C j Merge into a new cluster:

[0081]

[0082] S334. Update cluster information, merge clusters, and update the new cluster Central location:

[0083]

[0084] in, Is a new cluster The center position of the new cluster data points; is the vector representation of the data point z of the new cluster; C ij Is a cluster The number of data points in ;

[0085] S335. Repeat the iteration, and keep repeating S332-S334 until the division of the clusters no longer changes significantly or the preset number of iterations is reached;

[0086] S336. Dynamic adjustment: As new data is added, the membership u′(x) of the data point is recalculated, and the new data point x′ is divided into the corresponding cluster according to the new membership; if u′(x′)∈[u min ,u max ] and with an existing cluster C k If the cosine similarity of x is high, add x′ to C k ; Otherwise, create new clusters as needed;

[0087] S337. Reflect the actual operating status. Through continuous dynamic adjustment, the cluster can more accurately reflect the actual operating status of the wind turbine.

[0088] Further, the S4 includes:

[0089] S41. Construct a judgment matrix and conduct consistency test through the analytic hierarchy process (AHP) to determine the weight of each evaluation indicator;

[0090] S42. Calculate the comprehensive evaluation score by combining the membership and weight. Let the membership of the data point at different evaluation levels be u ik , the weight of each evaluation index is ω i , the comprehensive evaluation score is expressed as:

[0091]

[0092] Where n is the number of elements in the judgment matrix, m is the number of evaluation levels, and s ik It is the score function of the data point corresponding to the evaluation level k under the indicator i.

[0093] Furthermore, the evaluation indicators refer to specific indicators used to measure the operating status and performance of wind turbines. The evaluation indicators are determined by analyzing the design parameters and historical operating data of the wind turbines, and they can reflect the operating status and performance of the wind turbines under different conditions.

[0094] Further, the S41 includes:

[0095] S411. Establish a hierarchical model to decompose the evaluation problem of wind turbines into a target layer, a criterion layer and an indicator layer; the target layer is to evaluate the comprehensive operating status of wind turbines, the criterion layer may include performance, reliability, safety and other aspects, and the indicator layer is the specific evaluation indicator;

[0096] For the elements at the same level, we compare them pairwise to determine their relative importance. Let the importance comparison value of index i and index l be a il , satisfying a il >0, And a ii =1;

[0097] S412. Construct a judgment matrix A = (a il ) n×n , A is the judgment matrix, n is the number of elements in the judgment matrix,

[0098] S413. Calculate the maximum eigenvalue of the judgment matrix, including:

[0099] Initialize the eigenvector, choose an arbitrary non-zero initial vector

[0100] Iterative calculation, let vector Then the vector Normalize and get the feature vector

[0101] in, is a vector The module length; is the eigenvector obtained in the previous iteration;

[0102] Repeat the iteration and iterate the calculation steps until the eigenvector convergence;

[0103] Calculate the maximum eigenvalue. When the eigenvector converges, the maximum eigenvalue calculation formula is:

[0104]

[0105] Represents the feature vector Multiply it with the judgment matrix A, and then multiply the resulting vector with The transpose of the dot product is performed, and this value represents the current eigenvector Next, judge the "weighted influence" of matrix A; Represents the calculation of eigenvector The dot product with itself, which is actually the eigenvector The square of the length, indicating the eigenvector of “self-influence”;

[0106] S414. Normalize the feature vector to obtain the weight of each indicator;

[0107] S415. Perform consistency check, calculate consistency index CI, find the corresponding average random consistency index RI, calculate consistency ratio CR, if CR < 0.1, it is considered that the judgment matrix has satisfactory consistency, otherwise the judgment matrix needs to be readjusted;

[0108] in,

[0109] In the formula, λ max is the maximum eigenvalue; n is the number of elements in the judgment matrix.

[0110] Further, the S5 includes:

[0111] S51. Set the threshold value. Based on the results of the comprehensive evaluation in S4, analyze the historical data, calculate the mean μ and standard σ of the relevant indicators, and set the threshold value to F = μ ± kσ, where F is the set threshold value and k is a constant determined based on experience or actual needs;

[0112] S52. Real-time warning threshold,

[0113] S521. Combined with the dynamic clustering method, the sliding window and the quantile of the prediction residual of similar units are used as the real-time warning threshold. The size of the sliding window is set according to the actual situation. For example, a period of time in the past can be selected as the sliding window;

[0114] S522. For the data point of the wind turbine generator set, the prediction residual e between it and the data points of similar turbine generator sets is calculated within the sliding time window:

[0115]

[0116] Among them, y is the actual value, which is Predicted value;

[0117] The selection of similar units can be determined based on the unit’s performance indicators, operating parameters, geographical location, and other factors;

[0118] S523. Use statistical methods to calculate the quantiles of the prediction residuals, such as the median, the upper quartile Q3, and the lower quartile Q1;

[0119] S524. The obtained quantile is used as the real-time warning threshold, and when e>Q3 or e<Q1, a warning is triggered.

[0120] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are:

[0121] By comprehensively considering multiple factors and building an evaluation index system, we not only consider the design parameters and historical operating data of wind turbines, but also analyze the relationship between these factors, so as to have a more comprehensive understanding of the operating characteristics of wind turbines. This method of comprehensively considering multiple factors helps to improve the accuracy of anomaly detection and avoid misjudgment caused by the limitations of a single factor.

[0122] Taking full advantage of the similarity of units, the present invention fully considers the similarity of the states of multiple units in the anomaly detection process. By analyzing and processing the data of multiple units, the similarity features between units are mined, and these features are used to assist in anomaly detection. For example, in the data preprocessing stage, the data of similar units can be used to make a more accurate judgment on abnormal values; in the qualitative evaluation and comprehensive evaluation stages, the evaluation indicators and weights can be adjusted according to the similarity of the units, thereby improving the accuracy and reliability of the evaluation results.

[0123] Effectively deal with data quality issues. To address data quality issues, the present invention uses a series of data preprocessing methods. For example, using the local outlier factor (LOF) algorithm to preprocess the collected data can effectively identify and remove outliers in the data and improve the quality of the data. At the same time, by selecting appropriate wavelet basis functions and decomposition layers to denoise the data, the accuracy and reliability of the data are further improved, thereby providing a good data foundation for subsequent anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0124] Figure 1 A method flow chart of a wind turbine anomaly detection method considering similarity of multiple turbine states;

[0125] Figure 2 A method flow chart for selecting wavelet basis functions and the number of decomposition levels;

[0126] Figure 3 is a method flow chart of S3;

[0127] Figure 4 is a flow chart of the method of S4;

[0128] Figure 5 This is a flow chart of the method of S5. DETAILED DESCRIPTION

[0129] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in combination with the embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and are not used to limit the invention.

[0130] S1. Build an evaluation index system, analyze the design parameters and historical operation data of wind turbines, and identify the performance indicators and environmental factors of wind turbines;

[0131] S2. Data collection and preprocessing: The SCADA system is used to automatically and in real time collect the operating data of the wind turbines, and the local outlier factor (LOF) algorithm is used to preprocess the operating data of the wind turbines, select the wavelet basis function and the number of decomposition layers, and perform denoising on different types of data (such as vibration signals and temperature changes);

[0132] S3. Qualitative evaluation: Use the processed data in S2 to build a cloud model to describe the uncertainty and fuzziness of real-time operation data, calculate the membership of data points at different evaluation levels, and convert quantitative data into qualitative evaluation; dynamically adjust the cluster based on the calculation results of the membership to reflect the changes in the real-time operation status of the wind turbine;

[0133] S4. Comprehensive evaluation: using the real-time operating status of the wind turbines determined in S3, constructing a judgment matrix and performing consistency tests through the analytic hierarchy process (AHP), determining the weights of each evaluation index, and calculating the comprehensive evaluation score by combining the membership degree and weight;

[0134] S5. Anomaly detection: The anomaly detection threshold is set based on the results of comprehensive evaluation. Combined with the dynamic clustering method, the quantile of the sliding window and the prediction residual of similar units is used as the real-time warning threshold to detect the abnormal state of the wind turbine in real time.

[0135] In S1, the design parameters include blade length, hub height, rated power, generator type, etc., which directly affect the performance and operating characteristics of the wind turbine; the historical operating data includes parameters such as power output, speed, temperature, vibration, etc., which reflect the operating status and performance of the wind turbine under different environmental conditions.

[0136] In S1, the key performance indicators and environmental factors identification method:

[0137] The performance index of the wind turbine is Y, and the design parameters and environmental factors are X1, X2, ..., X n , then the multiple regression model can be expressed as:

[0138] Y=β0+β1X1+β2X2+…+β n X n +ε;

[0139] Among them, β0, β1, β2, β n is the regression coefficient, ε is the error term;

[0140] By performing regression analysis on historical operating data, the influence of various factors on performance indicators is determined. The size and sign of the regression coefficient reflect the importance and influence direction of the factor. For example, if the regression coefficient β iis positive and large, it means that factor X i It has a positive impact on the performance indicator Y, and the impact is large; conversely, if the regression coefficient is negative or small, it means that the impact of the factor is small or negative.

[0141] In S2, the SCADA system transmits the operating data of the wind turbine to the central control system through sensors and data acquisition equipment for real-time monitoring and analysis;

[0142] For example, the SCADA system can be installed in the nacelle, hub, tower and other parts of a wind turbine, collect temperature, vibration, wind speed and other parameters through sensors, and transmit the data to the central control system for processing and analysis.

[0143] In S2, the local outlier factor (LOF) algorithm is an algorithm for detecting outliers in data; if there are some outliers in the operating data of the wind turbine, these outliers will affect the subsequent analysis and evaluation results; the local outlier factor (LOF) algorithm preprocessing method is:

[0144] For a data point, first calculate its local reachable density (LRD);

[0145] Suppose the distance between point p and point o is d(p,o), and the k-distance neighborhood of point p is N k (p), which is the set of k points closest to point p; the k-distance neighborhood of point o is N k (o), that is, the set of k points closest to point o;

[0146] The reachable distance rd of point p k (p,o) reflects the reachability from point p to point o, and its expression is:

[0147] rd k (p,o)=max{k-dt(o),d(p,o)};

[0148] The reachable distance rd of point o k (o,p) reflects the reachability from point o to point p, and its expression is:

[0149] rd k (o,p)=max{k-dt(p),d(o,p)};

[0150] Where k-dt(o) is the distance from point o to its kth nearest neighbor; the calculation of the reachable distance comprehensively considers the k distance of point o and the actual distance between point p and point o; k-dt(p) is the distance from point p to its kth nearest neighbor; the calculation of the reachable distance comprehensively considers the k distance of point o and the actual distance between point p and point o

[0151] The local reachability density lrd of point p k (p) is:

[0152]

[0153] The local reachability density lrd of point p k (p) reflects the density of data points around point p. The larger the local reachable density, the denser the data points around point p. The smaller the local reachable density, the sparser the data points around point p.

[0154] The local reachability density lrd of point o k (o) is:

[0155]

[0156] The local reachability density lrd of point o k (o) reflects the density of data points around point o. The larger the local reachable density, the denser the data points around point o. The smaller the local reachable density, the sparser the data points around point o.

[0157] Then calculate the local outlier factor of point p:

[0158]

[0159] The local outlier factor of point p measures the degree of outlier of point p relative to its surrounding data points; if LOF k (p) is much greater than 1, then point p may be an outlier; the local outlier factor determines whether point p is an outlier by comparing the ratio of the local reachable density of point p to the local reachable density of the data points around it; if the local reachable density of point p is significantly lower than the local reachable density of the data points around it, then it is an outlier, which may be caused by sensor failure, data transmission error, etc.

[0160] In S2, the method of selecting the wavelet basis function and the number of decomposition layers includes:

[0161] S21. Wavelet basis function selection, select basis functions such as Morlet wavelet, Morlet wavelet has good localization characteristics in the time-frequency domain, and can better capture the mutation and high-frequency components in the vibration signal;

[0162] S22. Determine the number of decomposition layers. The number of decomposition layers is determined based on the sampling frequency and the highest frequency component of the signal. Suppose the sampling frequency of the signal is f s , the highest frequency component is f max , then according to the formula To estimate an initial decomposition level J;

[0163] S23. Denoising, applying wavelet transform to decompose the data, decomposing the signal into components of different frequencies;

[0164] S231. Wavelet decomposition, the original signal is s(t), and the wavelet coefficients d at different scales can be obtained through wavelet transform j,k and the scale factor c J,k ; j represents the decomposition number, and the value range of j is from J (the highest decomposition level) to 1; k represents the time index, and the wavelet transform formula is:

[0165]

[0166]

[0167] Among them, ψ j,k (t) represents the wavelet function ψ j,k (t); represents the scaling function;

[0168]

[0169]

[0170] S232.Threshold processing:

[0171] The noise component is removed by threshold processing method; assuming that the wavelet coefficient fixed threshold is λ, the processed wavelet coefficient for:

[0172]

[0173] The noise component is removed by threshold processing method; the scale coefficient is fixed to the threshold value λ′, then the scale coefficient after processing for:

[0174]

[0175] Among them, sign(d j,k ) means taking d j,k The symbol, sign(c J,k ) means taking c J,k Symbols;

[0176] S233. Reconstruct the signal, reconstruct the denoised signal by inverse wavelet transform; the reconstruction formula is:

[0177]

[0178] It means reconstructing the denoised signal through inverse wavelet transform; by combining the processed scaling coefficients and wavelet coefficients with the corresponding scaling function and wavelet function, the signal is reconstructed to obtain the denoised signal.

[0179] The S3 includes:

[0180] S31. Build a cloud model using the data processed in S2. Use three digital features, namely, expectation (Ex), entropy (En), and hyperentropy (He), to describe the uncertainty and fuzziness of real-time operation data. Expectation (Ex) represents the central value of a concept, entropy (En) reflects the degree of uncertainty of a concept, and hyperentropy (He) further describes the uncertainty of entropy.

[0181]

[0182]

[0183]

[0184] In the formula, Ex stands for expectation, which is the central tendency or average level of the data; En stands for entropy, which reflects the degree of uncertainty of the concept; He stands for hyperentropy, which further describes the uncertainty of entropy En; x i represents a data point; n represents the number of data points; En i Indicates the entropy value corresponding to each data point; En represents entropy En i The average value of

[0185] S32. Calculate the membership and qualitative evaluation, and use the normal cloud model to calculate the membership of each data point at different evaluation levels; by calculating the membership, convert the quantitative data into qualitative evaluation, that is, determine the membership of different evaluation levels;

[0186] The degree of membership indicates the degree to which a data point belongs to a certain evaluation level; for example, the operating status of a wind turbine can be divided into several levels such as "good", "average", and "poor", and then the degree of membership of each data point in these levels can be calculated;

[0187] The membership function calculation formula of the normal cloud model is:

[0188]

[0189]

[0190] In the formula, U(x) represents the membership function value; x represents the value of the data point, which is the operating data of the wind turbine obtained by sensors and data acquisition equipment (such as power output, speed, temperature, vibration and other parameters); En′ is used to adjust the shape or width of the function, indicating the probability or uncertainty of the center of the width distribution;

[0191] S33. Dynamically adjust the division of clusters, using the membership of different evaluation levels determined in S32 to divide data points with similar membership into the same cluster, so that the cluster can better reflect the actual operating status of the wind turbine;

[0192] The S33 includes:

[0193] S331. Assume that the membership threshold range is [u min ,u max ,], and classify the data points with membership within this range into a preliminary cluster C = {C1, C1, C3, …C n};

[0194] S332. Calculate the similarity between clusters. Use cosine similarity to calculate the similarity between clusters. The cosine similarity formula is:

[0195]

[0196] In the formula, C i and C j Represent two clusters respectively; and Represent cluster C i and cluster C j vector; |C i | and |C j | respectively represent cluster C i and cluster C j The number of data points in ; θ represents the angle between the two vectors; where,

[0197]

[0198]

[0199] in, and are the vector representations of data points x and y in the preliminary cluster C respectively;

[0200] S333. Merge similar clusters. If two clusters C i and C j If the cosine similarity of exceeds the preset threshold T, that is, cos(θ)>T, cluster C i and C j Merge into a new cluster:

[0201]

[0202] S334. Update cluster information, merge clusters, and update the new cluster Central location:

[0203]

[0204] in, Is a new cluster The center position of the new cluster data points; is the vector representation of the data point z of the new cluster; |C ij | is a cluster The number of data points in ;

[0205] S335. Repeat the iteration, and keep repeating S332-S334 until the division of the clusters no longer changes significantly or the preset number of iterations is reached;

[0206] S336. Dynamic adjustment: As new data is added, the membership u′(x) of the data point is recalculated, and the new data point x′ is divided into the corresponding cluster according to the new membership; if u′(x′)∈[u min ,u max ] and with an existing cluster C k If the cosine similarity of x is high (exceeds a certain threshold), then x′ is added to C k ; Otherwise, create new clusters as needed;

[0207] S337. Reflect the actual operating status. Through continuous dynamic adjustment, the cluster can more accurately reflect the actual operating status of the wind turbine. For example, if the operating status of the wind turbine changes, resulting in a change in the membership of the data point, the cluster is dynamically adjusted to capture this change in time and adjust the division of the cluster accordingly to ensure that the cluster can accurately reflect the current operating status of the wind turbine.

[0208] The S4 includes:

[0209] S41. Construct a judgment matrix and conduct consistency test through the analytic hierarchy process (AHP) to determine the weight of each evaluation indicator;

[0210] S42. Combine the membership and weight to calculate the comprehensive evaluation score. Suppose the membership of data points at different evaluation levels is (determined by S3), the weight of each evaluation indicator is ω i (determined by S41), the comprehensive evaluation score is expressed as:

[0211]

[0212] Among them, S is the comprehensive evaluation score, n is the number of elements in the judgment matrix, m is the number of evaluation levels, and s ik It is the score function of the data point corresponding to the evaluation level k under the indicator i.

[0213] The evaluation index refers to various specific indicators used to measure the operating status and performance of wind turbines. The evaluation index is determined by analyzing the design parameters and historical operating data of the wind turbines, and can reflect the operating status and performance of the wind turbines under different conditions;

[0214] The S41 includes:

[0215] S411. Establish a hierarchical model to decompose the evaluation problem of wind turbines into a target layer, a criterion layer and an indicator layer; the target layer is to evaluate the comprehensive operating status of the wind turbines, the criterion layer may include performance, reliability, safety and other aspects, and the indicator layer is specific evaluation indicators, such as power output, speed, temperature, vibration, etc.;

[0216] For the elements at the same level, we compare them pairwise to determine their relative importance. Let the importance comparison value of index i and index l be a il , satisfying a il >0, and

[0217] S412. Construct a judgment matrix A = (a il ) n×n , A is the judgment matrix, n is the number of elements in the judgment matrix,

[0218] S413. Calculate the maximum eigenvalue of the judgment matrix, including:

[0219] Initialize the eigenvector, choose an arbitrary non-zero initial vector

[0220] Iterative calculation, let vector Then the vector Normalize and get the feature vector

[0221] in, is a vector The module length;

[0222] Repeat the iteration and iterate the calculation steps until the eigenvector convergence;

[0223] Calculate the maximum eigenvalue. When the eigenvector converges, the maximum eigenvalue calculation formula is:

[0224]

[0225] Represents the feature vector Multiply it with the judgment matrix A, and then multiply the resulting vector with The transpose of the dot product is performed, and this value represents the current eigenvector Next, judge the "weighted influence" of matrix A; Represents the calculation of eigenvector The dot product with itself, which is actually the eigenvector The square of the length, indicating the eigenvector of “self-influence”;

[0226] S414. Normalize the feature vector to obtain the weight of each indicator;

[0227] S415. Perform consistency check, calculate consistency index CI, find the corresponding average random consistency index RI, calculate consistency ratio CR, if CR < 0.1, it is considered that the judgment matrix has satisfactory consistency, otherwise the judgment matrix needs to be readjusted;

[0228] in,

[0229] In the formula, λ max is the maximum eigenvalue; n is the number of elements in the judgment matrix.

[0230] The S5 includes:

[0231] S51. Set the threshold value. Based on the results of the comprehensive evaluation in S4, analyze the historical data, calculate the mean μ and standard σ of the relevant indicators, and set the threshold value to F = μ ± kσ, where F is the set threshold value and k is a constant determined based on experience or actual needs;

[0232] S52. Real-time warning threshold,

[0233] S521. Combined with the dynamic clustering method, the sliding window and the quantile of the prediction residual of similar units are used as the real-time warning threshold. The size of the sliding window is set according to the actual situation. For example, a period of time in the past can be selected as the sliding window;

[0234] S522. For the data point of the wind turbine generator set, the prediction residual e between it and the data points of similar turbine generator sets is calculated within the sliding time window:

[0235]

[0236] Among them, y is the actual value, which is Predicted value;

[0237] The selection of similar units can be determined based on the unit’s performance indicators, operating parameters, geographical location, and other factors;

[0238] S523. Use statistical methods to calculate the quantiles of the prediction residuals, such as the median, the upper quartile Q3, and the lower quartile Q1;

[0239] S524. The obtained quantile is used as a real-time warning threshold, for example, when e>Q3 or e<Q1, a warning is triggered.

[0240] The above description is a preferred embodiment of the invention and is not intended to limit the invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the invention should be included in the protection scope of the invention.

Claims

1. A wind turbine anomaly detection method considering the similarity of multiple turbine states, characterized in that: The following steps are involved: S1. Build an evaluation index system, analyze the design parameters and historical operation data of wind turbines, and identify the performance indicators and environmental factors of wind turbines; S2. Data collection and preprocessing: The SCADA system is used to automatically and in real time collect the operating data of the wind turbines, and the local outlier factor (LOF) algorithm is used to preprocess the operating data of the wind turbines, select the wavelet basis function and the number of decomposition layers, and perform denoising on different types of data; S3. Qualitative evaluation: Use the processed data in S2 to build a cloud model to describe the uncertainty and fuzziness of real-time operation data, calculate the membership of data points at different evaluation levels, and convert quantitative data into qualitative evaluation; dynamically adjust the cluster based on the calculation results of the membership to reflect the changes in the real-time operation status of the wind turbine; S4. Comprehensive evaluation: using the real-time operating status of the wind turbines determined in S3, constructing a judgment matrix and performing consistency tests through the analytic hierarchy process (AHP), determining the weights of each evaluation index, and calculating the comprehensive evaluation score by combining the membership degree and weight; S5. Anomaly detection: The anomaly detection threshold is set based on the results of comprehensive evaluation. Combined with the dynamic clustering method, the quantile of the sliding window and the prediction residual of similar units is used as the real-time warning threshold to detect the abnormal state of the wind turbine in real time.

2. The wind turbine anomaly detection method considering the similarity of multiple wind turbine states according to claim 1 is characterized in that: In S1, the method for identifying the performance indicators and environmental factors of the wind turbine generator set is as follows: The performance index of the wind turbine is Y, and the design parameters and environmental factors are X1, X2, ..., Xn, then the multivariate regression model can be expressed as: Y=β0+β1X1+β2X2+Λ+β n X n +e; Among them, β0, β1, β2, β n is the regression coefficient, ε is the error term; Through regression analysis of historical operation data, the influence of various factors on performance indicators is determined. The size and sign of the regression coefficient reflect the importance and influence direction of the factor. If the regression coefficient β i is positive and large, indicating that factor X i It has a positive impact on the performance indicator Y, and the impact is large; conversely, if the regression coefficient is negative or small, it means that the impact of the factor is small or negative.

3. The wind turbine anomaly detection method considering the similarity of multiple wind turbine states according to claim 1 is characterized in that: In S2, the SCADA system transmits the operating data of the wind turbine to the central control system through sensors and data acquisition equipment for real-time monitoring and analysis.

4. The wind turbine anomaly detection method considering the similarity of multiple wind turbine states according to claim 1, characterized in that: In S2, the local outlier factor (LOF) algorithm performs preprocessing on the operation data of the wind turbine generator set: First, the local reachable density (LRD) of the data points is calculated; The distance between point p and point o is d(p, o); the k-distance neighborhood of point p is Nk(p), which is the set of k points closest to point p; the k-distance neighborhood of point o is Nk(o), which is the set of k points closest to point o; The reachable distance rd of point p k (p,o) reflects the reachability from point p to point o, and its expression is: rd k (p,o)=max{k-dt(o),d(p,o)}; The reachable distance rd of point o k (o,p) reflects the reachability from point o to point p, and its expression is: rd k (o,p)=max{k-dt(p),d(o,p)}; Where k-dt(o) is the distance from point o to its kth nearest neighbor; the calculation of the reachable distance comprehensively considers the k distance of point o and the actual distance between point p and point o; k-dt(p) is the distance from point p to its kth nearest neighbor; the calculation of the reachable distance comprehensively considers the k distance of point o and the actual distance between point p and point o The local reachability density lrd of point p k (p) is: Local reachability density lrd of point p k (p) reflects the density of data points around point p. The larger the local reachable density, the denser the data points around point p. The smaller the local reachable density, the sparser the data points around point p. The local reachability density lrd of point o k (o) is: The local reachability density lrd of point o k (o) reflects the density of data points around point o. The larger the local reachable density, the denser the data points around point o. The smaller the local reachable density, the sparser the data points around point o. Then calculate the local outlier factor of point p: The local outlier factor of point p measures the degree of outlier of point p relative to its surrounding data points; if LOF k (p) is much greater than 1, then point p is an outlier; the local outlier factor determines whether point p is an outlier by comparing the ratio of the local reachable density of point p to the local reachable density of its surrounding data points; if the local reachable density of point p is significantly lower than the local reachable density of its surrounding data points, then it is an outlier.

5. The wind turbine anomaly detection method considering the similarity of multiple wind turbine states according to claim 1, characterized in that: In S2, the method of selecting the wavelet basis function and the number of decomposition layers includes: S21. Wavelet basis function selection: select Morlet wavelet basis function. Morlet wavelet has good localization characteristics in the time-frequency domain and can better capture the mutation and high-frequency components in the vibration signal. S22. Determine the number of decomposition layers. The number of decomposition layers is determined based on the sampling frequency and the highest frequency component of the signal. Suppose the sampling frequency of the signal is f s , the highest frequency component is f max , then according to the formula To estimate an initial decomposition level J; S23. Denoising, applying wavelet transform to decompose the data, decomposing the signal into components of different frequencies; S231. Wavelet decomposition, assuming the original signal is s(t), and the wavelet coefficients d at different scales are obtained through wavelet transform j,k and the scale factor c J,k ; j represents the decomposition layer, and the value range of j is from J to 1; k represents the time index, and the wavelet transform formula is: Among them, ψ j,k (t) represents the wavelet function ψ j,k (t); represents the scaling function; S232.Threshold processing: The noise component is removed by threshold processing method; the wavelet coefficient fixed threshold is λ, then the processed wavelet coefficient for: The noise component is removed by threshold processing method; the scale coefficient is fixed to the threshold value λ′, then the scale coefficient after processing for: Among them, sign(d j,k ) means taking d j,k The symbol, sign(c J,k ) means taking c J,k Symbols; S233. Reconstruct the signal, reconstruct the denoised signal through inverse wavelet transform, the reconstruction formula is: It means reconstructing the denoised signal through inverse wavelet transform; by combining the processed scaling coefficients and wavelet coefficients with the corresponding scaling function and wavelet function, the signal is reconstructed to obtain the denoised signal.

6. The wind turbine anomaly detection method considering the similarity of multiple turbine states according to claim 1, characterized in that: The S3 includes: S31. Build a cloud model using the data processed in S2. Use three digital features, namely, expected value Ex, entropy En, and super entropy He, to describe the uncertainty and fuzziness of real-time operation data. Expected value Ex represents the central value of a concept, entropy En reflects the degree of uncertainty of a concept, and super entropy He further describes the uncertainty of entropy En. In the formula, Ex stands for expectation, which is the central tendency or average level of the data; En stands for entropy, which reflects the degree of uncertainty of the concept; He stands for hyperentropy, which further describes the uncertainty of entropy En; x i represents a data point; n represents the number of data points; En i Represents the entropy value corresponding to each data point; represents the average value of entropy; S32. Calculate the membership and qualitative evaluation, and use the normal cloud model to calculate the membership of each data point at different evaluation levels; by calculating the membership, convert the quantitative data into qualitative evaluation, that is, determine the membership of different evaluation levels; The membership function calculation formula of the normal cloud model is: In the formula, U(x) represents the membership function value; x represents the value of the data point, and x is the operating data of the wind turbine obtained by sensors and data acquisition equipment; En′ is used to adjust the shape or width of the function, and represents the probability or uncertainty of the center of the width distribution; S33. Dynamically adjust the division of clusters, and use the membership of different evaluation levels determined by S32 to divide data points with similar membership into the same cluster, so that the cluster can better reflect the actual operating status of the wind turbine.

7. The wind turbine anomaly detection method considering the similarity of multiple turbine states according to claim 6 is characterized in that: The S33 includes: S331. Assume that the membership threshold range is [u min ,u max ,], the data points with membership within this range are classified into a preliminary cluster C = {C1, C1, C3, ΛC n }; S332. Calculate the similarity between clusters. Use cosine similarity to calculate the similarity between clusters. The cosine similarity formula is: In the formula, C i and C j Represent two clusters respectively; and Represent cluster C i and cluster C j vector; |C i | and |C j | respectively represent cluster C i and cluster C j The number of data points in ; θ represents the angle between the two vectors; where, in, and are the vector representations of data points x and y in the preliminary cluster C respectively; S333. Merge similar clusters. If two clusters C i and C j If the cosine similarity of exceeds the preset threshold T, that is, cos(θ)>T, cluster C i and C j Merge into a new cluster: S334. Update cluster information, merge clusters, and update the new cluster Central location: in, Is a new cluster The center position of the new cluster data points; is the vector representation of the data point z of the new cluster; |C ij | is a cluster The number of data points in ; S335. Repeat the iteration, and keep repeating S332-S334 until the division of the clusters no longer changes significantly or the preset number of iterations is reached; S336. Dynamic adjustment: As new data is added, the membership u′(x) of the data point is recalculated, and the new data point x′ is divided into the corresponding cluster according to the new membership; if u′(x′)∈[u min ,u max ] and with an existing cluster C k The cosine similarity of is high, then add x′ to C k ; Otherwise, create new clusters as needed; S337. Reflect the actual operating status. Through continuous dynamic adjustment, the cluster can more accurately reflect the actual operating status of the wind turbine.

8. The method for detecting abnormality of wind turbines considering the similarity of multiple turbine states according to claim 1, characterized in that: The S4 includes: S41. Construct a judgment matrix and conduct consistency test through the analytic hierarchy process (AHP) to determine the weight of each evaluation indicator; S42. Combine the membership and weight to calculate the comprehensive evaluation score. Suppose the membership of data points at different evaluation levels is The weight of each evaluation index is ω i , the comprehensive evaluation score is expressed as: Where n is the number of elements in the judgment matrix, m is the number of evaluation levels, and s ik It is the score function of the data point corresponding to the evaluation level k under the indicator i.

9. The wind turbine anomaly detection method considering the similarity of multiple turbine states according to claim 1, characterized in that: The S41 includes: S411. Establish a hierarchical model to decompose the evaluation problem of wind turbines into a target layer, a criterion layer and an indicator layer; the target layer is to evaluate the comprehensive operating status of wind turbines, the criterion layer includes performance, reliability and safety, and the indicator layer is the specific evaluation indicators; For the elements at the same level, two-by-two comparisons are made to determine the relative importance; the importance comparison value of index i and index l is satisfy and S412. Construct a judgment matrix A = (a il ) n×n , A is the judgment matrix, n is the number of elements in the judgment matrix, S413. Calculate the maximum eigenvalue of the judgment matrix, including: Initialize the eigenvector, choose an arbitrary non-zero initial vector Iterative calculation, let vector Then the vector Normalize and get the feature vector in, is a vector The module length; is the eigenvector obtained in the previous iteration; Repeat the iteration and iterate the calculation steps until the eigenvector convergence; Calculate the maximum eigenvalue. When the eigenvector converges, the maximum eigenvalue calculation formula is: Represents the feature vector Multiply it with the judgment matrix A, and then multiply the resulting vector with The transpose of the dot product is performed, and this value represents the current eigenvector Next, judge the "weighted influence" of matrix A; Represents the calculation of eigenvector The dot product with itself is the eigenvector The square of the length, indicating the eigenvector "Self-influence"; S414. Normalize the feature vector to obtain the weight of each indicator; S415. Perform consistency check, calculate consistency index CI, find the corresponding average random consistency index RI, and calculate consistency ratio CR; if CR < 0.1, it is considered that the judgment matrix has satisfactory consistency, otherwise the judgment matrix needs to be readjusted; in, In the formula, λ max is the maximum eigenvalue; n is the number of elements in the judgment matrix.

10. The wind turbine anomaly detection method considering the similarity of multiple wind turbine states according to claim 1, characterized in that: The S5 includes: S51. Set the threshold value. Based on the results of the comprehensive evaluation in S4, analyze the historical data, calculate the mean μ and standard σ of the relevant indicators, and set the threshold value to F = μ ± kσ, where F is the set threshold value and k is a constant determined based on experience or actual needs; S52. Real-time detection of warning thresholds, S521. Combined with the dynamic clustering method, the sliding window and the quantile of the prediction residual of similar units are used as the real-time warning threshold, and the size of the sliding window is set according to the actual situation; S522. For the data point of the wind turbine generator set, the prediction residual e between it and the data points of similar turbine generator sets is calculated within the sliding time window: Among them, y is the actual value, which is Predicted value; The selection of similar units is determined based on the unit’s performance indicators, operating parameters and geographical location; S523. Calculate quantiles of the prediction residuals using statistical methods, such as the median, the upper quartile Q3, and the lower quartile Q1; S524. The obtained quantile is used as the real-time warning threshold, and when e>Q3 or e<Q1, a warning is triggered.

Images