Wind turbine generator performance simulation analysis method and device based on time sequence analysis

Abstract

The invention relates to the technical field of data processing, and discloses a wind turbine generator performance simulation analysis method and device based on time sequence analysis. The method comprises the following steps: performing high-frequency sampling and self-adaptive filtering on pre-acquired operation data of a wind turbine generator to obtain a denoised multi-dimensional time sequence data set; performing wavelet packet decomposition and Hilbert-Huang transform on the multi-dimensional time sequence data set to obtain a time-frequency domain feature matrix; performing nonlinear manifold learning and sparse representation on the time-frequency domain feature matrix to obtain a low-dimensional feature vector; performing kernel principal component analysis and information entropy calculation on the low-dimensional feature vector to obtain a quantitative index of the performance of the wind turbine generator; performing dynamic time warping and sequence pattern mining on the quantitative indexes to obtain abnormal working condition characteristics of the wind turbine generator; and performing fuzzy rough set analysis and fusion processing on the abnormal working condition features to obtain performance evaluation data of the wind turbine generator. The efficiency and accuracy of wind turbine generator performance simulation analysis based on time sequence analysis are improved.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a method and device for simulating and analyzing the performance of a wind turbine generator system based on timing analysis. Background Art

[0002] Existing wind turbine performance analysis methods mainly rely on traditional data statistics and simple time series analysis techniques. These methods usually use a fixed sampling frequency to collect wind turbine operation data and use classic signal processing techniques for filtering and feature extraction. In terms of performance evaluation, threshold-based anomaly detection methods and simple regression models are often used to predict unit performance.

[0003] However, these traditional methods have limitations when dealing with complex wind turbine dynamic characteristics. They have difficulty effectively capturing the nonlinear behavior of wind turbines under different operating conditions, and cannot accurately identify and classify various abnormal operating conditions. In addition, existing methods often have problems of low computational efficiency and insufficient accuracy when dealing with high-dimensional and high-noise wind turbine data. Summary of the invention

[0004] The present application provides a method and device for wind turbine performance simulation analysis based on timing analysis, which are used to improve the efficiency and accuracy of wind turbine performance simulation analysis based on timing analysis.

[0005] In the first aspect, the present application provides a wind turbine performance simulation and analysis method based on time series analysis, and the wind turbine performance simulation and analysis method based on time series analysis includes: high-frequency sampling and adaptive filtering of pre-acquired wind turbine operation data to obtain a denoised multidimensional time series data set; wavelet packet decomposition and Hilbert-Huang transform of the multidimensional time series data set to obtain a time-frequency domain feature matrix; nonlinear manifold learning and sparse representation of the time-frequency domain feature matrix to obtain a low-dimensional feature vector of the wind turbine operation state; kernel principal component analysis and information entropy calculation of the low-dimensional feature vector to obtain quantitative indicators of wind turbine performance; dynamic time warping and sequence pattern mining of the quantitative indicators to obtain abnormal operating condition characteristics of the wind turbine; fuzzy rough set analysis and fusion processing of the abnormal operating condition characteristics to obtain performance evaluation data of the wind turbine.

[0006] In combination with the first aspect, in a first implementation method of the first aspect of the present application, the high-frequency sampling and adaptive filtering of the pre-acquired wind turbine operating data to obtain a denoised multidimensional time series data set includes: performing Fourier transform on the wind turbine operating data to obtain a frequency domain representation; processing the frequency domain representation through a Butterworth high-pass filter to obtain a high-frequency component; performing inverse Fourier transform on the high-frequency component to obtain a time domain high-frequency signal; performing wavelet threshold denoising on the time domain high-frequency signal to obtain a preliminary denoised signal; processing the preliminary denoised signal through a Kalman filter to obtain an optimized denoised signal; performing empirical mode decomposition on the optimized denoised signal to obtain a set of intrinsic mode functions; processing the intrinsic mode function through a Hilbert transform to obtain an instantaneous frequency and an instantaneous amplitude; reconstructing the original signal according to the instantaneous frequency and instantaneous amplitude to obtain a reconstructed signal; processing the reconstructed signal through an adaptive Wiener filter to obtain a final denoised signal; performing multi-dimensional combination and time alignment on the final denoised signal to obtain the denoised multidimensional time series data set.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present application, the multidimensional time series data set is subjected to wavelet packet decomposition and Hilbert-Huang transform to obtain a time-frequency domain feature matrix, including: performing discrete wavelet transform on the multidimensional time series data set to obtain multi-scale wavelet coefficients; processing the multi-scale wavelet coefficients by an optimal basis selection algorithm to obtain an optimal wavelet packet basis; decomposing the multidimensional time series data set according to the optimal wavelet packet basis to obtain wavelet packet coefficients; performing energy calculation on the wavelet packet coefficients to obtain energy distribution of each frequency band; processing the energy distribution of each frequency band by an empirical mode decomposition algorithm to obtain a set of intrinsic mode functions; performing Hilbert transform on the intrinsic mode function to obtain instantaneous frequency and instantaneous amplitude; constructing a Hilbert spectrum according to the instantaneous frequency and instantaneous amplitude to obtain a time-frequency energy distribution; processing the time-frequency energy distribution by a singular value decomposition algorithm to obtain a main eigenvector; normalizing the main eigenvector to obtain standardized features; arranging and combining the standardized features in chronological order to obtain the time-frequency domain feature matrix.

[0008] In combination with the first aspect, in a third implementation of the first aspect of the present application, the time-frequency domain feature matrix is ​​subjected to nonlinear manifold learning and sparse representation to obtain a low-dimensional feature vector of the operating state of the wind turbine, including: locally linearly embedding the time-frequency domain feature matrix to obtain an initial low-dimensional representation; processing the initial low-dimensional representation through a Laplace eigenmapping algorithm to obtain an optimized low-dimensional embedding; performing isotropic diffusion mapping on the optimized low-dimensional embedding to obtain manifold structure features; and visualizing the manifold structure features through a t-SNE algorithm. A two-dimensional scatter plot is obtained; a distance matrix between points is calculated according to the two-dimensional scatter plot to obtain a similarity relationship between data points; the similarity relationship is processed by a spectral clustering algorithm to obtain a data cluster division result; dictionary learning is performed on the original features according to the data cluster division result to obtain an over-complete dictionary; sparse coding is performed on the over-complete dictionary by an orthogonal matching pursuit algorithm to obtain sparse coefficients; principal component analysis is performed on the sparse coefficients to obtain main feature directions; the sparse coefficients are projected along the main feature directions to obtain a low-dimensional feature vector of the operating state of the wind turbine.

[0009] In combination with the first aspect, in a fourth implementation method of the first aspect of the present application, the low-dimensional feature vector is subjected to kernel principal component analysis and information entropy calculation to obtain a quantitative index of wind turbine performance, including: centralizing the low-dimensional feature vector to obtain zero-mean feature data; mapping the zero-mean feature data through a Gaussian kernel function to obtain a high-dimensional feature space representation; constructing a covariance matrix for the high-dimensional feature space representation to obtain a kernel matrix; performing eigenvalue decomposition on the kernel matrix to obtain eigenvalues ​​and eigenvectors; sorting the eigenvectors according to the eigenvalues ​​to obtain a principal component contribution rate; cumulatively summing the principal component contribution rates to obtain a cumulative contribution rate curve; determining the number of principal components according to the cumulative contribution rate curve to obtain a feature representation after dimensionality reduction; performing probability density estimation on the feature representation after dimensionality reduction to obtain a probability distribution function; calculating the Shannon entropy on the probability distribution function to obtain a feature information entropy value; and weightedly combining the information entropy value with the principal component contribution rate to obtain a quantitative index of wind turbine performance.

[0010] In combination with the first aspect, in the fifth implementation method of the first aspect of the present application, the dynamic time warping and sequence pattern mining of the quantitative indicators to obtain abnormal operating characteristics of the wind turbine generator set includes: segmenting the quantitative indicators into time series to obtain multiple subsequences; aligning the multiple subsequences through a dynamic time warping algorithm to obtain a warped time series; performing sliding window segmentation on the warped time series to obtain data segments of fixed length; performing discrete Fourier transform on the fixed-length data segments to obtain frequency domain feature representation; clustering the frequency domain feature representation through a self-organizing mapping algorithm to obtain feature pattern clusters; performing frequent pattern mining on the feature pattern clusters to obtain recurring subsequence patterns; calculating support and confidence for the recurring subsequence patterns to obtain an association rule set; pruning and merging the association rule set to obtain a streamlined rule representation; classifying the streamlined rule representation through a decision tree algorithm to obtain abnormal pattern recognition rules; labeling the original sequence according to the abnormal pattern recognition rules to obtain the abnormal operating characteristics of the wind turbine generator set.

[0011] In combination with the first aspect, in a sixth implementation method of the first aspect of the present application, the abnormal operating condition characteristics are subjected to fuzzy rough set analysis and fusion processing to obtain performance evaluation data of the wind turbine generator set, including: fuzzifying the abnormal operating condition characteristics to obtain a fuzzy feature set; classifying the fuzzy feature set through a fuzzy C-means clustering algorithm to obtain a fuzzy equivalence class; calculating the fuzzy lower approximation and upper approximation of the fuzzy equivalence class to obtain a fuzzy rough set representation; calculating the fuzzy granularity and fuzzy precision of the fuzzy rough set representation to obtain an uncertainty measure; constructing a fuzzy decision table according to the uncertainty measure to obtain conditional attributes and decision attributes; calculating the attribute importance of the conditional attributes and decision attributes to obtain a core attribute set; extracting rules according to the core attribute set to obtain fuzzy decision rules; performing conflict detection and resolution on the fuzzy decision rules to obtain a consistency rule base; optimizing the consistency rule base through a variable precision fuzzy rough set method to obtain a simplified rule set; evaluating and classifying abnormal operating conditions according to the simplified rule set to obtain performance evaluation data of the wind turbine generator set.

[0012] In a second aspect, the present application provides a wind turbine performance simulation analysis device based on timing analysis, the wind turbine performance simulation analysis device based on timing analysis comprising:

[0013] A filtering module is used to perform high-frequency sampling and adaptive filtering on the pre-acquired wind turbine operating data to obtain a denoised multi-dimensional time series data set;

[0014] A transformation module, used for performing wavelet packet decomposition and Hilbert-Huang transformation on the multidimensional time series data set to obtain a time-frequency domain feature matrix;

[0015] A learning module, used for performing nonlinear manifold learning and sparse representation on the time-frequency domain feature matrix to obtain a low-dimensional feature vector of the wind turbine operating state;

[0016] A calculation module, used for performing kernel principal component analysis and information entropy calculation on the low-dimensional feature vector to obtain quantitative indicators of wind turbine performance;

[0017] A mining module, used for performing dynamic time warping and sequence pattern mining on the quantitative indicators to obtain abnormal operating characteristics of the wind turbine;

[0018] The fusion module is used to perform fuzzy rough set analysis and fusion processing on the abnormal operating condition characteristics to obtain performance evaluation data of the wind turbine generator set.

[0019] The third aspect of the present application provides a wind turbine performance simulation and analysis device based on timing analysis, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the wind turbine performance simulation and analysis device based on timing analysis executes the above-mentioned wind turbine performance simulation and analysis method based on timing analysis.

[0020] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned wind turbine performance simulation analysis method based on timing analysis.

[0021] In the technical solution provided by the present application, by high-frequency sampling and adaptive filtering of the wind turbine operation data, it is possible to effectively capture transient changes and suppress noise interference, thereby improving the quality and reliability of the original data. The method of combining wavelet packet decomposition and Hilbert-Huang transform can not only realize multi-scale time-frequency analysis, but also accurately extract the instantaneous characteristics of the signal, thereby more comprehensively characterizing the dynamic behavior of the wind turbine. In the feature extraction link, the introduction of nonlinear manifold learning and sparse representation technology effectively reduces the dimension of the data while retaining key information, which greatly improves the efficiency and accuracy of subsequent analysis. The combination of kernel principal component analysis and information entropy calculation not only considers the main change trend of the data, but also quantifies its uncertainty, providing a more comprehensive and reliable indicator for the quantitative evaluation of wind turbine performance. The application of dynamic time warping and sequence pattern mining enables the method to process time series of different lengths and speeds, and find recurring patterns from them, which is of great significance for identifying abnormal working conditions and predicting potential faults. Finally, through fuzzy rough set analysis and fusion processing, the method can effectively deal with the uncertainty and ambiguity in the performance evaluation process of wind turbines, and improve the reliability and robustness of decision-making. Overall, this method constructs a complete analysis chain from data acquisition, feature extraction to performance evaluation through multiple innovative technical features, which significantly improves the accuracy, efficiency and adaptability of wind turbine performance simulation analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 A schematic diagram of an embodiment of a wind turbine performance simulation analysis method based on timing analysis in an embodiment of the present application;

[0024] Figure 2 Schematic diagram of an embodiment of a wind turbine performance simulation and analysis device based on timing analysis in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The embodiment of the present application provides a method and device for simulating and analyzing the performance of a wind turbine generator system based on timing analysis. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of a wind turbine performance simulation analysis method based on timing analysis includes:

[0027] Step S101, performing high-frequency sampling and adaptive filtering on the pre-acquired wind turbine operating data to obtain a denoised multi-dimensional time series data set;

[0028] Step S102, performing wavelet packet decomposition and Hilbert-Huang transform on the multidimensional time series data set to obtain a time-frequency domain feature matrix;

[0029] Step S103, performing nonlinear manifold learning and sparse representation on the time-frequency domain feature matrix to obtain a low-dimensional feature vector of the wind turbine operating state;

[0030] Step S104, performing kernel principal component analysis and information entropy calculation on the low-dimensional feature vector to obtain quantitative indicators of wind turbine performance;

[0031] Step S105, performing dynamic time warping and sequence pattern mining on the quantitative indicators to obtain abnormal operating characteristics of the wind turbine;

[0032] Step S106: Perform fuzzy rough set analysis and fusion processing on the abnormal operating condition characteristics to obtain performance evaluation data of the wind turbine generator set.

[0033] It is understandable that the execution subject of the present application may be a wind turbine performance simulation analysis device based on time series analysis, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0034] Specifically, the wind turbine performance simulation analysis method based on time series analysis of the present invention effectively solves the shortcomings of traditional methods in processing complex nonlinear dynamic characteristics and high-dimensional noise data through a series of innovative data processing and analysis steps. First, the pre-acquired wind turbine operation data is subjected to high-frequency sampling and adaptive filtering. This step uses a sampling frequency higher than that of traditional methods, such as 1000 samplings per second, to capture transient changes in the operation of the wind turbine. The adaptive filter can dynamically adjust its parameters according to the signal characteristics, effectively remove environmental noise and measurement errors, and thus obtain a high-quality multidimensional time series data set. The multidimensional time series data set is subjected to wavelet packet decomposition and Hilbert-Huang transform. Wavelet packet decomposition is an extension of wavelet transform. It decomposes the signal into multiple frequency bands, each with the same bandwidth. This method can more finely analyze the local characteristics of the signal. For example, for the vibration signal of a wind turbine, wavelet packet decomposition can effectively separate the characteristic frequencies of different components (such as blades and gearboxes). Hilbert-Huang transform is an adaptive signal processing method that can process nonlinear and non-stationary signals and extract the instantaneous frequency and instantaneous amplitude of the signal. By combining these two methods, a feature matrix containing rich time-frequency information is obtained. Nonlinear manifold learning and sparse representation are performed on the time-frequency domain feature matrix. Nonlinear manifold learning, such as isomap or local linear embedding (LLE), can discover low-dimensional manifold structures in high-dimensional data, while sparse representation can represent complex signals with a small number of basis functions. This step greatly reduces the dimensionality of the data while retaining key information. For example, for a wind turbine operating status containing hundreds of features, only 10-20 main features may be needed to accurately describe its working status. Kernel principal component analysis and information entropy calculation are performed on the low-dimensional feature vector. Kernel principal component analysis is a nonlinear extension of traditional principal component analysis, which can capture nonlinear relationships in data. Information entropy calculation quantifies the uncertainty of the data. The combination of these two methods enables a comprehensive evaluation of the performance status of wind turbines. For example, the main factors affecting the efficiency of wind turbines, such as wind speed and blade angle, can be identified through kernel principal component analysis, while information entropy can reflect the degree of change of these factors. Dynamic time warping and sequence pattern mining are performed on quantitative indicators. Dynamic time warping (DTW) algorithms can align time series of different lengths and speeds, which is very useful when comparing the performance of different periods or different units. Sequential pattern mining can discover recurring patterns, which is crucial for predicting potential failures. For example, it may be found that under a certain combination of wind speed and temperature, the vibration pattern of a wind turbine will change, which may indicate that certain components are about to fail.

[0035] Finally, the abnormal operating condition characteristics are analyzed and fused by fuzzy rough set. Fuzzy rough set theory combines the advantages of fuzzy sets and rough sets, and can effectively handle uncertain and imprecise information. Through this step, a robust decision support system can be established to provide a reliable basis for the performance evaluation of wind turbines. For example, a fuzzy decision table can be constructed based on multiple performance indicators (such as power generation efficiency, vibration level, temperature, etc.), and then the overall performance rating of the wind turbine can be obtained through fuzzy rough set analysis.

[0036] This method can comprehensively and accurately analyze the performance of offshore wind turbines. For example, for a 300MW offshore wind farm, including 50 6MW wind turbines, the operating status of each unit can be monitored in real time. Various sensor data, including wind speed, rotation speed, power output, vibration, temperature, etc., are collected through high-frequency sampling (such as 1000 times per second). After adaptive filtering, clear multi-dimensional time series data is obtained. Wavelet packet decomposition and Hilbert-Huang transform help identify the characteristic frequency of each unit, such as finding that the blades of a certain unit vibrate abnormally at a specific frequency. Nonlinear manifold learning and sparse representation compress features that may have hundreds of dimensions to about 20 dimensions, greatly improving the efficiency of subsequent analysis. Kernel principal component analysis and information entropy calculation make it possible to quantify the performance status of each unit. For example, it may be found that the power generation efficiency of unit 5 is 10% lower than the average level, and the uncertainty (entropy) of its operating parameters is higher than that of other units. Through dynamic time warping, the performance of different units under similar wind conditions can be compared, and sequence pattern mining helps discover a specific power fluctuation pattern that often appears before gearbox failure. Finally, fuzzy rough set analysis comprehensively considers multiple performance indicators and gives a comprehensive score for each unit, helping operation and maintenance personnel to prioritize the units with poor performance. With this method, wind farm operators can better understand the performance characteristics of each unit, identify potential problems in a timely manner, and optimize maintenance plans, thereby improving the operational efficiency and reliability of the entire wind farm.

[0037] In the embodiment of the present application, by high-frequency sampling and adaptive filtering of the wind turbine operation data, transient changes can be effectively captured and noise interference can be suppressed, thereby improving the quality and reliability of the original data. The method of combining wavelet packet decomposition and Hilbert-Huang transform can not only realize multi-scale time-frequency analysis, but also accurately extract the instantaneous characteristics of the signal, thereby more comprehensively characterizing the dynamic behavior of the wind turbine. In the feature extraction link, the introduction of nonlinear manifold learning and sparse representation technology effectively reduces the dimension of the data while retaining key information, which greatly improves the efficiency and accuracy of subsequent analysis. The combination of kernel principal component analysis and information entropy calculation not only considers the main change trend of the data, but also quantifies its uncertainty, providing a more comprehensive and reliable indicator for the quantitative evaluation of wind turbine performance. The application of dynamic time warping and sequence pattern mining enables the method to process time series of different lengths and speeds, and find recurring patterns from them, which is of great significance for identifying abnormal conditions and predicting potential faults. Finally, through fuzzy rough set analysis and fusion processing, the method can effectively deal with the uncertainty and ambiguity in the process of wind turbine performance evaluation, and improve the reliability and robustness of decision-making. Overall, this method constructs a complete analysis chain from data acquisition, feature extraction to performance evaluation through multiple innovative technical features, which significantly improves the accuracy, efficiency and adaptability of wind turbine performance simulation analysis.

[0038] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0039] (1) Perform Fourier transform on the wind turbine operating data to obtain frequency domain representation;

[0040] (2) Processing the frequency domain representation through a Butterworth high-pass filter to obtain high-frequency components;

[0041] (3) Perform inverse Fourier transform on the high-frequency component to obtain a high-frequency signal in the time domain;

[0042] (4) Perform wavelet threshold denoising on the high-frequency signal in the time domain to obtain a preliminary denoised signal;

[0043] (5) Processing the preliminary denoised signal through a Kalman filter to obtain an optimized denoised signal;

[0044] (6) performing empirical mode decomposition on the optimized denoised signal to obtain a set of intrinsic mode functions;

[0045] (7) Processing the intrinsic mode function through Hilbert transform to obtain the instantaneous frequency and instantaneous amplitude;

[0046] (8) reconstructing the original signal according to the instantaneous frequency and the instantaneous amplitude to obtain a reconstructed signal;

[0047] (9) Processing the reconstructed signal through an adaptive Wiener filter to obtain the final denoised signal;

[0048] (10) The final denoised signal is subjected to multi-dimensional combination and time alignment to obtain a denoised multi-dimensional time series data set.

[0049] Specifically, in the performance analysis of offshore wind turbines, the preprocessing of raw data is crucial. This method first performs Fourier transform on the wind turbine operation data to convert the time domain signal into a frequency domain representation. Fourier transform is a mathematical tool that decomposes the time domain signal into a superposition of sine waves of different frequencies, which can reveal the periodic characteristics hidden in the signal. For example, for a 6MW offshore wind turbine, after the blade speed signal is Fourier transformed, an obvious peak may be seen in the spectrum, corresponding to the rotation frequency of the blade (such as 15 revolutions per minute, i.e. 0.25Hz). The frequency domain representation is processed by a Butterworth high-pass filter to obtain a high-frequency component. The Butterworth high-pass filter is a filter that can smoothly attenuate low-frequency signals while retaining high-frequency signals. In the analysis of wind turbines, high-frequency components usually contain important information about the dynamic characteristics and potential faults of the unit. For example, the cutoff frequency may be set to 10Hz to filter out low-frequency interference such as wind speed changes and retain high-frequency vibration information of components such as gearboxes and bearings. The high-frequency component is inverse Fourier transformed to obtain a high-frequency signal in the time domain. This step converts the frequency domain information back to the time domain, making it possible to observe changes in high-frequency signals on a time scale. For example, it may be found that the high-frequency vibration signal suddenly increases during a certain period of time, which may indicate that there is a problem with the bearing or gearbox.

[0050] The high-frequency signal in the time domain is denoised by wavelet threshold to obtain a preliminary denoised signal. Wavelet threshold denoising is an effective noise suppression method that can remove random noise while retaining the important features of the signal. In practice, the db4 wavelet and soft threshold method may be selected, and the threshold is set to 3 times the standard deviation of the noise, which can effectively remove environmental noise while retaining the key characteristics of the unit operation. The preliminary denoised signal is processed by a Kalman filter to obtain an optimized denoised signal. The Kalman filter is a recursive estimator that can comprehensively consider measurement noise and system dynamics to provide an optimal estimate. In wind turbine analysis, the Kalman filter can effectively handle sensor drift and system state changes. For example, for wind speed signals, the Kalman filter can smooth short-term fluctuations while quickly responding to real wind speed changes.

[0051] Perform empirical mode decomposition on the optimized denoised signal to obtain a set of intrinsic mode functions. Empirical mode decomposition (EMD) is an adaptive signal processing method that decomposes complex signals into a series of intrinsic mode functions (IMFs). Each IMF represents an oscillation mode in the original signal. In wind turbine analysis, EMD can help separate the vibration characteristics of different components. For example, the first IMF may represent high-frequency noise, the second IMF may represent the vibration of the blades, and the third IMF may represent the swaying of the tower.

[0052] The intrinsic mode functions are processed by Hilbert transform to obtain the instantaneous frequency and instantaneous amplitude. The Hilbert transform can extract the instantaneous properties of the signal, which is particularly useful when analyzing non-stationary signals. By calculating the Hilbert transform of each IMF, the vibration frequency and intensity of each component of the wind turbine at different times can be obtained. For example, it may be found that under strong wind conditions, the vibration frequency of the blade increases slightly, while the amplitude increases significantly.

[0053] The original signal is reconstructed according to the instantaneous frequency and instantaneous amplitude to obtain a reconstructed signal. This step combines the analysis results of the previous steps to generate a clear signal containing key information. The reconstructed signal retains the main features of the original signal while removing most of the noise and interference.

[0054] The reconstructed signal is processed by an adaptive Wiener filter to obtain the final denoised signal. The adaptive Wiener filter can dynamically adjust its parameters according to the local statistical characteristics of the signal to further optimize the signal quality. In wind turbine analysis, this step can effectively deal with non-stationary noise, such as interference caused by waves or wind shear.

[0055] Finally, the final denoised signals are multi-dimensionally combined and time-aligned to obtain a denoised multi-dimensional time series dataset. This step integrates different types of signals (such as wind speed, power output, vibration, etc.) into a unified dataset and ensures that all data are synchronized in time.

[0056] For example, consider a 10MW offshore wind turbine in the North Sea. One week of operation data was collected, including signals such as wind speed, rotation speed, power output, tower vibration, and gearbox temperature, with a sampling frequency of 100Hz. First, each signal was Fourier transformed, and the fundamental frequency of the blade rotation speed (about 0.16Hz) and its harmonics were observed in the spectrum. After processing through a 20Hz Butterworth high-pass filter, the high-frequency components containing mechanical vibration information were retained. After inverse Fourier transform, the high-frequency signal in the time domain clearly showed the meshing frequency of the gearbox (about 300Hz). Wavelet threshold denoising used the db5 wavelet to effectively remove the influence of sea waves and wind noise. The Kalman filter further optimized the signal, especially when dealing with sudden changes in wind speed. The empirical mode decomposition produced 8 IMFs, of which the third IMF mainly reflected the vibration mode of the blade. The Hilbert transform revealed that under strong wind conditions (above 25m / s), the blade vibration frequency increased slightly (from 1.2Hz to 1.3Hz), while the amplitude increased by about 40%. The reconstructed signal clearly retains these key features. The adaptive Wiener filter effectively suppresses the remaining wave interference, which improves the signal-to-noise ratio of the tower vibration signal by about 6 dB. Finally, a high-quality multidimensional time series data set with 5 dimensions (wind speed, rotation speed, power, vibration, temperature) and a length of 604,800 sampling points (7 days × 24 hours × 3600 seconds × 100 Hz) is obtained.

[0057] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0058] (1) Perform discrete wavelet transform on multidimensional time series data sets to obtain multi-scale wavelet coefficients;

[0059] (2) The multi-scale wavelet coefficients are processed by the best basis selection algorithm to obtain the optimal wavelet packet basis;

[0060] (3) Decompose the multidimensional time series data set according to the optimal wavelet packet basis to obtain the wavelet packet coefficients;

[0061] (4) Calculate the energy of the wavelet packet coefficients to obtain the energy distribution of each frequency band;

[0062] (5) The energy distribution of each frequency band is processed by the empirical mode decomposition algorithm to obtain a set of intrinsic mode functions;

[0063] (6) Performing Hilbert transform on the intrinsic mode function to obtain the instantaneous frequency and instantaneous amplitude;

[0064] (7) constructing the Hilbert spectrum based on the instantaneous frequency and instantaneous amplitude to obtain the time-frequency energy distribution;

[0065] (8) The time-frequency energy distribution is processed by the singular value decomposition algorithm to obtain the main eigenvectors;

[0066] (9) Normalize the main feature vectors to obtain standardized features;

[0067] (10) Arrange and combine the standardized features in chronological order to obtain the time-frequency domain feature matrix.

[0068] Specifically, in the performance analysis of offshore wind turbines, in-depth processing of multidimensional time series data is a key step in extracting effective features. First, the multidimensional time series data set is subjected to discrete wavelet transform to obtain multi-scale wavelet coefficients. Discrete wavelet transform is a mathematical tool that decomposes a signal into different frequencies and time scales, which can capture the details and approximations of the signal at different scales. For example, for the vibration signal of an 8MW offshore wind turbine, the db4 wavelet may be used for 5-level decomposition to obtain 5-scale detail coefficients and 1 approximation coefficient.

[0069] The multi-scale wavelet coefficients are processed by the best basis selection algorithm to obtain the optimal wavelet packet basis. The best basis selection algorithm is a method to find the best representation in the wavelet packet tree. It selects the basis function that best suits the signal characteristics by minimizing a certain cost function (such as entropy). In wind turbine analysis, this step can help find the frequency band that best represents the dynamic characteristics of the unit. For example, by using Shannon entropy as a cost function, it may be found that certain nodes of the third-level decomposition are particularly suitable for describing the vibration characteristics of the gearbox.

[0070] The multidimensional time series data set is decomposed according to the optimal wavelet packet basis to obtain the wavelet packet coefficients. This step projects the original signal onto the optimal basis to obtain a set of coefficients that can best represent the signal characteristics. For the 8MW wind turbine mentioned above, 32 optimal basis functions may be finally selected, corresponding to different frequency bands and time positions.

[0071] Energy calculation is performed on the wavelet packet coefficients to obtain the energy distribution of each frequency band. Energy calculation is usually achieved by summing the squares of the coefficients, which reflects the energy concentration of the signal in each frequency band. In the example, it may be found that there is a significant energy concentration in the frequency band of 20-30Hz, which may correspond to the vibration frequency of a key component of the wind turbine (such as the main shaft).

[0072] The energy distribution of each frequency band is processed by the empirical mode decomposition algorithm to obtain a set of intrinsic mode functions. Empirical mode decomposition (EMD) is an adaptive signal processing method that decomposes complex signals into a series of inherent oscillation modes, called intrinsic mode functions (IMFs). In wind turbine analysis, EMD can help separate the contributions of different physical processes. For example, for the energy distribution signal, 5 IMFs may be obtained, of which the first IMF may represent high-frequency noise, the second IMF may represent blade vibration, and the third IMF may represent tower vibration.

[0073] Then, the intrinsic mode function is Hilbert transformed to obtain the instantaneous frequency and instantaneous amplitude. The Hilbert transform can extract the instantaneous properties of the signal, which is particularly useful when analyzing non-stationary signals. By calculating the Hilbert transform of each IMF, the vibration frequency and intensity of each component of the wind turbine at different times can be obtained. For example, it may be observed that when the wind speed suddenly changes, the instantaneous frequency of the tower vibration rises from 0.3Hz to 0.35Hz, while the amplitude increases by 20%.

[0074] Next, the Hilbert spectrum is constructed based on the instantaneous frequency and instantaneous amplitude to obtain the time-frequency energy distribution. The Hilbert spectrum is a three-dimensional representation that shows the change of signal energy with time and frequency. In the analysis of wind turbines, the Hilbert spectrum may clearly show that under strong wind conditions (such as wind speeds exceeding 20m / s), the energy in the 25-30Hz band increases significantly, which may indicate that the gearbox is under heavy load.

[0075] The time-frequency energy distribution is then processed through a singular value decomposition algorithm to obtain the main eigenvectors. Singular value decomposition (SVD) is a powerful matrix decomposition method that can decompose complex data structures into simpler and more meaningful components. In the example, SVD of the Hilbert spectrum may reveal 3-5 main eigenvectors, which may correspond to key factors such as wind speed variation, blade vibration, gearbox load, etc.

[0076] Next, the main feature vectors are normalized to obtain standardized features. Normalization ensures that features of different scales can be compared fairly. For example, the Z-score normalization method may be used to make each feature vector have a mean of 0 and a standard deviation of 1.

[0077] Finally, the standardized features are arranged and combined in chronological order to obtain the time-frequency domain feature matrix. This step integrates all the previously extracted features into a unified matrix, providing rich input for subsequent analysis.

[0078] For example, consider a 12MW offshore wind turbine located in the North Sea, which collects one day of operating data, including signals such as tower vibration, blade bending moment, gearbox temperature, etc., with a sampling frequency of 1kHz. First, a 5-level db5 wavelet decomposition is performed on each signal to obtain 6 groups of wavelet coefficients (5 details and 1 approximation). Through the optimal basis selection algorithm, it is found that in the 4th level decomposition, the nodes with a frequency range of 31.25-62.5Hz are particularly important for describing the gearbox vibration. According to this optimal basis, the original signal is decomposed by wavelet packet to obtain coefficients of 64 frequency bands.

[0079] Energy calculations show that there is a significant energy concentration in the frequency band of 50-60Hz, which is consistent with the characteristic frequency of the gearbox. EMD is performed on this energy distribution, and 4 IMFs are obtained, among which the second IMF (frequency is about 1Hz) captures the sway of the tower well. Hilbert transform reveals that when the wind speed suddenly increases from 10m / s to 15m / s, the instantaneous frequency of the tower vibration increases from 0.8Hz to 0.9Hz, and the amplitude increases by 30%.

[0080] The constructed Hilbert spectrum clearly shows that the energy in the 55-60Hz band increases sharply during 14:00-15:00, which is consistent with the strong wind weather recorded at that time. SVD is performed on this time-frequency energy distribution to extract five main eigenvectors, of which the first vector (corresponding to the largest singular value) mainly reflects the pattern of wind speed change, and the second vector captures information related to the periodic motion of the blades.

[0081] After normalization, a standardized feature matrix with a dimension of 5 (number of features) × 86400 (number of seconds) is obtained. This matrix comprehensively describes the dynamic behavior of the wind turbine in a day, including information on wind speed changes, structural response, mechanical vibration, and other aspects. For example, by analyzing the time variation of the first eigenvector, the daily variation pattern of wind speed can be clearly seen; while the periodic variation of the second eigenvector reflects the impact of blade rotation on the entire system. This time-frequency domain feature matrix provides rich and accurate input data for subsequent performance evaluation and fault diagnosis.

[0082] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0083] (1) Perform local linear embedding on the time-frequency domain feature matrix to obtain an initial low-dimensional representation;

[0084] (2) The initial low-dimensional representation is processed by the Laplace eigenmap algorithm to obtain an optimized low-dimensional embedding;

[0085] (3) Perform isotropic diffusion mapping on the optimized low-dimensional embedding to obtain manifold structural features;

[0086] (4) The manifold structure features are visualized using the t-SNE algorithm to obtain a two-dimensional scatter plot;

[0087] (5) Calculate the distance matrix between points based on the two-dimensional scatter plot to obtain the similarity relationship between data points;

[0088] (6) The similarity relationship is processed by the spectral clustering algorithm to obtain the data cluster division result;

[0089] (7) Perform dictionary learning on the original features according to the data cluster division results to obtain an overcomplete dictionary;

[0090] (8) Sparsely encode the overcomplete dictionary using an orthogonal matching pursuit algorithm to obtain sparse coefficients;

[0091] (9) Perform principal component analysis on the sparse coefficients to obtain the main feature directions;

[0092] (10) The sparse coefficients are projected along the main feature direction to obtain a low-dimensional feature vector of the wind turbine operating status.

[0093] Specifically, in the performance analysis of offshore wind turbines, deep processing of the time-frequency domain feature matrix is ​​the core step to extract key information. First, the time-frequency domain feature matrix is ​​locally linearly embedded (LLE) to obtain an initial low-dimensional representation. LLE is a nonlinear dimensionality reduction technique that assumes that each high-dimensional data point can be approximated by a linear combination of its neighboring points. In practical applications, for example, when processing a 100-dimensional time-frequency domain feature matrix of a 15MW offshore wind turbine (containing information such as wind speed, power output, temperature of each component, vibration, etc.), LLE may reduce it to 20 dimensions, retaining the main structural features of the data. The initial low-dimensional representation is processed by the Laplace eigenmap algorithm to obtain an optimized low-dimensional embedding. Laplace eigenmap is another nonlinear dimensionality reduction method that attempts to maintain local relationships between data points. In wind turbine data, this step may further optimize the 20-dimensional representation to 15 dimensions, making similar operating states closer in low-dimensional space. The optimized low-dimensional embedding is isotropically diffused to obtain manifold structural features. Diffusion mapping is based on random walk theory and can capture the multi-scale geometric structure of data. In this example, this step may map the 15-dimensional representation to a 10-dimensional manifold structure, clearly showing the evolution trajectory of the wind turbine state over time.

[0094] The manifold structure features are visualized using the t-SNE algorithm to obtain a two-dimensional scatter plot. t-SNE is a very effective high-dimensional data visualization technology that is particularly good at maintaining the local structure of data. Through t-SNE, the 10-dimensional manifold structure can be mapped to a 2-dimensional plane to intuitively display the relationship between different operating states. In this 2-dimensional scatter plot, several obvious clusters may be observed, such as the normal operation cluster, the low wind speed standby cluster, the high wind speed power limit cluster, and some small abnormal clusters.

[0095] Calculate the distance matrix between points based on the two-dimensional scatter plot to obtain the similarity relationship between data points. This step usually uses Euclidean distance to quantify the similarity between points. For one month of running data (assuming the sampling frequency is 1Hz), a 2,592,000×2,592,000 similarity matrix will be obtained.

[0096] The similarity relationship is processed by the spectral clustering algorithm to obtain the data clustering result. Spectral clustering is a clustering method based on graph theory, which uses the similarity relationship between data points for clustering. In the analysis, spectral clustering may divide the data into 10 clusters, 7 of which correspond to different normal operating conditions and 3 correspond to abnormal conditions (such as blade icing, gearbox overheating, yaw system failure).

[0097] Dictionary learning is performed on the original features according to the data clustering results to obtain an overcomplete dictionary. Dictionary learning is a representation learning method that attempts to find a set of basis vectors (called a dictionary) so that the original signal can be represented by a sparse linear combination of these basis vectors. In this example, an overcomplete dictionary containing 50 atoms may be obtained, each of which represents a basic operating mode or failure mode.

[0098] The overcomplete dictionary is sparsely encoded using the orthogonal matching pursuit (OMP) algorithm to obtain sparse coefficients. OMP is a greedy algorithm used to solve sparse representation problems. Through OMP, it may be found that only 5-10 non-zero coefficients are needed on average to well represent the operating status at any moment, which greatly reduces the dimension of the data.

[0099] Perform principal component analysis (PCA) on the sparse coefficients to obtain the main characteristic directions. PCA is a commonly used linear dimensionality reduction method that can find the main direction of change in the data. In the analysis, PCA may find that the first three principal components can explain more than 95% of the variance. These three principal components correspond to wind speed, power output and mechanical load respectively.

[0100] Finally, the sparse coefficients are projected along the main feature direction to obtain the low-dimensional feature vector of the wind turbine operating status. This step projects the high-dimensional sparse coefficients onto the main feature direction determined by PCA to obtain the final 3D feature vector.

[0101] In this 3D feature space, the operation trajectory of the wind turbine can be intuitively observed: the normal operating state forms a smooth surface, while various abnormal states are represented as points that deviate from this surface. For example, it may be observed that on a certain day, the feature vector suddenly deviates from the normal operating trajectory and moves to the vicinity of the previously identified "gearbox overheating" cluster. By tracing back the original data, it is found that the gearbox temperature has indeed increased, although it has not reached the traditional alarm threshold. This early warning capability can help operation and maintenance personnel take preventive measures before the problem worsens, greatly improving the operating efficiency and reliability of offshore wind turbines.

[0102] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0103] (1) Centralize the low-dimensional feature vector to obtain zero-mean feature data;

[0104] (2) Mapping the zero-mean feature data through a Gaussian kernel function to obtain a high-dimensional feature space representation;

[0105] (3) construct a covariance matrix for the high-dimensional feature space representation and obtain a kernel matrix;

[0106] (4) Perform eigenvalue decomposition on the kernel matrix to obtain eigenvalues ​​and eigenvectors;

[0107] (5) Sort the eigenvectors according to the eigenvalues ​​to obtain the principal component contribution rate;

[0108] (6) Cumulatively sum the principal component contribution rates to obtain a cumulative contribution rate curve;

[0109] (7) Determine the number of principal components based on the cumulative contribution rate curve and obtain the feature representation after dimensionality reduction;

[0110] (8) Estimating the probability density of the feature representation after dimensionality reduction to obtain the probability distribution function;

[0111] (9) Calculate the Shannon entropy of the probability distribution function to obtain the information entropy value of the feature;

[0112] (10) The information entropy value and the principal component contribution rate are weightedly combined to obtain the quantitative index of wind turbine performance.

[0113] Specifically, in the performance analysis of offshore wind turbines, in-depth processing of low-dimensional feature vectors is a key step in quantifying performance indicators. First, the low-dimensional feature vectors are centered to obtain zero-mean feature data. This step helps eliminate the scale differences between different features by subtracting the mean value of each feature and moving the data center to the origin. For example, for a 3-dimensional feature vector containing wind speed, power output, and vibration amplitude, after centralization, the mean value of each feature becomes 0.

[0114] Next, the zero-mean feature data is mapped through the Gaussian kernel function to obtain a high-dimensional feature space representation. The Gaussian kernel function is a commonly used kernel function that can implicitly map the original features to an infinite-dimensional feature space, which helps to capture nonlinear relationships. In practical applications, there is no need to explicitly calculate this high-dimensional representation, but to directly calculate the kernel matrix. For example, for two data points x and y, the Gaussian kernel function can be expressed as K(x,y)=exp(-||xy||^2 / (2σ^2)), where σ is the bandwidth parameter.

[0115] Then, a covariance matrix is ​​constructed for the high-dimensional feature space representation to obtain the kernel matrix. The kernel matrix contains the similarity information between all pairs of data points. For n data points, the kernel matrix is ​​an n×n symmetric matrix. For example, for a month of wind turbine operation data (assuming sampling once per second), the dimension of the kernel matrix will be 2,592,000×2,592,000.

[0116] Next, perform eigenvalue decomposition on the kernel matrix to obtain eigenvalues ​​and eigenvectors. Eigenvalue decomposition is an important matrix decomposition method that can reveal the main direction of change of the data. In this example, 2,592,000 eigenvalues ​​and corresponding eigenvectors may be obtained.

[0117] The eigenvectors are sorted according to the eigenvalues ​​to obtain the principal component contribution rate. The principal component contribution rate reflects the degree to which each eigenvector (principal component) explains the variability of the data. Usually, it is found that the first few principal components can explain most of the data variation. For example, it may be found that the first 10 principal components explain 95% of the data variation.

[0118] The cumulative contribution rate curve is obtained by cumulatively summing the principal component contributions. This curve shows how the cumulative explained data variation ratio changes as the number of principal components increases. Usually, this curve will show a shape that rises rapidly and then tends to be flat.

[0119] The number of principal components is determined based on the cumulative contribution rate curve to obtain the feature representation after dimensionality reduction. A common selection method is to select the number of principal components when the cumulative contribution rate reaches a certain threshold (such as 95%). In the example, the first 10 principal components may be selected as the feature representation after dimensionality reduction.

[0120] The probability density of the feature representation after dimensionality reduction is estimated to obtain the probability distribution function. This step usually uses non-parametric methods such as kernel density estimation to avoid making strong assumptions about the data distribution. The obtained probability distribution function describes the distribution of features in the dimensionality reduction space.

[0121] The Shannon entropy is calculated for the probability distribution function to obtain the characteristic information entropy value. Shannon entropy is an important indicator for measuring uncertainty in information theory. In the performance analysis of wind turbines, a higher entropy value may indicate the instability or increased complexity of the system state.

[0122] Finally, the information entropy value and the principal component contribution rate are weighted together to obtain a quantitative index of wind turbine performance. This combination takes into account the main change direction of the data (through the principal component contribution rate) and complexity (through information entropy), providing a comprehensive performance evaluation.

[0123] Let's take a specific example: suppose we want to analyze the operating data of a 15MW offshore wind turbine. The initial 3D feature vector includes the standardized wind speed, power output, and main shaft vibration amplitude. After centering, the mean of these features becomes 0. After applying the Gaussian kernel function (assuming σ = 1), a kernel matrix of 2,592,000×2,592,000 is obtained. After eigenvalue decomposition, it is found that the contribution rates of the first five principal components are 50%, 30%, 10%, 5%, and 2%, respectively. The cumulative contribution rate curve shows that the first four principal components can explain 95% of the data variation.

[0124] Select these four principal components as the feature representation after dimensionality reduction. Perform kernel density estimation on these four-dimensional data to obtain the probability distribution function. The calculated Shannon entropy value is 2.5 bits. This entropy value is weighted and combined with the contribution rate of the first four principal components (assuming the weights are 0.5 and 0.5 respectively), and the final performance quantification index is 3.725.

[0125] This indicator comprehensively reflects the operating status of the wind turbine: a higher value indicates that the system is in a complex or unstable state and may need further inspection; while a lower value indicates that the system is operating stably. By continuously monitoring the changes in this indicator, operation and maintenance personnel can promptly identify potential problems, optimize operation strategies, and improve the overall performance and reliability of offshore wind turbines.

[0126] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0127] (1) Segment the time series of quantitative indicators to obtain multiple subsequences;

[0128] (2) Align multiple subsequences using a dynamic time warping algorithm to obtain a regularized time series;

[0129] (3) Segment the regularized time series by sliding windows to obtain data segments of fixed length;

[0130] (4) Performing discrete Fourier transform on fixed-length data segments to obtain frequency domain feature representation;

[0131] (5) Clustering the frequency domain feature representation using a self-organizing mapping algorithm to obtain feature pattern clusters;

[0132] (6) Perform frequent pattern mining on the feature pattern clusters to obtain recurring subsequence patterns;

[0133] (7) Calculate the support and confidence of the repeated subsequence patterns to obtain the association rule set;

[0134] (8) Prune and merge the association rule sets to obtain a streamlined rule representation;

[0135] (9) Classifying the simplified rule representation through a decision tree algorithm to obtain abnormal pattern recognition rules;

[0136] (10) The original sequence is labeled according to the abnormal pattern recognition rules to obtain the abnormal operating characteristics of the wind turbine.

[0137] Specifically, in the performance analysis of offshore wind turbines, in-depth time series analysis of quantitative indicators is a key step in identifying abnormal operating conditions. First, the quantitative indicators are segmented into time series to obtain multiple subsequences. This step divides the long time series into multiple shorter subsequences by identifying mutation points or feature points in the time series. For example, for a monthly performance quantitative indicator series of a 15MW offshore wind turbine, it may be divided into 200 subsequences based on significant change points in wind speed or power output, and the average length of each subsequence is about 3.6 hours. Multiple subsequences are aligned using the dynamic time warping (DTW) algorithm to obtain a regularized time series. The DTW algorithm is a method for measuring the similarity of two time series, which can handle nonlinear deformations on the time scale. In the example, DTW can help align subsequences of different lengths, such as comparing a 2-hour low wind speed operation sequence with a 4-hour high wind speed operation sequence.

[0138] The regularized time series is segmented by sliding windows to obtain data segments of fixed length. Sliding windows are a commonly used time series data processing technique that can capture local features. Assuming that 30 minutes is selected as the window length and the step length is 5 minutes, 38 overlapping data segments will be obtained for a 3.6-hour subsequence. Discrete Fourier transform (DFT) is performed on the fixed-length data segments to obtain frequency domain feature representation. DFT can convert time domain signals into frequency domain representations and reveal the periodic characteristics of signals. In wind turbine analysis, this step can help identify important features such as blade rotation frequency and gear meshing frequency. For example, for a 30-minute data segment (assuming a sampling frequency of 1Hz), 1,800 complex frequency components will be obtained.

[0139] The frequency domain feature representation is clustered using the self-organizing map (SOM) algorithm to obtain feature pattern clusters. SOM is an unsupervised learning method that can map high-dimensional data to a low-dimensional space for visualization and clustering. In this case, a 10×10 SOM network may be used to map the 1800-dimensional frequency domain features to 100 neurons to form several feature pattern clusters.

[0140] Frequent pattern mining is performed on the characteristic pattern cluster to obtain recurring subsequence patterns. Frequent pattern mining is a technique for finding frequently occurring subsequences in sequence data. In wind turbine analysis, these frequent patterns may represent certain specific operating conditions or precursors to failures. For example, a pattern of "rapid increase in wind speed - rapid increase in power - slight increase in vibration amplitude" may be found to occur frequently.

[0141] The support and confidence of the recurring subsequence patterns are calculated to obtain the association rule set. The support indicates the frequency of a pattern in the entire data set, while the confidence indicates the probability that the appearance of one pattern will lead to the appearance of another pattern. For example, it may be found that the support of the rule "If the wind speed rises by more than 5m / s in 10 minutes, then the power output will increase by more than 20% in the next 5 minutes" is 0.05 (that is, this situation is observed in 5% of the time) and the confidence is 0.9 (that is, in 90% of the cases, a rapid increase in wind speed will lead to a rapid increase in power).

[0142] Prune and merge the association rule set to obtain a streamlined rule representation. This step aims to remove redundant or unimportant rules and improve the interpretability and practicality of the rule set. For example, two similar rules may be merged, or those with low support may be deleted.

[0143] The simplified rule representation is classified by the decision tree algorithm to obtain the abnormal pattern recognition rules. The decision tree is an intuitive classification method that makes decisions through a series of if-then rules. In the example, the decision tree may generate the following rules: "If the power output suddenly drops by more than 30%, the wind speed does not change significantly, and the gearbox temperature rises by more than 10°C, then it is determined to be a gearbox failure."

[0144] Finally, the original sequence is annotated according to the abnormal pattern recognition rules to obtain the abnormal operating characteristics of the wind turbine. This step applies the previously obtained rules to the original time series to identify possible abnormal points or abnormal intervals.

[0145] For example, suppose we analyze the operating data of a 20MW offshore wind turbine in the South China Sea. During a one-month monitoring period, multidimensional time series data including wind speed, power output, temperature of each component, vibration, etc. were collected with a sampling frequency of 1Hz. First, the data for the entire month was divided into 300 subsequences according to the significant change points of wind speed. These subsequences were aligned to a standard length (such as 4 hours) using the DTW algorithm. Then, these regularized sequences were segmented using a 30-minute sliding window (step size 5 minutes) to obtain thousands of data fragments.

[0146] After performing DFT on each data segment, 1800-dimensional frequency domain features were obtained. These features were clustered using a 10×10 SOM network to form 100 neurons representing different operating modes. Through frequent pattern mining, several important patterns were discovered, such as "rapid increase in wind speed-rapid increase in power-blade pitch angle adjustment", "stable wind speed-power fluctuation-slow increase in gearbox temperature", etc.

[0147] After calculating the support and confidence of these patterns, a series of association rules were obtained. After pruning and merging, the 20 most important rules were retained. Using these rules to build a decision tree, a classifier that can identify multiple abnormal working conditions was obtained. For example, one of the rules is: if the power output fluctuation exceeds 5% of the rated power and lasts for more than 10 minutes when the wind speed is stable, and the main shaft vibration amplitude increases by more than 20%, it is determined that there may be a transmission system failure.

[0148] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0149] (1) Perform fuzzy processing on the abnormal operating condition characteristics to obtain a fuzzy feature set;

[0150] (2) Classify the fuzzy feature set using the fuzzy C-means clustering algorithm to obtain fuzzy equivalence classes;

[0151] (3) Calculate the fuzzy lower approximation and upper approximation of the fuzzy equivalence class to obtain the fuzzy rough set representation;

[0152] (4) Calculate the fuzzy granularity and fuzzy precision of the fuzzy rough set representation to obtain the uncertainty measure;

[0153] (5) Construct a fuzzy decision table based on uncertainty measurement to obtain condition attributes and decision attributes;

[0154] (6) Calculate the attribute importance of conditional attributes and decision attributes to obtain the core attribute set;

[0155] (7) Extract rules based on the core attribute set to obtain fuzzy decision rules;

[0156] (8) Detect and resolve conflicts among fuzzy decision rules to obtain a consistent rule base;

[0157] (9) The consistency rule base is optimized by using the variable precision fuzzy rough set method to obtain a simplified rule set;

[0158] (10) Abnormal operating conditions are evaluated and classified according to a simplified rule set to obtain performance evaluation data of the wind turbine.

[0159] Specifically, in the performance analysis of offshore wind turbines, fuzzy-rough set analysis of abnormal operating conditions is a complex and critical process. First, the abnormal operating conditions are fuzzified to obtain fuzzy feature sets. This step converts precise values ​​into fuzzy sets to better handle the uncertainty in the data. For example, for a 15MW offshore wind turbine, the power output may be divided into three fuzzy sets of "low", "medium", and "high", and each value has a membership to these sets. Specifically, a power output of 5MW may have a membership of 0.8 to the "low" set, a membership of 0.2 to the "medium" set, and a membership of 0 to the "high" set.

[0160] The fuzzy feature set is classified by the fuzzy C-means clustering algorithm to obtain fuzzy equivalence classes. Fuzzy C-means clustering is a soft clustering method that allows data points to belong to multiple clusters to varying degrees. In this example, the abnormal operating condition features may be divided into five fuzzy equivalence classes: blade icing, gearbox failure, yaw system abnormality, generator overheating, and control system failure. Each data point has a membership vector belonging to these categories. The fuzzy lower approximation and upper approximation are calculated for the fuzzy equivalence classes to obtain a fuzzy rough set representation. The fuzzy lower approximation contains elements that definitely belong to the set, while the fuzzy upper approximation contains elements that may belong to the set. For example, for the equivalence class "gearbox failure", the fuzzy lower approximation may include those states with large vibration amplitude (>5mm / s) and high temperature (>80℃), while the fuzzy upper approximation may also include states where the vibration or temperature is only slightly abnormal.

[0161] The fuzzy granularity and fuzzy precision are calculated for the fuzzy rough set representation to obtain the uncertainty measure. The fuzzy granularity reflects the level of classification, while the fuzzy precision reflects the accuracy of the classification. For example, if the fuzzy granularity is 0.8 and the fuzzy precision is 0.75, this means that the classification scheme has a certain degree of effectiveness, but there is still room for improvement. A fuzzy decision table is constructed based on the uncertainty measure to obtain conditional attributes and decision attributes. In this table, conditional attributes may include wind speed, power output, vibration amplitude, etc., while the decision attribute is the type of abnormal working condition. For example, a row may represent: wind speed (high, 0.9), power output (low, 0.8), vibration amplitude (high, 0.7), temperature (high, 0.8), decision (gearbox failure, 0.9). The attribute importance is calculated for the conditional attributes and decision attributes to obtain the core attribute set. This step identifies the attributes that are most critical to the decision. In the example, it may be found that vibration characteristics, temperature characteristics, and power curve deviation are the three most important attributes, and their importance is 0.9, 0.85, and 0.8, respectively.

[0162] According to the core attribute set, rule extraction is performed to obtain fuzzy decision rules. These rules are in the form of IF-THEN, for example: "IF the vibration amplitude is high (>0.8) AND the gearbox temperature is high (>0.7) AND the power output is low (>0.6) THEN the probability of gearbox failure is high (0.9)".

[0163] Conflict detection and resolution are performed on fuzzy decision rules to obtain a consistent rule base. This step resolves possible contradictory rules. For example, if there are two rules that come to different conclusions under similar conditions, it may be necessary to introduce additional conditions to distinguish the two cases, or merge the rules.

[0164] The consistency rule base is optimized by the variable precision fuzzy rough set method to obtain a simplified rule set. This method allows a certain degree of classification error, thereby obtaining a more concise and robust rule set. For example, a 5% classification error may be allowed to simplify the original 50 rules into 15 most representative rules.

[0165] Finally, the abnormal conditions are evaluated and classified according to the simplified rule set to obtain the performance evaluation data of the wind turbine. This step applies the simplified rule set to the actual data to comprehensively evaluate the performance of the wind turbine.

[0166] For example, consider a 20MW offshore wind turbine in the East China Sea, which collects one month of operating data, including wind speed, power output, temperature of each component, vibration and other features. First, these features are fuzzified. For example, the power output is divided into three fuzzy sets: "low" (0-7MW), "medium" (5-15MW), and "high" (13-20MW). For a power output of 13MW, it may have a membership of 0.2 to the "medium" set and a membership of 0.8 to the "high" set.

[0167] Using the fuzzy C-means clustering algorithm, the abnormal conditions are divided into 5 categories. For each data point, a 5-dimensional membership vector is obtained, indicating the degree to which it belongs to these 5 categories. For example, a state may have a membership vector of [0.1, 0.7, 0.1, 0.05, 0.05], which is likely to be a gearbox failure.

[0168] After calculating the fuzzy lower and upper approximations, a fuzzy rough set representation is obtained. For example, for "gearbox failure", the fuzzy lower approximation includes those states with high vibration amplitude (>5mm / s, membership >0.8) and high gearbox temperature (>80℃, membership >0.7), while the fuzzy upper approximation also includes states where the vibration or temperature is only slightly abnormal.

[0169] The calculated fuzzy granularity is 0.82 and the fuzzy precision is 0.78, which indicates that the classification scheme has good effectiveness. Based on these metrics, a fuzzy decision table is constructed, which contains 8 conditional attributes and 1 decision attribute.

[0170] Through the calculation of attribute importance, it was found that vibration characteristics (importance 0.92), temperature characteristics (importance 0.88) and power curve deviation (importance 0.85) are the three most important attributes. Based on these core attributes, 40 fuzzy decision rules were extracted. After conflict detection and resolution, 35 consistency rules were obtained.

[0171] Using the variable precision fuzzy rough set method, allowing 3% classification error, the rule set was further simplified to 12 most representative rules. One of the rules is: "IF the vibration amplitude is high (membership > 0.85) AND the gearbox temperature is high (membership > 0.8) AND the power output is low (membership > 0.7) THEN the probability of gearbox failure is high (membership = 0.95)".

[0172] The above describes the wind turbine performance simulation analysis method based on timing analysis in the embodiment of the present application. The following describes the wind turbine performance simulation analysis device based on timing analysis in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a wind turbine performance simulation analysis device based on timing analysis includes:

[0173] The filtering module 201 is used to perform high-frequency sampling and adaptive filtering on the pre-acquired wind turbine operating data to obtain a denoised multi-dimensional time series data set;

[0174] A transformation module 202 is used to perform wavelet packet decomposition and Hilbert-Huang transform on the multi-dimensional time series data set to obtain a time-frequency domain feature matrix;

[0175] A learning module 203 is used to perform nonlinear manifold learning and sparse representation on the time-frequency domain feature matrix to obtain a low-dimensional feature vector of the wind turbine operating state;

[0176] A calculation module 204 is used to perform kernel principal component analysis and information entropy calculation on the low-dimensional feature vector to obtain a quantitative index of the performance of the wind turbine generator set;

[0177] A mining module 205 is used to perform dynamic time warping and sequence pattern mining on the quantitative indicators to obtain abnormal operating characteristics of the wind turbine generator set;

[0178] The fusion module 206 is used to perform fuzzy rough set analysis and fusion processing on the abnormal operating condition characteristics to obtain performance evaluation data of the wind turbine generator set.

[0179] Through the synergy of the above components, high-frequency sampling and adaptive filtering of wind turbine operation data can effectively capture transient changes and suppress noise interference, thereby improving the quality and reliability of the original data. The method of combining wavelet packet decomposition and Hilbert-Huang transform can not only realize multi-scale time-frequency analysis, but also accurately extract the instantaneous characteristics of the signal, thereby more comprehensively describing the dynamic behavior of the wind turbine. In the feature extraction link, the introduction of nonlinear manifold learning and sparse representation technology effectively reduces the dimension of the data while retaining key information, which greatly improves the efficiency and accuracy of subsequent analysis. The combination of kernel principal component analysis and information entropy calculation not only considers the main change trend of the data, but also quantifies its uncertainty, providing a more comprehensive and reliable indicator for the quantitative evaluation of wind turbine performance. The application of dynamic time warping and sequence pattern mining enables the method to process time series of different lengths and speeds and find recurring patterns from them, which is of great significance for identifying abnormal conditions and predicting potential faults. Finally, through fuzzy rough set analysis and fusion processing, this method can effectively deal with the uncertainty and ambiguity in the process of wind turbine performance evaluation, and improve the reliability and robustness of decision-making. Overall, this method constructs a complete analysis chain from data acquisition, feature extraction to performance evaluation through multiple innovative technical features, which significantly improves the accuracy, efficiency and adaptability of wind turbine performance simulation analysis.

[0180] The present application also provides a wind turbine performance simulation and analysis device based on timing analysis, the wind turbine performance simulation and analysis device based on timing analysis includes a memory and a processor, the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the wind turbine performance simulation and analysis method based on timing analysis in the above-mentioned embodiments.

[0181] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the wind turbine performance simulation analysis method based on timing analysis.

[0182] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0183] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A wind turbine performance simulation analysis method based on timing analysis, characterized in that: The wind turbine performance simulation analysis method based on timing analysis includes: Perform high-frequency sampling and adaptive filtering on the pre-acquired wind turbine operating data to obtain a denoised multi-dimensional time series data set; Performing wavelet packet decomposition and Hilbert-Huang transform on the multidimensional time series data set to obtain a time-frequency domain feature matrix; Performing nonlinear manifold learning and sparse representation on the time-frequency domain feature matrix to obtain a low-dimensional feature vector of the wind turbine operating state; Performing kernel principal component analysis and information entropy calculation on the low-dimensional feature vector to obtain quantitative indicators of wind turbine performance; Performing dynamic time warping and sequence pattern mining on the quantitative indicators to obtain abnormal operating characteristics of the wind turbine generator set; The abnormal operating condition characteristics are subjected to fuzzy rough set analysis and fusion processing to obtain performance evaluation data of the wind turbine generator set.

2. The wind turbine performance simulation analysis method based on timing analysis according to claim 1 is characterized in that: The high-frequency sampling and adaptive filtering of the pre-acquired wind turbine operating data to obtain a denoised multi-dimensional time series data set includes: Performing Fourier transformation on the wind turbine operating data to obtain frequency domain representation; Processing the frequency domain representation through a Butterworth high-pass filter to obtain a high-frequency component; Performing inverse Fourier transform on the high-frequency component to obtain a time-domain high-frequency signal; Performing wavelet threshold denoising on the time domain high-frequency signal to obtain a preliminary denoised signal; Processing the preliminary denoised signal through a Kalman filter to obtain an optimized denoised signal; Performing empirical mode decomposition on the optimized denoised signal to obtain a set of intrinsic mode functions; Processing the intrinsic mode function through Hilbert transform to obtain instantaneous frequency and instantaneous amplitude; Reconstructing the original signal according to the instantaneous frequency and the instantaneous amplitude to obtain a reconstructed signal; Processing the reconstructed signal through an adaptive Wiener filter to obtain a final denoised signal; The final denoised signal is subjected to multi-dimensional combination and time alignment to obtain the denoised multi-dimensional time series data set.

3. The wind turbine performance simulation analysis method based on timing analysis according to claim 1 is characterized in that: The multidimensional time series data set is subjected to wavelet packet decomposition and Hilbert-Huang transform to obtain a time-frequency domain feature matrix, including: Performing discrete wavelet transform on the multidimensional time series data set to obtain multi-scale wavelet coefficients; Processing the multi-scale wavelet coefficients by an optimal basis selection algorithm to obtain an optimal wavelet packet basis; Decomposing the multidimensional time series data set according to the optimal wavelet packet basis to obtain wavelet packet coefficients; Performing energy calculation on the wavelet packet coefficients to obtain energy distribution of each frequency band; Processing the energy distribution of each frequency band by an empirical mode decomposition algorithm to obtain a set of intrinsic mode functions; Performing Hilbert transform on the intrinsic mode function to obtain instantaneous frequency and instantaneous amplitude; Constructing a Hilbert spectrum according to the instantaneous frequency and the instantaneous amplitude to obtain a time-frequency energy distribution; Processing the time-frequency energy distribution by a singular value decomposition algorithm to obtain a main eigenvector; Normalizing the main feature vector to obtain standardized features; The standardized features are arranged and combined in time order to obtain the time-frequency domain feature matrix.

4. The wind turbine performance simulation analysis method based on timing analysis according to claim 1 is characterized in that: The nonlinear manifold learning and sparse representation of the time-frequency domain feature matrix to obtain a low-dimensional feature vector of the wind turbine operating state includes: Performing local linear embedding on the time-frequency domain feature matrix to obtain an initial low-dimensional representation; Processing the initial low-dimensional representation through a Laplace eigenmapping algorithm to obtain an optimized low-dimensional embedding; Performing isotropic diffusion mapping on the optimized low-dimensional embedding to obtain manifold structure features; The manifold structure features are visualized using a t-SNE algorithm to obtain a two-dimensional scatter plot; Calculate the distance matrix between points according to the two-dimensional scatter plot to obtain the similarity relationship between data points; Processing the similarity relationship by using a spectral clustering algorithm to obtain a data cluster division result; Performing dictionary learning on the original features according to the data cluster division result to obtain an overcomplete dictionary; Performing sparse coding on the overcomplete dictionary by using an orthogonal matching pursuit algorithm to obtain sparse coefficients; Performing principal component analysis on the sparse coefficients to obtain main feature directions; The sparse coefficients are projected along the main feature direction to obtain a low-dimensional feature vector of the operating state of the wind turbine generator set.

5. The wind turbine performance simulation analysis method based on timing analysis according to claim 1 is characterized in that: The method of performing kernel principal component analysis and information entropy calculation on the low-dimensional feature vector to obtain quantitative indicators of wind turbine performance includes: Centralizing the low-dimensional feature vector to obtain zero-mean feature data; Mapping the zero-mean feature data through a Gaussian kernel function to obtain a high-dimensional feature space representation; constructing a covariance matrix for the high-dimensional feature space representation to obtain a kernel matrix; Performing eigenvalue decomposition on the kernel matrix to obtain eigenvalues ​​and eigenvectors; Sorting the eigenvectors according to the eigenvalues ​​to obtain the principal component contribution rate; Cumulatively summing the principal component contribution rates to obtain a cumulative contribution rate curve; Determine the number of principal components according to the cumulative contribution rate curve to obtain a feature representation after dimensionality reduction; Performing probability density estimation on the feature representation after dimensionality reduction to obtain a probability distribution function; Calculating the Shannon entropy for the probability distribution function to obtain an information entropy value of the feature; The information entropy value and the principal component contribution rate are weightedly combined to obtain a quantitative index of the wind turbine performance.

6. The wind turbine performance simulation analysis method based on timing analysis according to claim 1 is characterized in that: The dynamic time warping and sequence pattern mining of the quantitative indicators are performed to obtain abnormal operating characteristics of the wind turbine generator set, including: Segmenting the quantitative index into time series to obtain multiple subsequences; Aligning the multiple subsequences using a dynamic time warping algorithm to obtain a warped time series; Perform sliding window segmentation on the regularized time series to obtain data segments of fixed length; Performing discrete Fourier transform on the fixed-length data segment to obtain a frequency domain feature representation; Clustering the frequency domain feature representations using a self-organizing mapping algorithm to obtain feature pattern clusters; Performing frequent pattern mining on the characteristic pattern cluster to obtain recurring subsequence patterns; Calculating support and confidence for the repeated subsequence patterns to obtain an association rule set; Pruning and merging the association rule sets to obtain a streamlined rule representation; Classifying the simplified rule representations through a decision tree algorithm to obtain abnormal pattern recognition rules; The original sequence is labeled according to the abnormal pattern recognition rule to obtain the abnormal operating condition characteristics of the wind turbine generator set.

7. The wind turbine performance simulation analysis method based on timing analysis according to claim 1 is characterized in that: The fuzzy rough set analysis and fusion processing of the abnormal operating condition characteristics are performed to obtain the performance evaluation data of the wind turbine generator set, including: Performing fuzzy processing on the abnormal operating condition characteristics to obtain a fuzzy feature set; Classifying the fuzzy feature set by using a fuzzy C-means clustering algorithm to obtain fuzzy equivalence classes; Calculating a fuzzy lower approximation and an upper approximation for the fuzzy equivalence class to obtain a fuzzy rough set representation; Calculating fuzzy granularity and fuzzy precision for the fuzzy rough set representation to obtain uncertainty measurement; Construct a fuzzy decision table according to the uncertainty measurement to obtain condition attributes and decision attributes; Calculating the attribute importance of the condition attributes and decision attributes to obtain a core attribute set; Extracting rules according to the core attribute set to obtain fuzzy decision rules; Performing conflict detection and resolution on the fuzzy decision rules to obtain a consistent rule base; The consistency rule base is optimized by using a variable precision fuzzy rough set method to obtain a simplified rule set; Abnormal operating conditions are evaluated and classified according to the simplified rule set to obtain performance evaluation data of the wind turbine generator set.

8. A wind turbine performance simulation analysis device based on time series analysis, characterized in that: The wind turbine performance simulation analysis device based on timing analysis includes: A filtering module is used to perform high-frequency sampling and adaptive filtering on the pre-acquired wind turbine operating data to obtain a denoised multi-dimensional time series data set; A transformation module, used for performing wavelet packet decomposition and Hilbert-Huang transformation on the multidimensional time series data set to obtain a time-frequency domain feature matrix; A learning module, used for performing nonlinear manifold learning and sparse representation on the time-frequency domain feature matrix to obtain a low-dimensional feature vector of the wind turbine operating state; A calculation module, used for performing kernel principal component analysis and information entropy calculation on the low-dimensional feature vector to obtain quantitative indicators of wind turbine performance; A mining module, used for performing dynamic time warping and sequence pattern mining on the quantitative indicators to obtain abnormal operating characteristics of the wind turbine; The fusion module is used to perform fuzzy rough set analysis and fusion processing on the abnormal operating condition characteristics to obtain performance evaluation data of the wind turbine.

9. A wind turbine performance simulation and analysis device based on timing analysis, characterized in that: The wind turbine performance simulation analysis device based on timing analysis includes: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the wind turbine performance simulation analysis device based on timing analysis executes the wind turbine performance simulation analysis method based on timing analysis as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the wind turbine performance simulation analysis method based on timing analysis as described in any one of claims 1 to 7 is implemented.

Images