Wind power abnormality detection method and system based on diffusion model

Abstract

This invention provides a method and system for detecting wind power anomalies based on a diffusion model, belonging to the field of wind power fault diagnosis technology. The method includes: real-time acquisition of multi-dimensional time-series operational data of each wind turbine in a wind farm cluster to obtain a raw operational dataset; based on the raw operational dataset, data cleaning and time window alignment are performed through edge computing nodes to obtain a state feature sequence; the state feature sequence is synchronously transmitted to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades to acquire time-series data of tower structural strain, nacelle vibration spectrum characteristics, and blade acoustic signature signals, respectively, constructing a multi-physics coupling parameter set. This invention achieves comprehensive, dynamic, and refined identification of wind power anomalies.

Description

Technical Field

[0001] This invention relates to the field of wind power fault diagnosis technology, and in particular to a wind power anomaly detection method and system based on a diffusion model. Background Technology

[0002] Existing wind power anomaly detection technologies mostly rely on fixed threshold judgment rules or shallow machine learning fitting models, and rely on single electrical power time series data for analysis. The feature dimensions are thin, making it difficult to adapt to the multi-physics field coupling detection needs under complex mountain wind fields, low temperature freezing damage, turbulent disturbances and other conditions.

[0003] In actual operation and maintenance scenarios, when a wind turbine blade experiences low-temperature icing failure, leading to aerodynamic performance failure, traditional detection methods are difficult to integrate the acoustic fingerprint signal at the blade root and the strain time series data of the tower structure to carry out multi-source heterogeneous information linkage analysis. Furthermore, it is difficult to complete the spatial consistency correction of the feature sequence, which can easily lead to the real power generation anomaly being masked by environmental noise, resulting in missed fault detection. When early, slight mechanical wear of the gearbox induces latent anomalies such as gradual power decay, existing technologies generally neglect the mining of nacelle vibration spectrum characteristics. At the same time, they do not consider the inherent response delay and spatial distribution differences of multiple sensing units, making it difficult to construct a virtual hyperellipsoidal tensor field that characterizes the global structural response of the wind turbine. It is also difficult to extract structural dynamic weighting factors that are adapted to the operating conditions through surface curvature features. Ultimately, this results in the weak latent fault characteristics being submerged, making it difficult to achieve early warning.

[0004] Existing conventional detection schemes lack the ability to eliminate spatial observation data deviations caused by wind turbine structural deformation, and they also fail to introduce a diffusion model inverse denoising mechanism to achieve dynamic fitting and adaptive modeling of power probability distribution. The resulting theoretical power benchmark is rigid and inflexible, making it difficult to match real-time dynamic operating conditions. Ultimately, they fail to meet the core requirements of high-precision, full-condition, and finely classified real-time online anomaly detection and fault tracing for large-scale wind farm clusters. Summary of the Invention

[0005] This invention provides a method and system for detecting wind power anomalies based on a diffusion model, enabling comprehensive, dynamic, and refined identification of wind power anomalies.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a wind power anomaly detection method based on a diffusion model, the method comprising: Step 1: Collect multi-dimensional time-series operation data of each wind turbine in the wind farm cluster in real time to obtain the raw operation dataset; Based on the raw operation dataset, perform data cleaning and time window alignment processing through edge computing nodes to obtain the state feature sequence. Step 2: The state feature sequence is synchronously transmitted to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades to acquire the strain time series data of the tower structure, the vibration spectrum characteristics of the nacelle, and the acoustic signature signal of the blades, respectively, and to construct a multi-physics field coupling parameter set. Step 3: Based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set, a virtual hyperellipsoidal tensor field describing the structural response of the wind turbine is obtained; non-uniform rational B-spline fitting is performed on the virtual hyperellipsoidal tensor field to obtain the curvature extrema. Step 4: Divide the high strain gradient region and the low strain transmission region according to the curvature extreme points to obtain the division result; use the division result as the structural dynamic weighting factor to perform spatial consistency correction on the state feature sequence to obtain the corrected structural coupling state sequence. Step 5: The corrected structural coupling state sequence is used as a conditional input to the pre-trained diffusion model. The probability distribution of the wind turbine output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve. Step 6: Based on the comparison between the theoretical power distribution curve and the real-time grid-connected active power, calculate the degree of deviation of the actual power value from the theoretical power distribution to obtain the power anomaly residual sequence; based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, determine the anomaly type and obtain the detection result.

[0007] Furthermore, multi-dimensional time-series operational data of each wind turbine in the wind farm cluster are collected in real time to obtain the raw operational dataset. Based on the raw operational dataset, data cleaning and time window alignment are performed through edge computing nodes to obtain a state feature sequence, including: The original operational dataset is constructed by acquiring timestamp data of wind turbine impeller speed, nacelle temperature, gearbox vibration amplitude, real-time grid-connected active power, wind speed and direction, ambient temperature and humidity, and blade pitch angle at a preset sampling frequency. The missing values ​​in the original running dataset are filled in by linear interpolation through edge computing nodes. By removing high-frequency noise interference, the data is cleaned and the cleaned data in each dimension is obtained. Based on a unified time benchmark, the nearest neighbor resampling method is used to align the time windows of the cleaned data in each dimension, mapping the multi-dimensional time series data with different sampling frequencies to the same time grid to obtain the state feature sequence.

[0008] Furthermore, the state feature sequence is synchronously transmitted to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades, respectively acquiring the time-series data of the tower structure strain, the nacelle vibration spectrum characteristics, and the blade acoustic signature signals, constructing a multi-physics coupling parameter set, including: The state characteristic sequence is distributed to the fiber optic strain sensor at the bottom of the tower, the triaxial accelerometer at the top of the nacelle, and the acoustic emission sensor at the blade root. The tower deformation is monitored in real time using a fiber optic strain sensor at the bottom of the tower, and the strain time series data of the tower structure is obtained by demodulating the wavelength offset. Vibration signals were collected using a three-axis accelerometer on the top of the cabin, and frequency domain amplitude and phase information were extracted using fast Fourier transform to obtain the cabin vibration spectrum characteristics. The aerodynamic noise and structural impact signals are captured by an acoustic emission sensor at the blade root. The energy spectral density is extracted by wavelet packet decomposition to obtain the blade acoustic signature signal. The time-series data of tower structure strain, nacelle vibration spectrum characteristics, and blade acoustic signature signals are time-stamped and vector-stitched to construct a multi-physics field coupled parameter set.

[0009] Furthermore, based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set, a virtual hyperellipsoidal tensor field describing the structural response of the wind turbine is obtained; non-uniform rational B-spline fitting is performed on the virtual hyperellipsoidal tensor field to obtain the curvature extrema, including: The initial tensor matrix is ​​mapped to a three-dimensional physical coordinate system with the bottom of the tower as the origin. The lengths of the major axis, minor axis and semi-major axis of the virtual hyperellipsoid tensor field are determined based on the eigenvalues ​​of the initial tensor matrix, thus obtaining the virtual hyperellipsoid tensor field describing the structural response of the wind turbine. Set the initial set of control vertices and node vectors for the non-uniform rational B-spline surface, calculate the sum of squared Euclidean distances from the actual data points of the sensor to the surface of the virtual hyperellipsoid tensor field, and adjust the positions of the control vertices and the weights of the node vectors through an iterative optimization algorithm until the sum of squared Euclidean distances converges to a preset error threshold, thus completing the surface reconstruction. Based on the reconstructed surface, the coefficients of the first and second basic forms at discrete grid points on the surface are calculated, and the Gaussian curvature values ​​of each discrete grid point are derived. By iterating through the Gaussian curvature values ​​of all discrete grid points, the coordinates of the positions with the maximum and minimum curvature values ​​are selected using the neighborhood comparison method to obtain the curvature extrema points.

[0010] Furthermore, based on the curvature extrema, high-strain gradient regions and low-strain transfer regions are divided, resulting in a partitioning result. This partitioning result is used as a structural dynamic weighting factor to perform spatial consistency correction on the state characteristic sequence, yielding a corrected structural coupling state sequence, including: Using the curvature value of the curvature extremum point as the benchmark threshold, the surface mesh of the virtual hyperellipsoid tensor field is divided into a first set of surfaces with curvature values ​​higher than the benchmark threshold and a second set of surfaces with curvature values ​​lower than the benchmark threshold, which correspond to the high strain gradient region and the low strain transmission region, respectively, and the division result is obtained. Based on the partitioning results, the spatial volume enclosed by the first surface set and the second surface set in the three-dimensional physical coordinate system is calculated to obtain the region volume value. The total energy of the sensing signals in each region is counted, and the ratio of the total energy to the region volume value is calculated to obtain the energy distribution density of each region. The energy distribution density of each region is normalized to obtain the confidence coefficient corresponding to each region. The confidence coefficient is then weighted and summed with the preset regional importance weight to obtain the structural dynamic weight factor. Extract the feature vectors corresponding to the spatial positions of the tower bottom, nacelle top and blade root in the state feature sequence, and perform element-wise multiplication of the structural dynamic weight factor with the feature vectors to obtain the spatial deviation compensation value. By subtracting the spatial deviation compensation value from the eigenvector, spatial consistency correction is completed, resulting in the corrected structural coupling state sequence.

[0011] Furthermore, the corrected structural coupling state sequence is used as a conditional input to the pre-trained diffusion model. The probability distribution of the wind turbine's output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve, including: The corrected structural coupling state sequence is input into the multilayer perceptron layer of the conditional coding network to obtain the conditional embedding vector. By using conditional embedding vectors, a random noise matrix that follows a standard normal distribution is obtained. This random noise matrix is ​​then used as the initial input sample for the reverse process of the diffusion model, and the total time step for reverse denoising is set. In each inverse denoising time step, the current input sample, the encoding vector of the current time step, and the conditional embedding vector are concatenated and input into the noise prediction network to obtain the predicted noise component of the current step. Based on the preset variance scheduling coefficient, a subtraction operation is performed on the current input sample to remove the predicted noise component, and a new random noise term is added to obtain the updated input sample, until all time steps are completed. Repeat the independent reverse denoising iteration process multiple times, collect all the final output samples to form a power distribution sample set, and calculate the mean and standard deviation of the power distribution sample set at each time point; The expected power trajectory is constructed based on the mean, and the upper and lower confidence boundaries are calculated based on the standard deviation and confidence coefficient. The combination yields the theoretical power distribution curve that characterizes the expected output power of the wind turbine and the confidence interval under the current operating conditions.

[0012] Furthermore, based on a comparison between the theoretical power distribution curve and the real-time grid-connected active power, the deviation of the actual power value from the theoretical power distribution is calculated, resulting in a power anomaly residual sequence. Based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, the anomaly type is determined, yielding the detection results, including: The expected power trajectory and upper and lower confidence boundaries are extracted from the theoretical power distribution curve. The real-time grid-connected active power is subtracted from the expected power trajectory point by point to obtain the power anomaly residual sequence. Set a fixed-length sliding time window, perform sliding traversal on the power anomaly residual sequence, and calculate the linear fitting slope, variance energy value and maximum absolute deviation value of the residual data in each window; If the slope of the linear fit is consistently below the negative threshold and the variance energy value is below the fluctuation threshold, it is determined to be a latent anomaly of power asymptotic decay, and the determination result is obtained. If the maximum absolute deviation value exceeds the deviation threshold but the duration is less than the time threshold, and the variance energy value exhibits high-frequency abrupt change characteristics, it is determined to be a false anomaly of power data jump, so as to obtain the determination result; If the maximum absolute deviation value continues to exceed the upper and lower confidence boundaries and the duration is longer than the time threshold, and the power deviation from the theoretical wind speed range is determined in conjunction with wind speed data, it is judged as a real power generation anomaly of the power deviation type, so as to obtain the judgment result. By combining the judgment results from each window, a detection result is obtained that includes the anomaly type, occurrence time, and confidence level.

[0013] Secondly, a wind power anomaly detection system based on a diffusion model includes: The acquisition module is used to collect multi-dimensional time-series operation data of each wind turbine in the wind farm cluster in real time to obtain the raw operation dataset. Based on the raw operation dataset, data cleaning and time window alignment are performed by edge computing nodes to obtain the state feature sequence. The module is used to synchronously transmit the state feature sequence to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades, respectively to acquire the strain time series data of the tower structure, the vibration spectrum characteristics of the nacelle, and the acoustic signature signal of the blades, and to construct a multi-physics field coupling parameter set. The calculation module is used to obtain a virtual hyperellipsoidal tensor field describing the structural response of the wind turbine unit based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set; and to perform non-uniform rational B-spline fitting on the virtual hyperellipsoidal tensor field to obtain the curvature extrema. The correction module is used to divide the high strain gradient region and the low strain transmission region according to the curvature extreme point to obtain the division result; the division result is used as the structural dynamic weight factor to perform spatial consistency correction on the state feature sequence to obtain the corrected structural coupling state sequence. The fitting module is used to input the corrected structural coupling state sequence as a condition into the pre-trained diffusion model. The probability distribution of the wind turbine output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve. The processing module is used to compare the theoretical power distribution curve with the real-time grid-connected active power, calculate the degree of deviation of the actual power value from the theoretical power distribution, and obtain the power anomaly residual sequence; based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, the anomaly type is determined and the detection result is obtained.

[0014] The above-described solution of the present invention has at least the following beneficial effects: By collecting multi-dimensional time-series operational data from each wind turbine in a wind farm cluster in real time and performing data cleaning and time window alignment via edge computing nodes, a coupled parameter set is constructed by simultaneously fusing multi-physics field data from three heterogeneous sensing units: the bottom of the tower, the top of the nacelle, and the root of the blades. A virtual hyperellipsoidal tensor field is then constructed by combining the response delay and spatial distribution of the sensing units, and curvature extrema are extracted through non-uniform rational B-spline fitting. The partitioning results are used as dynamic weighting factors to correct the state feature sequence. The corrected feature sequence is then used as conditional input to pre-train a diffusion model, which undergoes inverse denoising and dynamic fitting of the power probability distribution. The technology combines the time-varying characteristics of residual sequences with amplitude morphology to determine anomaly types, thus overcoming the technical problems of existing technologies that rely on a single data dimension, cannot integrate multi-physics coupling characteristics, are difficult to eliminate spatial observation biases, have rigid and inflexible theoretical power benchmarks, and suffer from missed detection of hidden faults and difficulty in accurately distinguishing anomaly types. This achieves the technical effect of realizing full-domain, dynamic, and refined identification of wind power anomalies, reducing the false positive and false negative rates, meeting the high-precision, full-condition real-time online anomaly detection and fault tracing needs of large wind farm clusters, and improving the health management and operation and maintenance efficiency of wind turbine units. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of a wind power anomaly detection method based on a diffusion model provided in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of a wind power anomaly detection system based on a diffusion model provided in an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art.

[0018] like Figure 1 As shown, embodiments of the present invention propose a wind power anomaly detection method based on a diffusion model, the method comprising the following steps: Step 1: Collect multi-dimensional time-series operation data of each wind turbine in the wind farm cluster in real time to obtain the raw operation dataset; Based on the raw operation dataset, perform data cleaning and time window alignment processing through edge computing nodes to obtain the state feature sequence. Step 2: The state feature sequence is synchronously transmitted to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades to acquire the strain time series data of the tower structure, the vibration spectrum characteristics of the nacelle, and the acoustic signature signal of the blades, respectively, and to construct a multi-physics field coupling parameter set. Step 3: Based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set, a virtual hyperellipsoidal tensor field describing the structural response of the wind turbine is obtained; non-uniform rational B-spline fitting is performed on the virtual hyperellipsoidal tensor field to obtain the curvature extrema. Step 4: Divide the high strain gradient region and the low strain transmission region according to the curvature extreme points to obtain the division result; use the division result as the structural dynamic weighting factor to perform spatial consistency correction on the state feature sequence to obtain the corrected structural coupling state sequence. Step 5: The corrected structural coupling state sequence is used as a conditional input to the pre-trained diffusion model. The probability distribution of the wind turbine output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve. Step 6: Based on the comparison between the theoretical power distribution curve and the real-time grid-connected active power, calculate the degree of deviation of the actual power value from the theoretical power distribution to obtain the power anomaly residual sequence; based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, determine the anomaly type and obtain the detection result.

[0019] In this embodiment of the invention, multi-dimensional time-series operation data of each wind turbine in a wind farm cluster are collected in real time. Data cleaning and time window alignment are performed using edge computing nodes to obtain a state feature sequence. This state feature sequence is simultaneously transmitted to three heterogeneous sensing units at the bottom of the tower, the top of the nacelle, and the root of the blades to obtain multi-physics data and construct a multi-physics coupling parameter set. A virtual hyperellipsoidal tensor field is constructed by combining the response delay characteristics and spatial distribution of each sensing unit, and curvature extrema are obtained through non-uniform rational B-splines. Regions are divided based on these curvature extrema and used as structural dynamic weighting factors to correct the spatial consistency of the state feature sequence. The corrected structural coupling state sequence is then used as a conditional input to a pre-trained diffusion model, and a reverse denoising process is performed. The method of dynamically fitting the probability distribution of wind turbine output power to obtain the theoretical power distribution curve, and finally comparing it with the real-time grid-connected active power to calculate the residual sequence and determine the anomaly type based on its characteristics, overcomes the technical problems of existing wind power anomaly detection technologies, such as reliance on a single data dimension, inability to integrate multi-physics coupling features, difficulty in eliminating spatial observation bias, rigid and inflexible theoretical power benchmarks, easy to miss hidden faults, and inability to accurately distinguish anomaly types. It achieves full-domain, dynamic, and refined identification of wind power anomalies, effectively reduces the false judgment and missed detection rates, adapts to the multi-condition requirements of complex wind farms, meets the high-precision real-time online anomaly detection and fault tracing requirements of large wind farm clusters, and improves the technical effect of wind turbine health management and operation and maintenance efficiency.

[0020] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1 involves acquiring timestamp data on wind turbine rotor speed, nacelle temperature, gearbox vibration amplitude, real-time grid-connected active power, wind speed and direction, ambient temperature and humidity, and blade pitch angle at a preset sampling frequency to form the original operational dataset. Specifically, this includes: conducting full-area data collection for all wind turbine generators within the wind farm cluster, pre-setting a fixed and uniform data collection interval as the standard sampling basis, and continuously acquiring data according to this preset sampling frequency; sequentially collecting real-time rotor speed data, real-time ambient temperature data inside the nacelle cavity, vibration amplitude data generated during gearbox operation, and real-time active power transmitted to the grid by the generator unit under the operating conditions of each wind turbine generator unit. The system collects power numerical data, real-time airflow velocity and wind direction data at the wind farm site, real-time air temperature and humidity data of the external natural environment of the turbine, and real-time adjustment angle data of the wind turbine blades. Simultaneously, it binds and records the precise acquisition time information corresponding to each data point, forming a single time-series monitoring data point with complete time traceability information. It then aggregates and collects all dimensions of time-series monitoring data generated at different acquisition times from a single wind turbine, and integrates and collects the monitoring data from all wind turbines within the wind farm cluster. The data is arranged systematically according to turbine number and acquisition time sequence, and finally, a complete original operating dataset covering all turbines and all operating conditions is formed.

[0021] Step 1.2 involves using edge computing nodes to fill in missing values ​​in the original operational dataset using linear interpolation. This removes high-frequency noise interference, completing data cleaning and yielding cleaned data for each dimension. Specifically, this includes: retrieving the complete original operational dataset from edge computing nodes deployed at the wind farm site; performing standardized data cleaning in stages; and conducting missing data completion, using linear interpolation throughout. This involves locating all missing data points in the original operational dataset, extracting the corresponding monitoring values ​​for the same dimension from two consecutive valid acquisition times before and after each missing data point, calculating the average change in that dimension per unit time, and calculating the time span between the missing data point and the previous valid acquisition time. This completes the filling of all missing data. High-frequency noise filtering is performed using a sliding window filtering process. A statistical window with a fixed data capacity is set, and the window slides sequentially along the entire time series data at fixed steps. Within each sliding window, all values ​​of the original monitoring data within the window are accumulated to calculate the average value of the window data. This average value is then used to replace the original monitoring data corresponding to the center position of the window. By continuously sliding and updating window by window, environmental interference and high-frequency noise mixed in the data are removed, eliminating data distortion caused by random abnormal fluctuations. After completing the two core processing steps of missing value imputation and high-frequency noise filtering, all optimized and corrected monitoring data of each dimension are integrated and arranged in a unified manner, finally obtaining cleaned and complete operational data of each dimension without data gaps or noise interference.

[0022] Step 1.3, based on a unified time benchmark, uses the nearest neighbor resampling method to align the time windows of the cleaned data across all dimensions, mapping multi-dimensional time-series operational data with different sampling frequencies to the same time grid to obtain a state feature sequence. Specifically, this includes: establishing a unified standard time measurement benchmark for the entire wind farm; formulating fixed, unique, and uniformly accurate basic time grid partitioning rules; determining the core reference standard for the entire time-series alignment; for each dimension of operational data after cleaning and optimization, using the nearest neighbor resampling matching method to perform cross-frequency time alignment correction based on the original data acquisition time nodes; and mapping all multi-dimensional time-series operational data originally generated at different acquisition frequencies into a pre-defined unified standard time grid framework. In this process, for each reference time node within the standard time grid, the original valid monitoring data with the shortest time interval from that reference node is precisely selected, and the selected valid monitoring data is directly bound to the corresponding current reference time node. Strictly following this matching logic, the time node adaptation and conversion of all differentiated sampling frequency monitoring data is completed node by node and dimension by dimension, ensuring that all multi-dimensional time series operation data are accurately and neatly distributed on the same set of standard time grids, achieving complete synchronization and unification of all monitoring data in the time dimension. All the neat monitoring data that have completed time window alignment and correction are systematically spliced ​​and integrated according to the chronological order and the corresponding relationship with the unit, finally generating a standardized state feature sequence with standardized format, synchronized time sequence, and unified dimensions.

[0023] In this embodiment of the invention, a raw operational dataset is constructed by collecting multi-dimensional timestamped data such as wind turbine impeller speed and nacelle temperature at a preset sampling frequency. The original data is cleaned by using linear interpolation to fill missing values ​​and sliding window filtering algorithm to remove high-frequency noise through edge computing nodes. Then, based on a unified time reference, the nearest neighbor resampling method is used to align the time windows of the data in each dimension and map data with different sampling frequencies to the same time grid. Therefore, the invention overcomes the problems of missing values ​​and high-frequency noise interference in the original operational data, as well as the inconsistency of sampling frequencies and time synchronization of data in each dimension, which prevent the formation of standardized and high-quality feature sequences. This results in obtaining a clean, synchronized, and standardized state feature sequence.

[0024] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Distribute the status feature sequence to the fiber Bragg grating strain sensors at the bottom of the tower barrel, the triaxial acceleration sensors at the top of the nacelle, and the acoustic emission sensors at the root of the blade. Specifically, after the generation of the standardized status feature sequence is completed, the system starts the global data synchronization and distribution process, and synchronously pushes the integrated complete status feature sequence to three types of dedicated sensing devices deployed and installed in advance according to the unified transmission instruction. The first type of device is the fiber Bragg grating strain sensor fixedly installed at the bottom of the tower barrel of the wind turbine generator set, the second type of device is the triaxial acceleration sensor fixedly installed at the core monitoring point at the top of the nacelle, and the third type of device is the acoustic emission sensor fixedly attached to the key stress position at the root of the wind turbine blade. The entire distribution process maintains complete synchronization of the transmission timing, ensuring that the three types of sensing units start the linkage acquisition operation based on the same status feature sequence, and realizing the coordinated triggering of the electrical characteristics and structural mechanics monitoring of the unit operation.

[0025] Step 2.2: Use the fiber Bragg grating strain sensors at the bottom of the tower barrel to monitor the deformation of the tower barrel in real time, and demodulate the wavelength offset to obtain the tower barrel structure strain time series data. Specifically, after the fiber Bragg grating strain sensors deployed at the bottom of the tower barrel receive the synchronous trigger instruction, they continuously and uninterruptedly collect the microscopic deformation signals generated by the tower barrel shell under the action of wind load, self-weight deformation, and environmental temperature difference; the grating period inside the fiber Bragg grating will change synchronously with the physical deformation of the tower barrel, thereby causing the reflected light wave wavelength in the incident light propagation process to shift and change; complete the strain value conversion in combination with the inherent calibration coefficient of the fiber Bragg grating strain sensing; complete the wavelength demodulation and strain conversion continuously at each fixed sampling interval, and arrange and integrate the strain values corresponding to each moment in chronological order, and completely generate continuous and uninterrupted tower barrel structure strain time series data, retaining the original time series change characteristics throughout the process without performing early noise reduction and compression processing to ensure the complete retention of the structural deformation detail information.

[0026] Step 2.3: Use the triaxial acceleration sensors at the top of the nacelle to collect vibration signals, and extract the frequency domain amplitude and phase information through fast Fourier transform to obtain the nacelle vibration spectrum characteristics. Specifically, after the triaxial acceleration sensors deployed at the top of the nacelle receive the synchronous acquisition instruction, they synchronously capture the full-dimensional vibration time domain simulation signals caused by gearbox transmission, main shaft rotation, and nacelle aerodynamic disturbance during the operation of the nacelle, complete the analog-to-digital conversion of the original vibration time domain signal, convert the continuous analog electrical signal into an operable discrete digital vibration sequence, perform fast Fourier transform operation on the discrete vibration sequence, and realize the complete conversion of the time domain vibration signal to the frequency domain characteristics. Based on the complex frequency domain results, calculate the amplitude parameter and phase parameter corresponding to each frequency component separately. The amplitude calculation formula: Phase calculation formula: , where represents the The complex frequency domain results corresponding to each frequency component For complex numbers The real part, For complex numbers The imaginary part of the spectral density is used to summarize and integrate the amplitude and phase information corresponding to all frequency components in the full frequency range in frequency order, forming a complete spectral characteristic of the cabin vibration that reflects the frequency distribution law of the cabin vibration.

[0027] Step 2.4: Acoustic emission sensors at the blade root capture aerodynamic noise and structural impact signals. Energy spectral density is extracted via wavelet packet decomposition to obtain the blade acoustic signature signal. Specifically, the acoustic emission sensors pre-embedded at the blade root continuously capture aerodynamic noise signals generated during blade operation, as well as acoustic pulse signals caused by minute cracks in the blade shell, internal structural impacts, and edge icing. The resulting raw continuous acoustic time-domain signal is pre-processed in segments, and then a multi-level wavelet packet decomposition algorithm is used to decompose and analyze the signal layer by layer, breaking down the raw acoustic signal into sub-band wavelet packets corresponding to different frequency bands. The reconstructed signal of a single sub-band wavelet packet is defined as... , This represents the hierarchical number corresponding to the wavelet packet decomposition. This represents a continuous-time variable. The energy spectral density is calculated segment by segment for the sub-band signal of each sub-band. The core calculation formula for the energy spectral density of a single sub-band is: In the formula, Representing the The energy value corresponding to each wavelet packet, and the integration interval. arrive It represents a fixed time interval for a single signal analysis; it summarizes the energy distribution results of all decomposed subbands, and completes continuous splicing according to time sequence to form a blade acoustic pattern signal that intuitively reflects the dynamic change law of blade acoustic energy, accurately highlighting the abrupt changes in acoustic energy caused by blade icing, local cracking, and aerodynamic anomalies.

[0028] Step 2.5 involves time-stamp matching and vector concatenation of the tower structure strain time-series data, nacelle vibration spectrum characteristics, and blade acoustic signature signals with the state feature sequences to construct a multi-physics coupling parameter set. Specifically, this includes: uniformly calibrating the underlying timestamps of the three types of sensing data—tower structure strain time-series data, nacelle vibration spectrum characteristics, and blade acoustic signature signals—using the time base of the previously generated state feature sequences as the sole reference standard; traversing all data sampling nodes, accurately binding and matching the tower strain value, nacelle spectrum amplitude and phase parameters, and blade sub-band energy spectral density values ​​corresponding to each unified moment with the corresponding operating parameter values ​​within the state feature sequence at that moment, ensuring a one-to-one correspondence between all physical feature data at the same moment; after completing the accurate time matching, constructing standardized feature data vectors. Following the complete timeline order, sequentially arranging and summarizing the fused feature vectors constructed at all moments, completing the vector concatenation and integration of all-dimensional data, ultimately constructing a complete multi-physics coupling parameter set covering mechanical structure deformation, vibration frequency, acoustic characteristics, and equipment operating parameters.

[0029] In this embodiment of the invention, the state feature sequence is distributed to three sets of heterogeneous sensors at the bottom of the tower, the top of the nacelle, and the root of the blades. The fiber optic strain sensor is used to demodulate the wavelength offset to obtain the time-series data of the tower structure strain, the triaxial accelerometer is used to extract the vibration spectrum features of the nacelle through fast Fourier transform, and the acoustic emission sensor is used to extract the acoustic pattern signal of the blades through wavelet packet decomposition. Then, the three types of physical signals are timestamped and vector-stitched with the state feature sequence to construct a multi-physics coupling parameter set. Therefore, this method overcomes the technical problems of existing technologies that rely only on single electrical operation data, lack relevant features of wind turbine structural mechanics, cannot achieve multi-physics data fusion, and are difficult to correlate structural faults with power anomalies. This method enriches the feature dimensions of anomaly detection and achieves a deep correlation between the electrical operation status and structural health status of the wind turbine.

[0030] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Map the initial tensor matrix to a three-dimensional physical coordinate system with the tower bottom as the origin. Determine the lengths of the major axis, minor axis, and semi-major axis of the virtual hyperellipsoidal tensor field based on the eigenvalues ​​of the initial tensor matrix to obtain the virtual hyperellipsoidal tensor field describing the structural response of the wind turbine. Specifically, this includes: extracting the initial tensor matrix corresponding to the global structural response based on the multiphysics coupling parameter set obtained in the previous step; defining a dedicated three-dimensional physical coordinate system for the entire wind turbine's physical space; strictly setting the center position of the tower bottom as the origin of the entire coordinate system; defining the vertical upward direction as the longitudinal coordinate axis; the horizontal extension of the tower's radial direction as the first transverse coordinate axis; and the horizontal extension of the impeller's rotational section as the second transverse coordinate axis; and mapping the strain data, vibration data, acoustic signature, and structural parameters contained within the initial tensor matrix to a complete system. The system is imported into a newly established three-dimensional physical coordinate system, where each set of parameters in the tensor matrix corresponds to the structural response characteristics of a fixed spatial point within the coordinate system. The three sets of core eigenvalues ​​are then sorted in order of magnitude: the largest eigenvalue corresponds to the overall major axis of the virtual hyperellipsoid tensor field, the middle eigenvalue corresponds to the first minor axis, and the smallest eigenvalue corresponds to the second semi-axis. Centered on the origin, a spatial ellipsoidal distribution is constructed based on the determined major, minor, and semi-axis lengths. Each spatial gradient within the ellipsoid corresponds to the structural deformation response intensity at different locations within the wind turbine, ultimately generating a complete virtual hyperellipsoidal tensor field that characterizes the stress deformation and vibration transmission patterns of the wind turbine's entire spatial structure.

[0031] Step 3.2: Set the initial set of control vertices and node vectors for the non-uniform rational B-spline surface. Calculate the sum of squared Euclidean distances from the actual sensor data points to the surface of the virtual hyperellipsoidal tensor field. Adjust the control vertex positions and node vector weights through an iterative optimization algorithm until the sum of squared Euclidean distances converges to a preset error threshold, completing the surface reconstruction. Specifically, this includes: for the constructed virtual hyperellipsoidal tensor field, pre-define the initial set of control vertices required for the formation of the non-uniform rational B-spline surface, and configure basic node vectors that adapt to the overall surface contour distribution. Control vertices are used to constrain the overall trend of the surface, and node vectors are used to divide the data interpolation intervals within the surface. Mark all actual physical data points collected and calculated by the three types of sensor units deployed on-site in the three-dimensional physical coordinate system. Calculate the Euclidean space distance between each actual sensor data point and the surface of the virtual hyperellipsoidal tensor field. Then, square all distance values ​​and sum them up. The overall calculation formula for the sum of squared Euclidean distances is set as follows: In the formula, The sum of the squared Euclidean distances between all sensor data points and the projected points on the ellipsoidal surface represents the total error, which is used to characterize the overall deviation of the surface fitting. This represents the total number of actual sensor data points involved in this fitting calculation; The sequence number representing the sensor data point ranges from one to the total number of all data points. , , Representing the first The first horizontal coordinate, the second horizontal coordinate, and the vertical coordinate of each actual sensing data point in the three-dimensional physical coordinate system; , , Representing the surface of the virtual hyperellipsoidal tensor field, respectively, and the first The three-dimensional coordinates of the corresponding points in the vertical projection are formed by the actual sensor data points.

[0032] A gradient descent iterative optimization algorithm is introduced to continuously adjust the spatial placement of the surface control vertices, while simultaneously correcting the corresponding weight ratio parameters within the node vectors. After each round of parameter adjustment, the sum of squared Euclidean distances is recalculated, and the total error obtained from the current calculation is continuously compared with the error threshold set in advance by the system. The parameter optimization and error verification process is continuously executed in a loop until the sum of squared Euclidean distances continuously decreases and stabilizes within the preset error threshold range. At this point, all iterative adjustment operations are stopped, the final control vertex positions and node vector weight parameters are locked, and the high-precision non-uniform rational B-spline surface reconstruction operation that conforms to the actual sensor data distribution characteristics is completed.

[0033] Step 3.3: Based on the reconstructed surface, calculate the coefficients of the first and second fundamental forms at discrete grid points on the surface, and derive the Gaussian curvature values ​​for each discrete grid point. Specifically, this includes: performing fine-grained meshing on the reconstructed surface after iterative optimization, uniformly dividing the complete and continuous surface into a densely and regularly arranged set of discrete grid nodes, ensuring that the grid nodes can cover the entire surface contour of the virtual hyperellipsoidal tensor field without leaving any computational blind spots; for each independent discrete grid node on the surface, sequentially solve for the coefficients of the first and second fundamental forms corresponding to the surface differential geometry. The first fundamental form characterizes the arc length and area variation characteristics of local points on the surface, and the second fundamental form characterizes the bending, concave, and convex deformation characteristics of local points on the surface; based on the two sets of fundamental coefficients, derive the average curvature and principal curvature parameters of the surface corresponding to a single grid node, and then combine the principal curvature values ​​to complete the unified solution of the Gaussian curvature. The core formula for Gaussian curvature calculation is: , in the formula, The Gaussian curvature value obtained from solving for the current location of a single discrete grid node directly reflects the overall intensity of the bending deformation of the surface at that point. The value of the first principal curvature obtained from solving in two mutually perpendicular orthogonal directions represents the current mesh node position of the surface, which is the maximum bending curvature of the surface at that point. The value of the second principal curvature obtained from solving in two mutually perpendicular orthogonal directions represents the current mesh node position of the surface, and is the minimum bending curvature of the surface at that point.

[0034] By traversing point by point, the Gaussian curvature values ​​of all discrete grid nodes on the reconstructed surface are calculated and recorded one by one, and a complete Gaussian curvature distribution data table covering the entire tensor field surface is established to accurately restore the differences in bending deformation strength in different regions of the surface.

[0035] Step 3.4 involves iterating through the Gaussian curvature values ​​of all discrete grid points and using a neighborhood comparison method to identify the coordinates of the points with the maximum and minimum curvature values, thus obtaining curvature extrema. Specifically, after calculating the Gaussian curvature values ​​for all discrete grid nodes, a global curvature extrema screening process is performed. The standard neighborhood comparison method is used to identify the points. For each individual discrete grid node, the Gaussian curvature values ​​of all grid nodes within its immediate vicinity are retrieved to construct a local curvature comparison interval for the current node. The curvature value of the current central node is compared one by one with the curvature values ​​of all neighboring nodes. If the Gaussian curvature value of the current node is significantly higher than that of all neighboring nodes, the node is marked as a curvature maximum point; if the Gaussian curvature value of the current node is significantly lower than that of all neighboring nodes, the node is marked as a curvature minimum point. The neighborhood loop comparison of all discrete mesh nodes on the reconstructed surface is continuously performed to completely screen out all curvature maxima and minima. Simultaneously, the spatial coordinate information of the three-dimensional physical coordinate system of each extreme point is accurately retained. Finally, all curvature extreme points reflecting the key positions of the deformation abrupt change of the virtual hyperellipsoid tensor field surface are obtained, providing a precise spatial reference for strain region division and dynamic weight factor calculation.

[0036] In this embodiment of the invention, the initial tensor matrix is ​​mapped to a three-dimensional physical coordinate system with the bottom of the tower as the origin. The axis parameters of the virtual hyperellipsoidal tensor field are determined based on the eigenvalues ​​of the initial tensor matrix to construct a virtual hyperellipsoidal tensor field characterizing the structural response of the wind turbine. An initial set of control vertices and node vectors of a non-uniform rational B-spline surface are set. The surface reconstruction is completed by iteratively optimizing and adjusting the weights of the control vertices and node vectors. Then, the basic form coefficients of the discrete grid points of the reconstructed surface are calculated, the Gaussian curvature values ​​are derived, and the curvature extrema are selected by the neighborhood comparison method. Therefore, this invention overcomes the technical problems of existing technologies that cannot construct a tensor field that accurately characterizes the global structural response of the wind turbine, cannot accurately extract structural curvature features through surface fitting, and cannot reflect the differences in the wind turbine structural strain distribution, thus affecting the accuracy of region division and weight correction. This invention achieves the accurate construction of a virtual hyperellipsoidal tensor field characterizing the global structural response of the wind turbine, realizes high-precision surface reconstruction, and accurately extracts curvature extrema reflecting the structural strain distribution.

[0037] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Using the curvature values ​​of the curvature extrema as a baseline threshold, the surface mesh of the virtual hyperellipsoidal tensor field is divided into a first set of surfaces with curvature values ​​higher than the baseline threshold and a second set of surfaces with curvature values ​​lower than the baseline threshold, corresponding to the high strain gradient region and the low strain transfer region, respectively. The division results are obtained by: extracting the Gaussian curvature values ​​corresponding to all curvature extrema points obtained in the previous steps; calculating the mean of all extrema point curvature values; setting the final mean as the unified baseline threshold used for global surface division; and using this baseline threshold as the judgment boundary to perform a global comparison and classification operation on all discrete surface meshes covering the surface layer of the virtual hyperellipsoidal tensor field. The actual Gaussian curvature values ​​corresponding to each surface mesh node in the tensor field are traversed. All mesh nodes with actual curvature values ​​higher than a preset benchmark threshold are integrated and aggregated to form a complete first surface set. The spatial region corresponding to this set is a high strain gradient region with drastic structural deformation and large stress transmission gradient. At the same time, all mesh nodes with actual curvature values ​​lower than the preset benchmark threshold are uniformly integrated and aggregated to form a complete second surface set. The spatial region corresponding to this set is a low strain transmission region with gentle structural deformation and uniform and stable stress transmission. The two characteristic regions are accurately segmented through global curvature comparison and classification, and a clear surface mesh division result is output.

[0038] Step 4.2: Based on the partitioning results, calculate the spatial volume enclosed by the first and second surface sets in the three-dimensional physical coordinate system to obtain the regional volume value. Also, calculate the total energy of the sensing signals within each region and the ratio of the total energy to the regional volume value to obtain the energy distribution density of each region. Specifically, this includes: based on the partitioned first and second surface sets, performing integral calculations on the three-dimensional space enclosed by each surface set within the three-dimensional physical coordinate system established with the tower bottom as the origin, to obtain the actual volume of the corresponding three-dimensional space and the exclusive regional volume value for each of the two regions. For the high strain gradient region and the low strain transmission region, calculate the total effective energy of the structural strain signal, vibration spectrum signal, and acoustic signature signal collected by all deployed sensing units within the region. Accumulate and summarize the total energy corresponding to the sensing signals within each independent region. Substitute the total energy corresponding to the high strain gradient region and the low strain transmission region with the regional volume value to complete the accurate calculation of the energy distribution density of the two regions, quantitatively characterizing the degree of dense distribution of sensing energy within different structural stress regions.

[0039] Step 4.3: Normalize the energy distribution density of each region to obtain the confidence coefficient corresponding to each region. Then, weight and sum the confidence coefficients with preset regional importance weights to obtain a structural dynamic weight factor. Specifically, this includes: performing global normalization on the two sets of energy distribution densities obtained from solving the high strain gradient region and the low strain transmission region; eliminating the numerical magnitude difference caused by volume and absolute energy; uniformly converting the energy characteristics of the two regions to a standard value range of zero to one to obtain their respective confidence coefficients; pre-setting fixed regional importance weights based on the criticality of the wind turbine structure; setting higher weights for high strain gradient regions corresponding to structural deformation fault identification, and setting basic weights for low strain transmission regions corresponding to conventional steady-state monitoring; and multiplying the confidence coefficient of each region sequentially with its corresponding preset regional importance weight, summing the results, and weighting and fusing to obtain a unified structural dynamic weight factor adapted to the real-time structural deformation state.

[0040] Step 4.4: Extract feature vectors corresponding to the spatial positions of the tower bottom, nacelle top, and blade root from the state feature sequence. Perform element-wise multiplication of the structural dynamic weighting factor with the feature vectors to obtain the spatial deviation compensation value. Specifically, this involves: from the complete state feature sequence generated after edge processing, accurately extracting exclusive feature vectors that match the three physical points according to their spatial correlation. Each set of feature vectors contains the temporal features of the corresponding operating parameters and basic state data. Use the structural dynamic weighting factor generated by the previous fusion as a global correction coefficient and perform element-wise multiplication with each set of extracted spatially corresponding feature vectors. Perform weighted calculation by matching the value of each parameter within the feature vector. The element-wise multiplication compensation calculation uses basic operational relationships. In the formula This represents the single spatial deviation compensation value obtained from a single calculation. Represents a unified structural dynamic weighting factor across the entire domain. The original feature parameter values ​​corresponding to each item within the feature vector are represented. All feature parameters are converted and solved according to the principle of bitwise operation. All individual compensation values ​​are summarized to form a complete set of overall spatial deviation compensation values ​​that are adapted to each spatial point, accurately reflecting the inherent deviation of the observation data caused by the difference in wind turbine structural deformation.

[0041] Step 4.5: Subtract the spatial deviation compensation value from the feature vector to complete the spatial consistency correction, obtaining the corrected structural coupling state sequence. Specifically, this includes: performing deviation elimination and correction operations on each of the original spatially corresponding feature vectors obtained from the three locations at the bottom of the tower, the top of the nacelle, and the root of the blades. The core correction operation relationship is set as follows: In the formula This represents the individual corrected feature value obtained after bias removal. This represents the original feature parameter values ​​that were originally retained within the feature vector. This represents the single spatial deviation compensation value obtained from the weighted calculation in the preceding steps; it performs subtraction correction operations on all characteristic parameters of all spatial points one by one to eliminate the observation offset error caused by the response delay due to the spatial distribution difference of the wind turbine structural deformation sensor; it re-integrates and splices all characteristic parameters that have completed single-point correction according to their original time sequence and spatial correlation, and unifies and integrates the structural operation characteristics, mechanical deformation characteristics, and acoustic vibration characteristics to finally generate a corrected structural coupling state sequence that eliminates spatial observation deviations and conforms to the actual unit structural linkage law.

[0042] In this embodiment of the invention, the curvature of the extreme curvature point is used as the benchmark threshold to divide the high and low strain gradient regions. The spatial volume and total energy of the sensing signal in each region are calculated and the energy distribution density is solved. After normalization, the confidence coefficient is obtained and then combined with preset weights to generate a structural dynamic weight factor. Then, the spatial deviation compensation value is obtained by matching the corresponding spatial feature vector. The spatial consistency correction of the feature sequence is completed by deducting the compensation value. Therefore, this invention overcomes the technical problems of uneven spatial distribution of multiple sensing units, spatial offset of feature sequence caused by differences in wind turbine structural strain, inability of fixed weights to adapt to real-time structural deformation, difficulty in eliminating errors in heterogeneous fusion data, and easy distortion and weakening of fault features. This invention achieves accurate feature correction by matching the actual structural stress distribution of the wind turbine, effectively eliminates spatial observation deviation interference, optimizes the accuracy of multi-physics feature fusion, and generates a high-fidelity structural coupling state sequence.

[0043] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1 involves inputting the corrected structural coupling state sequence into the multilayer perceptron layer of the conditional coding network to obtain a conditional embedding vector. Specifically, this includes: inputting the complete structural coupling state sequence obtained after spatial consistency correction into the constructed conditional coding network and then into the multilayer perceptron hierarchical structure configured within the network; the multilayer perceptron performs layer-by-layer feature compression fusion and high-dimensional semantic mapping processing on the tower deformation features, nacelle vibration features, blade acoustic signature features, and unit steady-state operation features contained within the structural coupling state sequence, according to preset layer-by-layer mapping rules; the multilayer perceptron uses built-in fully connected weight parameters and activation functions to perform nonlinear transformations on the original sequence features, eliminating redundant feature interference while strengthening multi-physics coupling and correlation features; through multi-layer progressive feature extraction and dimension regularization operations, the high-dimensional, long-span structural coupling state sequence is uniformly transformed into a fixed-dimensional feature vector representing real-time unit operating condition information across the entire domain. This feature vector is the conditional embedding vector required for the diffusion model operation, ensuring the complete retention of operating condition correlation features throughout the process.

[0044] Step 5.2: Obtain a random noise matrix that follows a standard normal distribution through the conditional embedding vector. Use the random noise matrix as the initial input sample for the reverse process of the diffusion model, and set the total time step for reverse denoising. Specifically, this includes: using the generated conditional embedding vector as the operating condition constraint benchmark, constructing a random noise matrix adapted to the power time series dimension and feature dimension according to the data generation rules of the standard normal distribution. The basic probability calculation relationship of the standard normal distribution is set as P(x)=2π1e−2x2, where P(x) represents the probability density value corresponding to the random variable. All noise data in the matrix generated based on this distribution law conform to the distribution characteristics of zero mean and fixed standard deviation of one. Use the constructed random noise matrix directly as the initial input sample for the reverse denoising operation of the diffusion model. At the same time, based on the wind power time series sampling interval and the rate of change of operating conditions, uniformly set the total number of iteration steps required for the reverse denoising process. This fixed number of steps is the total time step for reverse denoising, ensuring that the noise removal process gradually conforms to the dynamic change law of power.

[0045] Step 5.3: In each inverse denoising time step, the current input sample, the encoding vector of the current time step, and the conditional embedding vector are concatenated and input into the noise prediction network to obtain the predicted noise component of the current step. Specifically, this includes: after the inverse denoising process is started, standardization operations are performed for each independent inverse denoising time step to extract the exclusive time encoding vector corresponding to the current iteration step, while retrieving the conditional embedding vector that remains unchanged throughout the process and the input sample data participating in the calculation at the current step. The three types of feature data are concatenated and fused according to the feature dimension alignment rules to form integrated feature data containing real-time sample information, time step identification information, and global working condition constraint information. The concatenated integrated feature data is completely input into the pre-trained and converged noise prediction network. The noise prediction network, relying on the deep convolutional mapping and fully connected fitting capabilities, accurately calculates the interference components superimposed on the samples within the current inverse iteration step. This interference component is the predicted noise component corresponding to the current step, realizing the accurate location and quantification extraction of invalid noise features in each iteration.

[0046] Step 5.4: Based on the preset variance scheduling coefficient, perform a subtraction operation on the current input sample to remove the predicted noise component, and add a new random noise term to obtain the updated input sample. This process continues until all time steps are completed. Specifically, this includes: pre-configuring a fixed variance scheduling coefficient that fits the characteristics of wind power fluctuations. This coefficient is used to control the amplitude of noise removal and the intensity of the new random disturbance in each round, ensuring that the denoising process is stable and does not cause feature distortion; and performing basic denoising operations based on the predicted noise component and variance scheduling coefficient obtained from the current step. The core update calculation formula is as follows: In the formula To complete the denoising update at the current step, the next inverse time step corresponds to the new input sample data. This refers to the original input sample data that is currently being processed in the reverse denoising time step. A fixed positive variance scheduling coefficient is preset to control the attenuation magnitude of noise removal in a single operation. This represents the precise noise component predicted by the network at the current inverse denoising time step. To control the superposition intensity of newly added small-amplitude noise, a fixed supplementary disturbance scheduling coefficient is preset. To maintain the continuity of the sample distribution, an additional small amount of random noise is added. After completing the single-step sample update according to the calculation logic, the updated sample is brought into the next inverse denoising time step to continue the loop calculation. The complete process of noise prediction sample update is repeated until all the preset total inverse denoising time steps are completed, and a single complete inverse denoising deduction is completed.

[0047] Step 5.5 involves repeatedly executing multiple independent reverse denoising iterations to collect all final output samples, forming a power distribution sample set. The mean and standard deviation of the power distribution sample set at each time point are calculated. Specifically, since a single reverse denoising iteration contains random distribution fluctuation errors, multiple independent complete reverse denoising iterations are repeatedly executed according to unified calculation rules. Each iteration retains the final output power fitting sample data. All independently derived output samples are collected and integrated to form a complete power distribution sample set covering the range of random distribution fluctuations. For all sample values ​​at each corresponding time point within the power distribution sample set, mean and standard deviation calculations are performed sequentially. The formula for calculating the mean at a single time point is: In the formula The average output value of all independently projected power samples at the same time point represents the standard expected power at that moment. This represents the total number of independent reverse denoising simulations performed throughout the entire process. The sequence numbering for the independent deduction experiments. For the first The sub-independent inverse extrapolation yields the fitted power sample values ​​at the current fixed time node; the formula for calculating the standard deviation at a single time node is... In the formula The standard deviation of the power sample data at the current time point represents the reasonable discrete fluctuation range of power output under normal operating conditions. This represents the total number of independent reverse denoising simulations performed throughout the entire process. The sequence numbering for the independent deduction experiments. For the first The fitted power sample value output by the sub-inverse inference at the current fixed time point. The average output value is calculated for all power samples at the current time point; the mean and standard deviation are solved globally for each time point, and a complete time series mean data table and standard deviation data table are established.

[0048] Step 5.6: Construct the expected power trajectory based on the mean, and calculate the upper and lower confidence boundaries according to the standard deviation and confidence coefficient. Combine these to obtain the theoretical power distribution curve representing the expected output power of the wind turbine under the current operating conditions and the confidence interval. Specifically, this includes: using the power mean data calculated at all time points as the core basis, connecting the power mean points corresponding to each time point one by one in chronological order to form a continuous curve. This curve is the expected power trajectory. This trajectory accurately reflects the standard output level of the wind turbine under normal operating conditions and is the core benchmark for determining whether the power is abnormal. Combining the confidence standard commonly used in the wind power testing industry, a fixed confidence coefficient is pre-configured. The role of this coefficient is to define the reasonable fluctuation range of the normal power generation output of the wind turbine, ensuring the accuracy of the anomaly judgment and industry adaptability. For each time node, based on the calculated standard deviation for that node and a preset confidence coefficient, the upper and lower confidence boundaries for that time node are obtained through corresponding calculations. Specifically, the calculation logic for the upper confidence boundary is to add the power mean of that time node to the product of the confidence coefficient and the standard deviation of that node to obtain the maximum allowable power output of the wind turbine at the current time node. The calculation logic for the lower confidence boundary is to subtract the product of the confidence coefficient and the standard deviation of that node from the power mean of that time node to obtain the minimum allowable power output of the wind turbine at the current time node. After completing the calculation of the upper and lower confidence boundaries for all time nodes, the previously constructed power expectation trajectory is combined and integrated with the upper and lower confidence boundaries calculated for each time node in chronological order. Through this combination and integration, a complete theoretical power distribution curve is finally generated.

[0049] In this embodiment of the invention, the corrected structural coupling state sequence is used to generate a conditional embedding vector through a multilayer perceptron of a conditional coding network. A standard normally distributed random noise matrix is ​​constructed based on the conditional embedding vector as the initial sample for inverse denoising of the diffusion model, and the total time step is configured. At each inverse time step, the current sample, the time step coding vector, and the conditional embedding vector are concatenated and input into a noise prediction network to obtain the predicted noise component. Noise is removed by combining a preset variance scheduling coefficient, and new noise is added to complete the iterative update. Then, through multiple independent inverse denoising iterations, the samples are aggregated to calculate the mean and standard deviation. Finally, the expected power trajectory is constructed based on the mean and combined with the standard deviation. The technique of using confidence coefficients to define upper and lower confidence boundaries to generate theoretical power distribution curves overcomes the technical problems of traditional detection models, which can only output fixed power values ​​at single points, cannot construct dynamic probability distributions and adaptive confidence intervals, are difficult to combine with wind turbine structural coupling feature constraints for modeling, have rigid and fixed theoretical power benchmarks, and have poor ability to adapt to complex wind conditions and nonlinear fluctuations. This results in the inability to accurately quantify the degree of power deviation and the easy occurrence of misjudgments and omissions. As a result, it achieves dynamic fitting of the power probability distribution based on the real-time structural state of the wind turbine, generates theoretical power curves and confidence intervals that fit the changing laws of the entire operating condition, and accurately defines the normal power fluctuation range.

[0050] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1: Extract the expected power trajectory and upper and lower confidence boundaries from the theoretical power distribution curve. Perform point-by-point subtraction between the real-time grid-connected active power and the expected power trajectory to obtain the power anomaly residual sequence. Specifically, this includes: accurately extracting the core expected power trajectory and the upper and lower confidence boundaries from the previously generated theoretical power distribution curve, ensuring that these three elements maintain a completely consistent time series rhythm with the real-time grid-connected active power, with each time node precisely aligned. Synchronously retrieve the wind turbine grid-connected active power data collected by the wind farm real-time monitoring system. This data uses the same time scale and sampling interval as the expected power trajectory, truly and accurately reflecting the current actual output status of the wind turbine, avoiding comparison deviations caused by inconsistent time dimensions. Perform point-by-point calculation of the power anomaly residual. Specifically, subtract the actual monitored value of the real-time grid-connected active power at each time node from the value of the expected power trajectory at the corresponding time node. By performing this point-by-point subtraction operation, the expected power component of the wind turbine under normal operating conditions is effectively removed, clearly highlighting the deviation between the actual output and the theoretical expectation. This deviation is the power anomaly residual. After completing the residual calculation for all time nodes, the power anomaly residual values ​​corresponding to each time node are arranged and integrated one by one according to the chronological order to form a complete power anomaly residual sequence.

[0051] Step 6.2: Set a fixed-length sliding time window and perform a sliding traversal on the power anomaly residual sequence. Within each window, calculate the linear fitting slope, variance energy value, and maximum absolute deviation value of the residual data. Specifically, this includes: combining the fluctuation characteristics of wind power time series data, the typical duration of various anomalies, and practical experience in wind farm operation and maintenance, reasonably setting a fixed-length sliding time window; the window length setting needs to consider two aspects: on the one hand, it should capture subtle changes in early minor latent anomalies, and on the other hand, it should effectively filter out instantaneous false anomalies. It is usually set to a fixed duration covering multiple sampling periods, and the specific duration can be flexibly adjusted according to the actual operating conditions of different wind farms; after setting the sliding time window, start from the starting time node of the power anomaly residual sequence and slide it point by point. The entire residual sequence is traversed globally. Each time a sampling point is slid, all power anomaly residual data within the current window are locked. For each residual data point within the window, three core feature parameters are calculated: the linear fitting slope, the variance energy value, and the maximum absolute deviation value. The linear fitting slope is calculated as follows: using time within the current window as the independent variable and the residual value as the dependent variable, a linear regression is performed on all residual data within the window to obtain the slope of the fitted line. This slope reflects the overall trend of the residual data within the current window. If the slope is negative, it indicates that the residuals are generally decreasing, corresponding to the actual grid-connected active power of the wind turbine gradually falling below the theoretical expected power. If the slope is positive, it indicates that the residuals are generally increasing, corresponding to the actual grid-connected active power of the wind turbine gradually exceeding the theoretical expected power.

[0052] The fluctuation intensity of the residuals is quantified by calculating the variance of all residual data within the current window. The larger the variance energy value, the more drastic the residual fluctuation within the current window, and vice versa, accurately reflecting the stability of the residuals. All residual data within the current window are traversed, and the absolute value of each residual value is extracted. The largest absolute value is the maximum absolute deviation value within the current window, which directly reflects the maximum deviation between the actual power of the wind turbine and the theoretical expected power within the current window. After the calculation of the three core characteristic parameters of each window is completed, the corresponding parameter results are recorded in a timely manner. The sliding window continues to repeat the above calculation process until the sliding window has traversed the entire power anomaly residual sequence.

[0053] Step 6.3: If the linear fitting slope remains below the negative threshold and the variance energy value is below the fluctuation threshold, it is determined to be a latent anomaly of gradual power decay. The determination result includes: based on historical operation and maintenance data of the wind turbine, fault characteristics of key components such as gearboxes, and anomaly judgment standards in the wind power industry, two fixed thresholds are pre-set: a negative threshold and a fluctuation threshold. The negative threshold is set to a fixed negative value to determine the severity of the residual decline trend; the fluctuation threshold is set to a smaller fixed value to determine the stability of residual fluctuations. Both thresholds are set with the core objective of accurately capturing latent anomalies caused by early, slight mechanical wear. To ensure no subtle anomalies are overlooked, a dual-condition judgment is performed on the linear fitting slope and variance energy value calculated for each sliding window. The linear fitting slope is continuously lower than a preset negative threshold. If this condition is met, it indicates that the overall power anomaly residual is showing a continuous downward trend, corresponding to the actual grid-connected active power of the wind turbine gradually falling below the theoretical expected power, and the downward trend is stable. This is one of the core characteristics of gradual power decay. The variance energy value is also judged to be lower than a preset fluctuation threshold. If this condition is met, it indicates that the fluctuation amplitude of the power anomaly residual is small, the change is gradual, and there are no drastic fluctuations. This is highly consistent with the gradual power decay characteristics caused by early slight mechanical wear.

[0054] These types of anomalies are latent anomalies, characterized by slow changes and small fluctuations, making them easily overlooked by traditional detection methods. However, the aforementioned dual-condition judgment can effectively highlight these subtle anomaly characteristics. When a sliding window simultaneously meets both conditions, and multiple consecutive adjacent windows maintain the same judgment result, random deviations from a single window can be ruled out, and a latent anomaly of gradual power attenuation in the wind turbine during that time period can be determined. The start and end times of the corresponding window, as well as key parameters such as the linear fitting slope and variance energy value of the current window, are recorded simultaneously to form a preliminary judgment result for this type of anomaly, achieving early warning of latent anomalies.

[0055] Step 6.4: If the maximum absolute deviation value exceeds the deviation threshold but the duration is less than the time threshold, and the variance energy value exhibits high-frequency abrupt change characteristics, it is determined to be a false anomaly of power data jump, so as to obtain the judgment result. Specifically, it includes: combining the operation and maintenance experience of wind farm sensors and the characteristics of data transmission, three fixed thresholds are preset, namely the deviation threshold and the time threshold. At the same time, the judgment criteria for the high-frequency abrupt change characteristics of the variance energy value are clarified, such as the variance energy value changing more than a preset multiple between adjacent windows. The deviation threshold is set to a value slightly higher than the normal power fluctuation range to determine the degree of deviation of the residual amplitude. The time threshold is set to a relatively short fixed duration, usually 1-2 sampling periods, to determine the duration of the anomaly. Through these three thresholds and judgment criteria, it is possible to effectively distinguish between instantaneous data jumps and actual wind turbine faults. For each sliding window, the maximum absolute deviation value, variance energy value, and the corresponding duration of the anomaly are used to perform a triple condition judgment. The maximum absolute deviation value is judged to see if it exceeds the preset deviation threshold. If this condition is met, it means that the actual power of the wind turbine in the current window has deviated significantly from the theoretical expected power. The duration of this deviation is judged to see if it is lower than the preset time threshold. If this condition is met, it means that this deviation is instantaneous and has not formed a continuous and stable abnormal state. It is a false deviation caused by non-equipment fault factors such as instantaneous sensor faults and data transmission interference. The variance energy value of the current window is judged to see if it shows high-frequency abrupt change characteristics. If this condition is met, it means that the residual fluctuation is instantaneous and drastic, which is consistent with the characteristics of data jumps, rather than a stable deviation caused by a real wind turbine equipment fault.

[0056] When a sliding window simultaneously meets all three of the above conditions, the power anomaly occurring within that time period can be determined as a false anomaly of power data jump. This type of anomaly is not a power generation failure of the wind turbine equipment itself, but rather caused by interference during data acquisition or transmission. Therefore, it is unnecessary to trigger a fault alarm, avoiding unnecessary alarms that would increase the workload of maintenance personnel. Simultaneously, the occurrence time, duration, and corresponding maximum absolute deviation value, variance energy value, and other key parameters of this false anomaly are recorded to form a preliminary judgment result for this type of anomaly.

[0057] Step 6.5: If the maximum absolute deviation value continuously exceeds the upper and lower confidence boundaries and the duration is longer than the time threshold, and the power deviation from the theoretical wind speed range is determined in conjunction with wind speed data, it is judged as a real power generation anomaly of the power deviation type, to obtain the judgment result. Specifically, this includes: using the preset deviation threshold and time threshold to ensure the consistency of the judgment standard; simultaneously, retrieving the wind speed data monitored in real time at the wind farm, and combining it with the characteristics of the wind turbine power curve to calculate the theoretical power range corresponding to the current wind speed, that is, the power range that the wind turbine should be in for normal power generation at the current wind speed; for the maximum absolute deviation value calculated for each sliding window, and the corresponding anomaly duration and real-time wind speed data, a triple condition judgment is performed; the maximum absolute deviation value is determined. If the large absolute deviation value continuously exceeds the upper and lower confidence boundaries of the theoretical power distribution curve, it indicates that the deviation between the actual power and the theoretical expected power has exceeded the normal fluctuation range, suggesting an anomaly. Next, it is determined whether the duration of this deviation exceeds a preset time threshold. If this condition is met, it indicates that the deviation is stable and continuous, not an instantaneous jump, exhibiting fault characteristics and ruling out the possibility of instantaneous interference. Finally, combined with real-time wind speed data, it is determined whether the actual grid-connected active power of the wind turbine deviates from the theoretical power range at the corresponding wind speed. This condition is used to exclude normal power fluctuations caused by sudden wind speed changes, ensuring that the identified anomaly is a power generation fault of the wind turbine itself, rather than a normal output fluctuation caused by wind speed changes.

[0058] When a sliding window simultaneously meets all three conditions mentioned above, and multiple consecutive adjacent windows maintain the same judgment result, random deviations can be ruled out, and it can be determined that there is a genuine power generation anomaly of the wind turbine during that time period. This type of anomaly is caused by a fault in the wind turbine equipment itself, which will affect the normal power generation of the wind turbine. It is necessary to trigger a fault alarm in a timely manner and arrange for maintenance personnel to investigate and handle the situation. Simultaneously record the occurrence time, duration, maximum deviation, and corresponding key parameters such as wind speed data and theoretical power range of this genuine anomaly to form a preliminary judgment result for this type of anomaly.

[0059] Step 6.6: Integrate the judgment results from each window to obtain detection results including anomaly type, occurrence time, and confidence level. Specifically, this includes: after completing anomaly judgment for all sliding windows, performing full-domain integration and verification of the preliminary judgment results for all windows, focusing on eliminating occasional single-window anomaly judgment results to avoid misjudgments caused by instantaneous data deviations; simultaneously, retaining anomaly information where judgment results from multiple consecutive windows are consistent to ensure the reliability and accuracy of anomaly judgment; for each type of anomaly judged, including latent anomalies such as gradual power decay, false anomalies such as power data jumps, and real power generation anomalies such as power deviations, summarizing the corresponding anomaly information; and clarifying the anomaly. The specific category of an anomaly is determined, accurately identifying its occurrence and end times. The occurrence time is based on the start time of the consecutive anomaly window, and the end time is based on the end time of the consecutive anomaly window. Simultaneously, the confidence level for each anomaly type is calculated. The confidence level is calculated by dividing the number of consecutive windows identified as belonging to that anomaly type by the total number of all relevant windows corresponding to that anomaly, and then converting the result to a percentage. A higher confidence level indicates a higher accuracy in anomaly identification, providing priority reference for operations and maintenance personnel in handling anomalies. All integrated anomaly information, including anomaly type, occurrence time, end time, and confidence level, is then uniformly organized to form a complete detection result.

[0060] In this embodiment of the invention, the expected power trajectory and upper and lower confidence boundaries are extracted from the theoretical power distribution curve. The real-time grid-connected active power is subtracted from the expected trajectory point by point to generate a power anomaly residual sequence. A fixed-length sliding time window is set to traverse the residuals and calculate the linear fitting slope, variance energy value, and maximum absolute deviation value. Based on the slope threshold, deviation amplitude, duration, and variance characteristics, the anomalies are classified into gradual power decay hidden anomalies, data jump false anomalies, and power deviation from actual power generation anomalies. Simultaneously, wind speed data is used to complete the operational condition verification. Finally, the entire window information is integrated and output, including the anomaly type, occurrence time, and identification confidence level. This technology overcomes the limitations of traditional single-threshold comparison, which cannot distinguish between multiple types of power anomalies, is difficult to identify early, gradual, latent faults, and is prone to misjudging sensor data jumps as real power generation faults. It also lacks operational condition correlation verification and cannot accurately trace the source of detection results. This technology enables refined classification and discrimination of wind power anomalies, effectively filters false interference signals, accurately captures weak latent faults, improves the rigor of judgment by relying on wind speed linkage verification, outputs standardized and implementable detection conclusions, reduces the probability of fault misjudgment and missed detection, and provides reliable data support for precise operation and maintenance, fault tracing, and proactive maintenance of wind farms.

[0061] like Figure 2 As shown, embodiments of the present invention also provide a wind power anomaly detection system based on a diffusion model, comprising: The acquisition module is used to collect multi-dimensional time-series operation data of each wind turbine in the wind farm cluster in real time to obtain the raw operation dataset. Based on the raw operation dataset, data cleaning and time window alignment are performed by edge computing nodes to obtain the state feature sequence. The module is used to synchronously transmit the state feature sequence to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades, respectively to acquire the strain time series data of the tower structure, the vibration spectrum characteristics of the nacelle, and the acoustic signature signal of the blades, and to construct a multi-physics field coupling parameter set. The calculation module is used to obtain a virtual hyperellipsoidal tensor field describing the structural response of the wind turbine unit based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set; and to perform non-uniform rational B-spline fitting on the virtual hyperellipsoidal tensor field to obtain the curvature extrema. The correction module is used to divide the high strain gradient region and the low strain transmission region according to the curvature extreme point to obtain the division result; the division result is used as the structural dynamic weight factor to perform spatial consistency correction on the state feature sequence to obtain the corrected structural coupling state sequence. The fitting module is used to input the corrected structural coupling state sequence as a condition into the pre-trained diffusion model. The probability distribution of the wind turbine output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve. The processing module is used to compare the theoretical power distribution curve with the real-time grid-connected active power, calculate the degree of deviation of the actual power value from the theoretical power distribution, and obtain the power anomaly residual sequence; based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, the anomaly type is determined and the detection result is obtained.

[0062] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0063] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A wind power anomaly detection method based on a diffusion model, characterized in that, The method includes: Step 1: Collect multi-dimensional time-series operation data of each wind turbine in the wind farm cluster in real time to obtain the raw operation dataset; Based on the raw operation dataset, perform data cleaning and time window alignment processing through edge computing nodes to obtain the state feature sequence. Step 2: The state feature sequence is synchronously transmitted to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades to acquire the strain time series data of the tower structure, the vibration spectrum characteristics of the nacelle, and the acoustic signature signal of the blades, respectively, and to construct a multi-physics field coupling parameter set. Step 3: Based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set, a virtual hyperellipsoidal tensor field describing the structural response of the wind turbine is obtained; non-uniform rational B-spline fitting is performed on the virtual hyperellipsoidal tensor field to obtain the curvature extrema. Step 4: Divide the high strain gradient region and the low strain transmission region according to the curvature extreme points to obtain the division result; use the division result as the structural dynamic weighting factor to perform spatial consistency correction on the state feature sequence to obtain the corrected structural coupling state sequence. Step 5: The corrected structural coupling state sequence is used as a conditional input to the pre-trained diffusion model. The probability distribution of the wind turbine output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve. Step 6: Based on the comparison between the theoretical power distribution curve and the real-time grid-connected active power, calculate the degree of deviation of the actual power value from the theoretical power distribution to obtain the power anomaly residual sequence; based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, determine the anomaly type and obtain the detection result.

2. The wind power anomaly detection method based on a diffusion model according to claim 1, characterized in that, Step 1: Collect multi-dimensional time-series operation data of each wind turbine in the wind farm cluster in real time to obtain the raw operation dataset; Based on the original running dataset, data cleaning and time window alignment are performed through edge computing nodes to obtain a state feature sequence, including: The original operational dataset is constructed by acquiring timestamp data of wind turbine impeller speed, nacelle temperature, gearbox vibration amplitude, real-time grid-connected active power, wind speed and direction, ambient temperature and humidity, and blade pitch angle at a preset sampling frequency. The missing values ​​in the original running dataset are filled in by linear interpolation through edge computing nodes. By removing high-frequency noise interference, the data is cleaned and the cleaned data in each dimension is obtained. Based on a unified time benchmark, the nearest neighbor resampling method is used to align the time windows of the cleaned data in each dimension, mapping the multi-dimensional time series data with different sampling frequencies to the same time grid to obtain the state feature sequence.

3. The wind power anomaly detection method based on a diffusion model according to claim 2, characterized in that, Step 2: The state feature sequence is synchronously transmitted to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades to acquire the time-series data of the tower structure strain, the nacelle vibration spectrum characteristics, and the blade acoustic signature signals, respectively, and to construct a multi-physics coupling parameter set, including: The state characteristic sequence is distributed to the fiber optic strain sensor at the bottom of the tower, the triaxial accelerometer at the top of the nacelle, and the acoustic emission sensor at the blade root. The tower deformation is monitored in real time using a fiber optic strain sensor at the bottom of the tower, and the strain time series data of the tower structure is obtained by demodulating the wavelength offset. Vibration signals were collected using a three-axis accelerometer on the top of the cabin, and frequency domain amplitude and phase information were extracted using fast Fourier transform to obtain the cabin vibration spectrum characteristics. The aerodynamic noise and structural impact signals are captured by an acoustic emission sensor at the blade root. The energy spectral density is extracted by wavelet packet decomposition to obtain the blade acoustic signature signal. The time-series data of tower structure strain, nacelle vibration spectrum characteristics, and blade acoustic signature signals are time-stamped and vector-stitched to construct a multi-physics field coupled parameter set.

4. The wind power anomaly detection method based on a diffusion model according to claim 3, characterized in that, Step 3: Based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set, obtain the virtual hyperellipsoidal tensor field describing the structural response of the wind turbine. Non-uniform rational B-spline fitting is performed on the virtual hyperellipsoidal tensor field to obtain the curvature extrema, including: The initial tensor matrix is ​​mapped to a three-dimensional physical coordinate system with the bottom of the tower as the origin. The lengths of the major axis, minor axis and semi-major axis of the virtual hyperellipsoid tensor field are determined based on the eigenvalues ​​of the initial tensor matrix, thus obtaining the virtual hyperellipsoid tensor field describing the structural response of the wind turbine. Set the initial set of control vertices and node vectors for the non-uniform rational B-spline surface, calculate the sum of squared Euclidean distances from the actual data points of the sensor to the surface of the virtual hyperellipsoid tensor field, and adjust the positions of the control vertices and the weights of the node vectors through an iterative optimization algorithm until the sum of squared Euclidean distances converges to a preset error threshold, thus completing the surface reconstruction. Based on the reconstructed surface, the coefficients of the first and second basic forms at discrete grid points on the surface are calculated, and the Gaussian curvature values ​​of each discrete grid point are derived. By iterating through the Gaussian curvature values ​​of all discrete grid points, the coordinates of the positions with the maximum and minimum curvature values ​​are selected using the neighborhood comparison method to obtain the curvature extrema points.

5. The wind power anomaly detection method based on a diffusion model according to claim 4, characterized in that, Step 4: Divide the high strain gradient region and the low strain transmission region according to the curvature extremum point to obtain the division result; Using the partitioning results as structural dynamic weighting factors, spatial consistency correction is applied to the state feature sequence to obtain the corrected structural coupling state sequence, including: Using the curvature value of the curvature extremum point as the benchmark threshold, the surface mesh of the virtual hyperellipsoid tensor field is divided into a first set of surfaces with curvature values ​​higher than the benchmark threshold and a second set of surfaces with curvature values ​​lower than the benchmark threshold, which correspond to the high strain gradient region and the low strain transmission region, respectively, and the division result is obtained. Based on the partitioning results, the spatial volume enclosed by the first surface set and the second surface set in the three-dimensional physical coordinate system is calculated to obtain the region volume value. The total energy of the sensing signals in each region is counted, and the ratio of the total energy to the region volume value is calculated to obtain the energy distribution density of each region. The energy distribution density of each region is normalized to obtain the confidence coefficient corresponding to each region. The confidence coefficient is then weighted and summed with the preset regional importance weight to obtain the structural dynamic weight factor. Extract the feature vectors corresponding to the spatial positions of the tower bottom, nacelle top and blade root in the state feature sequence, and perform element-wise multiplication of the structural dynamic weight factor with the feature vectors to obtain the spatial deviation compensation value. By subtracting the spatial deviation compensation value from the eigenvector, spatial consistency correction is completed, resulting in the corrected structural coupling state sequence.

6. The wind power anomaly detection method based on the diffusion model according to claim 5, characterized in that, Step 5: The corrected structural coupling state sequence is used as a conditional input to the pre-trained diffusion model. The probability distribution of the wind turbine output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve, including: The corrected structural coupling state sequence is input into the multilayer perceptron layer of the conditional coding network to obtain the conditional embedding vector. By using conditional embedding vectors, a random noise matrix that follows a standard normal distribution is obtained. This random noise matrix is ​​then used as the initial input sample for the reverse process of the diffusion model, and the total time step for reverse denoising is set. In each inverse denoising time step, the current input sample, the encoding vector of the current time step, and the conditional embedding vector are concatenated and input into the noise prediction network to obtain the predicted noise component of the current step. Based on the preset variance scheduling coefficient, a subtraction operation is performed on the current input sample to remove the predicted noise component, and a new random noise term is added to obtain the updated input sample, until all time steps are completed. Repeat the independent reverse denoising iteration process multiple times, collect all the final output samples to form a power distribution sample set, and calculate the mean and standard deviation of the power distribution sample set at each time point; The expected power trajectory is constructed based on the mean, and the upper and lower confidence boundaries are calculated based on the standard deviation and confidence coefficient. The combination yields the theoretical power distribution curve that characterizes the expected output power of the wind turbine and the confidence interval under the current operating conditions.

7. The wind power anomaly detection method based on a diffusion model according to claim 6, characterized in that, Step 6: Based on the comparison between the theoretical power distribution curve and the real-time grid-connected active power, calculate the degree of deviation of the actual power value from the theoretical power distribution to obtain the power anomaly residual sequence; Based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, the anomaly type is determined, and the detection results are obtained, including: The expected power trajectory and upper and lower confidence boundaries are extracted from the theoretical power distribution curve. The real-time grid-connected active power is subtracted from the expected power trajectory point by point to obtain the power anomaly residual sequence. Set a fixed-length sliding time window, perform sliding traversal on the power anomaly residual sequence, and calculate the linear fitting slope, variance energy value and maximum absolute deviation value of the residual data in each window; If the slope of the linear fit is consistently below the negative threshold and the variance energy value is below the fluctuation threshold, it is determined to be a latent anomaly of power asymptotic decay, and the determination result is obtained. If the maximum absolute deviation value exceeds the deviation threshold but the duration is less than the time threshold, and the variance energy value exhibits high-frequency abrupt change characteristics, it is determined to be a false anomaly of power data jump, so as to obtain the determination result; If the maximum absolute deviation value continues to exceed the upper and lower confidence boundaries and the duration is longer than the time threshold, and the power deviation from the theoretical wind speed range is determined in conjunction with wind speed data, it is judged as a real power generation anomaly of the power deviation type, so as to obtain the judgment result. By combining the judgment results from each window, a detection result is obtained that includes the anomaly type, occurrence time, and confidence level.

8. A wind power anomaly detection system based on a diffusion model, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to collect multi-dimensional time-series operating data of each wind turbine in the wind farm cluster in real time to obtain the raw operating dataset. Based on the original running dataset, data cleaning and time window alignment are performed through edge computing nodes to obtain a state feature sequence. The module is used to synchronously transmit the state feature sequence to three sets of heterogeneous sensing units deployed at the bottom of the tower, the top of the nacelle, and the root of the blades, respectively to acquire the strain time series data of the tower structure, the vibration spectrum characteristics of the nacelle, and the acoustic signature signal of the blades, and to construct a multi-physics field coupling parameter set. The calculation module is used to obtain a virtual hyperellipsoidal tensor field describing the structural response of the wind turbine unit based on the response delay characteristics and spatial distribution of each sensing unit in the multiphysics coupling parameter set; and to perform non-uniform rational B-spline fitting on the virtual hyperellipsoidal tensor field to obtain the curvature extrema. The correction module is used to divide the high strain gradient region and the low strain transmission region based on the curvature extremum point to obtain the division result; The partitioning result is used as a structural dynamic weighting factor to perform spatial consistency correction on the state feature sequence, resulting in the corrected structural coupling state sequence. The fitting module is used to input the corrected structural coupling state sequence as a condition into the pre-trained diffusion model. The probability distribution of the wind turbine output power is dynamically fitted through the inverse denoising process of the diffusion model to obtain the theoretical power distribution curve. The processing module is used to compare the theoretical power distribution curve with the real-time grid-connected active power, calculate the degree of deviation of the actual power value from the theoretical power distribution, and obtain the power anomaly residual sequence; based on the time-varying characteristics and amplitude shape of the power anomaly residual sequence, the anomaly type is determined and the detection result is obtained.

Images