Fan operation abnormal vibration monitoring method and system based on multi-sensor fusion

Abstract

The invention discloses a fan operation abnormal vibration monitoring method and system based on multi-sensor fusion, and relates to the technical field of fan operation monitoring, and the method comprises the following steps: deploying a plurality of types of sensors on a fan, collecting the data of each sensor in real time, carrying out the synchronous processing of the sensor data through employing an IEEE1588 protocol, carrying out the preprocessing through combining wavelet denoising, and carrying out the monitoring of the abnormal vibration of the fan. Extracting feature data by using a principal component PCA combined independent component ICA analysis method; according to the method, multiple types of sensors are deployed on the fan, the actual conditions of the offshore wind field are considered, fine processing and feature extraction are carried out after multi-source signals are collected in real time, the abnormal vibration condition of the fan in the coastal or offshore wind field under the complex environment is accurately monitored and reliably recognized, and therefore the early abnormal features of fan operation can be captured in time; normal changes and real fault anomalies caused by environmental factors are effectively distinguished, and early discovery of faults is further realized.

Description

Technical Field

[0001] This invention relates to the field of wind turbine operation monitoring technology, specifically to a method and system for monitoring abnormal vibrations in wind turbine operation based on multi-sensor fusion. Background Technology

[0002] Coastal or offshore wind turbines operate in extremely harsh environments, facing complex conditions such as high salt spray, high humidity, and strong winds and waves for extended periods. These environmental factors can affect the vibration characteristics of the wind turbines.

[0003] For example, on the one hand, the periodic changes of strong winds and tides at sea have a combined effect on wind turbines. Strong winds cause the turbine blades to vibrate, and the rise and fall of tides alter the hydrodynamic environment around the turbine foundation. The interaction between the two may lead to modulation of low-frequency vibrations in the turbine. This low-frequency modulation can mask early fault characteristics of the turbine, and traditional monitoring methods are unable to detect early signs of faults, resulting in delayed fault detection and increasing the risk of turbine damage. On the other hand, wind turbine blades are usually made of glass fiber / carbon fiber composite materials. These composite materials are hygroscopic. When operating in the high humidity environment of the coast for a long time, the blades will absorb moisture, leading to changes in the material's elastic modulus, increased density, and enhanced damping characteristics, resulting in a toughening effect. This, in turn, changes the blade's natural frequency and vibration response characteristics. At the same time, it also changes the propagation speed and attenuation performance of sound waves in the acoustic monitoring of the blades. Many changes may affect the system's ability to misdiagnose blade modal changes caused by humidity saturation as damage such as internal structural delamination or web cracking, which may cause false alarms. This makes it difficult for maintenance personnel to accurately determine whether the turbine has actually failed, increasing maintenance costs and workload.

[0004] Existing technologies have the following shortcomings: Firstly, existing technologies mostly use single-type sensors, lacking deep fusion of multi-physical field data, and cannot comprehensively acquire information on the wind turbine's operating status. Secondly, insufficient consideration of environmental factors leads to a high false alarm rate, causing significant challenges to operation and maintenance. Finally, the lack of adaptive learning capabilities prevents adaptation to dynamic changes in the marine environment, resulting in unstable monitoring performance under different environmental conditions. Therefore, there is an urgent need to propose a method and system for monitoring abnormal vibrations in wind turbines based on multi-sensor fusion. This system should accurately monitor and reliably identify abnormal vibrations of wind turbines in complex environments in coastal or offshore wind farms, effectively distinguishing between normal changes caused by environmental factors and genuine faults, thereby reducing false alarm rates, enabling early fault detection and accurate diagnosis, and improving the safety and reliability of wind turbine operation.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for monitoring abnormal vibrations in wind turbine operation based on multi-sensor fusion. This invention solves the problems in the background art by deploying multiple types of sensors on the wind turbine to extract feature data, establishing a scene feature template library for complex marine environments, using physical modeling combined with data-driven modeling to establish an accurate scene model, and comprehensively applying vibration analysis technology, acoustic detection technology, and machine learning diagnostic algorithms to diagnose faults based on quantified abnormal features.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring abnormal vibrations in wind turbine operation based on multi-sensor fusion, comprising the following steps:

[0008] S1. Deploy multiple types of sensors on the wind turbine to collect data from each sensor in real time. Use the IEEE1588 protocol to synchronize and process the sensor data. Combine wavelet denoising for preprocessing and then use principal component analysis (PCA) and independent component analysis (ICA) to extract feature data.

[0009] S2. Establish a scene feature template library for complex marine environments, compare the extracted feature data with the scene feature template library, use the cosine similarity algorithm to identify the scene where the wind turbine is located, define the feature threshold of the abnormal early warning mechanism, and trigger the abnormal early warning when the extracted feature data exceeds the feature threshold.

[0010] S3. Based on the scene identified by the extracted feature data, a precise scene model is established using physical modeling combined with data-driven modeling. The extracted feature data is input into the corresponding scene model to simulate normal operation. The abnormal feature indicators are compared with the actual data to quantify the abnormal features, refine the scene classification, set abnormal feature thresholds, and determine the abnormal type and severity based on the comparison and quantification of abnormal feature indicators.

[0011] S4. By comprehensively utilizing vibration analysis technology, acoustic detection technology, and machine learning diagnostic algorithms, fault diagnosis is performed on quantified abnormal features to generate diagnostic results that accurately locate the fault. A scene adaptive mechanism is developed to adaptively adjust diagnostic parameters according to scene characteristics and automatically trigger corresponding processing decisions.

[0012] S5. Based on the system feedback of diagnostic results and processing decisions, establish a dual-loop feedback mechanism to continuously optimize system performance. At the same time, build a knowledge base, analyze and update the knowledge base regularly, and optimize the system's monitoring and diagnostic capabilities.

[0013] Optionally, the data acquisition steps for each sensor are as follows:

[0014] For the various types of sensors deployed on the wind turbine, the sensor types and the corresponding number of sensor types are counted. Among them, the sensor types are divided into acceleration sensors, displacement sensors, acoustic sensors, anemometers, tidal sensors and temperature and humidity sensors.

[0015] The various types of sensors were statistically analyzed and assigned type and quantity numbers to identify each sensor deployed on the wind turbine.

[0016] Multiple types of sensors are activated to collect data over a continuous period of time, forming a time series of sensor data.

[0017] Data from various sensors is integrated into a multi-sensor dataset.

[0018] Optionally, the data synchronization processing and wavelet denoising preprocessing steps are as follows:

[0019] For each sensor, there is a time offset and network transmission delay between the local timestamp and the master clock timestamp;

[0020] The time offset and network transmission delay are calculated by using the interleaved signaling mechanism of the IEEE 1588PTP protocol on the data from each sensor.

[0021] Based on time offset and network transmission delay correction of time series data from each sensor, the timestamps of all sensor data are unified to a unified timestamp on a master clock to generate a time-aligned multi-sensor dataset.

[0022] Wavelet denoising preprocessing is performed on time-aligned multi-sensor datasets. Wavelet basis functions are selected and decomposed into wavelet coefficients of different scales through wavelet transform.

[0023] Apply a threshold function to the wavelet coefficients to remove wavelet coefficients with small amplitudes, and obtain the processed detail coefficients.

[0024] Wavelet reconstruction is performed using the processed detail coefficients to obtain the denoised multi-sensor dataset.

[0025] Optionally, the feature data extraction steps are as follows:

[0026] The multi-sensor dataset, after synchronization and denoising, is standardized and organized into an observation matrix;

[0027] Based on the observation matrix, the covariance matrix is ​​calculated using principal component analysis (PCA).

[0028] Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues;

[0029] Select the eigenvectors corresponding to the top N largest eigenvalues ​​as principal components, and project the standardized observation matrix onto the principal components to obtain the principal component score matrix.

[0030] The principal component score matrix is ​​input into the independent component ICA analysis method, and the FastICA algorithm is used to iteratively solve the separation matrix. The FastICA algorithm requires initializing the separation matrix first, and then iteratively updating the initial separation matrix until convergence, thus obtaining the independent component matrix.

[0031] The independent component matrix is ​​used as the feature vector to form the extracted feature data.

[0032] Optionally, the steps for establishing the scene feature template library are as follows:

[0033] Based on long-term historical data of wind turbines in a healthy state, and covering different scenarios such as combined strong wind and tidal effects, blade moisture absorption, and complex scenarios, multi-sensor data were obtained during the actual operation of wind turbines.

[0034] We still use the principal component analysis (PCA) combined with independent component analysis (ICA) to extract feature vectors from the data collected for each scenario.

[0035] The extracted feature vectors are clustered using the unsupervised clustering method K-Means to match feature patterns;

[0036] The feature vectors are divided into three clusters, and the centroid of each cluster constitutes a scene feature template, including the strong wind-tidal combined effect feature pattern, the leaf moisture absorption feature pattern, and the composite scene feature pattern.

[0037] For each scene feature template, calculate the Euclidean distance distribution from all feature vectors within the cluster to the centroid, and set a dynamic threshold to determine whether the data points of the new feature vector are abnormal with respect to the scene feature template.

[0038] Analyze the statistical distribution of all feature vector data points falling into each cluster with their corresponding scene feature templates, and mark the data points of abnormal scenes.

[0039] Optionally, the steps of the cosine similarity algorithm for scene recognition are as follows:

[0040] Calculate the cosine similarity between the data points of the new feature vector and each template in the scene feature template library;

[0041] Select the scene corresponding to the template with the highest similarity as the scene where the wind turbine is currently located;

[0042] After identifying the scene, calculate the Euclidean distance between the data points of the new feature vector and the centroid of the scene feature template;

[0043] The Euclidean distance is compared with a dynamic threshold. When the Euclidean distance exceeds the dynamic threshold, the scene is marked as abnormal and an abnormality warning is triggered.

[0044] Optionally, the steps for establishing the accurate scene model are as follows:

[0045] A coupled model of wind turbine and marine environment and a model of moisture absorption and mechanical performance change were established by using physical modeling combined with data-driven modeling.

[0046] Among them, for establishing a coupled wind turbine-ocean environment model:

[0047] Collect the dynamic behavior of the wind turbine under the combined action of strong wind and tide, including environmental information vectors and the state vector of wind turbine operation;

[0048] Based on environmental information vectors and state vectors, a force vector function generated by strong winds is constructed to calculate the aerodynamic load vector of the aerodynamic forces acting on the wind turbine. At the same time, a force vector function generated by tidal forces is constructed to calculate the hydrodynamic load vector of the forces of waves and ocean currents acting on the wind turbine foundation structure.

[0049] Based on aerodynamic load vectors and hydrodynamic load vectors, and combined with data-driven inertial forces, damping forces, and elastic forces during wind turbine operation, a coupled dynamic behavior model is obtained.

[0050] For the hygroscopic-mechanical property change model:

[0051] Collect data on the moisture concentration, ambient temperature, and humidity absorbed by the composite material of the wind turbine blades;

[0052] Based on the theoretical water diffusion coefficient of water diffusion dynamics, Fick's second law is used to construct the diffusion equation of water in the leaf from the high concentration region to the low concentration region inside the composite material, calculate the water concentration field, and define the boundary conditions according to the ambient humidity.

[0053] Based on the moisture concentration field, combined with the equivalent elastic modulus and damping coefficient of the blade material, a moisture absorption-mechanical property change model is established to obtain the mechanical property degradation relationship between mechanical properties and moisture absorption.

[0054] The established wind turbine-ocean environment coupling model and moisture absorption-mechanical performance change model will simulate the operating state of the wind turbine in the scenario and output the feature vector predicted by the scenario model.

[0055] Optionally, the steps for generating the diagnostic results are as follows:

[0056] Receive quantified abnormal feature indicators, combine them with extracted feature data, and list and statistically analyze the types of faults and normal states that may occur in the wind turbine operation status to build a status identification framework.

[0057] Vibration analysis technology, acoustic detection technology, and machine learning diagnostic algorithms are used as evidence sources (BPA) for fault diagnosis. Fault diagnosis probabilities are assigned to different state types in the state recognition framework, including generating vibration analysis evidence sources by analyzing the spectrum, envelope spectrum, and modal parameters of vibration signals.

[0058] By analyzing the amplitude, count, energy, and location information of acoustic emission signals, acoustic detection evidence sources are generated.

[0059] By inputting the feature data vector and abnormal feature indicators extracted in real time during wind turbine operation into the probability distribution output by a pre-trained multi-class XGBooster, a source of machine learning evidence is generated.

[0060] The decision-level fusion combination rule using DS evidence theory integrates vibration analysis evidence sources, acoustic detection evidence sources, and machine learning evidence sources to obtain a comprehensive quality function, and then formulates processing decisions based on the comprehensive quality function.

[0061] The diagnostic parameters for scene recognition are dynamically adjusted using a scene adaptation mechanism, the relationship between scene features and diagnostic parameters is mapped, and diagnostic results are generated.

[0062] Optionally, the quantification steps of the abnormal feature indicators are as follows:

[0063] Real-time online monitoring of coastal / offshore wind turbine operation is conducted. The actual measured synchronously denoised feature data is compared and analyzed with the predicted feature vector output after simulating the normal operation of the wind turbine by a precise scenario model. Abnormal feature indicators of low frequency modulation index (LMI), moisture absorption influence factor (MIF), and composite anomaly index (CAI) are quantitatively calculated.

[0064] For the low-frequency modulation index (LMI), the residual signal of the difference between the real-time monitored feature data and the predicted feature vector output by the accurate scene model is subjected to Hilbert transform to obtain the envelope. The envelope is then subjected to spectral analysis to obtain the spectrum of the actual monitored signal and the spectrum of the accurate scene model simulated signal. The energy affected by the tidal / wave characteristic frequency is found, and the low-frequency modulation index (LMI) is calculated.

[0065] For the moisture absorption impact factor (MIF), a weight matrix is ​​set based on prior knowledge of the humidity sensitivity of different scene characteristics, and the moisture absorption impact factor (MIF) is calculated based on the residual signal using Mahalanobis distance.

[0066] For the composite anomaly index CAI, the weighted average method is used to calculate the composite anomaly index CAI based on the low-frequency modulation index LMI and the moisture absorption effect factor MIF.

[0067] The wind turbine operation abnormal vibration monitoring system based on multi-sensor fusion includes a multi-source data acquisition module: multiple types of sensors, such as acceleration sensors, displacement sensors, acoustic sensors, anemometers, tidal sensors, and temperature and humidity sensors, are deployed on the nacelle, tower, and blades of the wind turbine. The system uses the IEEE1588 protocol to synchronize the microsecond-level time of all sensors and collects sensor data on vibration, acoustics, and environment in real time when the wind turbine is operating in coastal / offshore conditions.

[0068] Data preprocessing and feature extraction module: Receives sensor data and extracts feature data using wavelet denoising combined with principal component analysis and PCA-ICA algorithm;

[0069] Scene modeling and anomaly feature quantification module: Construct a scene feature template library, identify scenes by comparing feature data, establish scene models using physical modeling combined with data-driven modeling, quantify anomaly feature indicators, further refine scene classification, and set anomaly feature thresholds to assess anomaly type and severity;

[0070] Multi-technology integrated fault diagnosis module: It comprehensively utilizes vibration analysis technology, acoustic detection technology and machine learning diagnostic algorithms to diagnose faults based on abnormal characteristic indicators, develops a scene adaptive mechanism, adjusts diagnostic parameters according to scene characteristics, and triggers corresponding processing decisions.

[0071] Feedback optimization knowledge base management module: Establish a dual-loop feedback mechanism to feed back diagnostic results and processing decisions to various modules of the system. After building and updating the knowledge base, adjust sensor data acquisition strategies, feature data extraction methods, scene model construction, and anomaly threshold settings based on the feedback information.

[0072] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0073] This invention deploys multiple types of sensors on wind turbines, taking into account the actual conditions of offshore wind farms. It collects multi-source signals in real time, performs fine processing and feature extraction, accurately monitors and reliably identifies abnormal vibrations of wind turbines in complex environments along the coast or at sea. This allows for timely detection of early abnormal characteristics in wind turbine operation, effectively distinguishing between normal changes caused by environmental factors and genuine faults, thus enabling early fault detection and providing sufficient time for wind turbine maintenance and repair. By establishing a scenario feature template library and quantifying abnormal feature indicators, it effectively distinguishes between the influence of environmental factors and actual structural damage, significantly reducing false alarm rates and improving the adaptability of the wind turbine system in complex coastal / offshore environments. Employing a multi-technology integrated diagnostic strategy, it organically combines vibration analysis, acoustic detection, and machine learning algorithms. Based on multi-evidence fusion, it achieves precise fault location and identification, improving the accuracy of fault diagnosis and providing strong support for the safe and reliable operation of offshore wind turbines. Furthermore, through a dual-loop feedback optimization mechanism and a periodic model update mechanism, it enhances the monitoring performance of the wind turbine system while enriching the fault case database to adapt to the dynamic changes in the offshore environment. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0075] Figure 1 This is a flowchart of the method for monitoring abnormal vibration during wind turbine operation according to the present invention.

[0076] Figure 2 This is a block diagram of the wind turbine abnormal operation vibration monitoring system of the present invention. Detailed Implementation

[0077] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0078] Example 1

[0079] This invention provides, for example Figure 1 The method for monitoring abnormal vibrations in wind turbine operation based on multi-sensor fusion, as shown, includes the following steps:

[0080] S1. Deploy multiple types of sensors on the wind turbine, including acceleration sensors installed on the nacelle, tower and blades, displacement sensors installed on the foundation, acoustic sensors installed on the blades, as well as anemometers, tidal sensors and temperature and humidity sensors to monitor the environment. Collect data from each sensor in real time, use the IEEE1588 protocol for sensor data synchronization processing, combine wavelet denoising for preprocessing, and then use principal component analysis (PCA) combined with independent component analysis (ICA) to extract feature data.

[0081] Specifically, the data acquisition steps for each sensor are as follows:

[0082] For the various types of sensors deployed on the wind turbine, the sensor types and their corresponding quantities were counted. The sensor types were categorized into acceleration sensors, displacement sensors, acoustic sensors, anemometers, tidal sensors, and temperature and humidity sensors, and were calibrated as follows: ;

[0083] The total number of corresponding sensor types is: ;

[0084] The statistically analyzed sensors were then assigned type and quantity numbers to label each sensor deployed on the wind turbine. For example, the first acceleration sensor installed on the wind turbine was labeled as... Then the first one installed on the fan One accelerometer, labeled as ;

[0085] The first displacement sensor installed on the wind turbine is marked as Then the first one installed on the fan One displacement sensor, labeled as ;

[0086] The first acoustic sensor installed on the wind turbine is marked as Then the first one installed on the fan One acoustic sensor, labeled as ;

[0087] The first anemometer installed on the wind turbine is marked as Then the first one installed on the fan One anemometer, marked as ;

[0088] The first tidal sensor installed on the wind turbine is marked as Then the first one installed on the fan One tide sensor, marked as ;

[0089] The first temperature sensor installed on the fan is marked as Then the first one installed on the fan Unit temperature sensor, marked as ;

[0090] The first humidity sensor installed on the fan is marked as Then the first one installed on the fan Individual humidity sensor, marked as ;

[0091] Multiple types of sensors are activated to acquire data over a continuous period of time, forming a time series of sensor data. The expression for each sensor data is as follows: , , , , , , ,and In the formula, Represented as statistics One acceleration sensor In the Accelerometer data collected at each sampling time point Represented as statistics Displacement sensor In the Displacement sensor data collected at each sampling time, Represented as statistics Acoustic sensor In the Acoustic sensor data collected at each sampling time. Represented as statistics Anemometer In the Anemometer data collected at each sampling time Represented as statistics Tide sensor In the Tide sensor data collected at each sampling time. Represented as statistics Temperature sensor In the Temperature sensor data collected at each sampling time, Represented as statistics One humidity sensor In the Humidity sensor data collected at each sampling time. This represents the total number of sampling times. Represented as the timestamp of the sampling point;

[0092] The data from each sensor is integrated into a multi-sensor dataset, where the expression for the multi-sensor dataset is: In the formula, It is represented as a multi-sensor dataset.

[0093] Specifically, the steps for data synchronization processing and wavelet denoising preprocessing are as follows:

[0094] For each sensor, there is a time offset and network transmission delay between the local timestamp and the master clock timestamp;

[0095] The time offset and network transmission delay were calculated using the interleaved signaling mechanism of the IEEE 1588PTP protocol on the data from each sensor. The formula for calculating the time offset is as follows: In the formula, Represented as time offset, This represents the time when the synchronization message is received from the clock. This represents the timestamp at which the master clock sent the synchronization message request. This indicates the time when the master clock receives the delay request message. This represents the time when the delay request message is sent from the clock; the formula for calculating network transmission delay is... In the formula, This is expressed as network transmission delay;

[0096] Based on time offset and network transmission delay correction of time series data from various sensors, the timestamps of all sensor data are unified to a single timestamp on a master clock. This is used to eliminate timing errors caused by communication delays and differences in sampling start times, generating a time-aligned multi-sensor dataset. The expression for the unified timestamp is as follows: In the formula, This is represented as a corrected, unified timestamp. Represented as a local timestamp from the clock; the expression for a time-aligned multi-sensor dataset is... ,and , , , , , , In the formula, Represented as a time-aligned multi-sensor dataset, These are represented as time-aligned data from each sensor.

[0097] Wavelet denoising preprocessing is performed on a time-aligned multi-sensor dataset. Wavelet basis functions are selected, and the dataset is decomposed into wavelet coefficients of different scales through wavelet transform. The expression for the wavelet coefficients is as follows: In the formula, Represented as wavelet coefficients, Represented as a scale parameter. Represented as translation parameters, Represented as the wavelet mother function, It is represented as the complex conjugate of the wavelet mother function;

[0098] Applying a threshold function to the wavelet coefficients removes wavelet coefficients with small amplitudes, yielding the processed detail coefficients. The expression for the threshold function is: In the formula, This is represented as the new wavelet coefficients after thresholding. Represented as a soft threshold function, This is represented as the threshold for distinguishing between signal and noise;

[0099] Wavelet reconstruction is performed using the processed detail coefficients to obtain the denoised multi-sensor dataset, where the expression for the denoised multi-sensor dataset is: In the formula, This is represented as the denoised output multi-sensor dataset. It is represented as the wavelet admissibility constant.

[0100] Specifically, the steps for extracting feature data are as follows:

[0101] The multi-sensor dataset, after synchronization and denoising, is standardized and organized into an observation matrix, where the expression for the observation matrix is: In the formula, Represented as an observation matrix, This is expressed as the number of sampling points. Represented as the first The denoised output multi-sensor dataset collected from each sampling point Represented as matrix transpose;

[0102] Based on the observation matrix, the covariance matrix is ​​calculated using principal component analysis (PCA). The formula for calculating the covariance matrix is ​​as follows: ,and In the formula, Represented as a covariance matrix, Represented as a mean vector, Represented as the first The denoised output multi-sensor dataset collected from each sampling point, and ;

[0103] The covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues. The expression for covariance matrix decomposition is: In the formula, Represented as an eigenvalue matrix, Represented as an eigenvector matrix;

[0104] Select the eigenvectors corresponding to the top N largest eigenvalues ​​as principal components, and project the standardized observation matrix onto the principal components to obtain the principal component score matrix. The expression for the principal component score matrix is: In the formula, This represents the principal component score matrix. It is represented as the set of the directions of the first N principal components, where N represents the number of principal components to be retained;

[0105] The principal component score matrix is ​​input into the independent component analysis (ICA) method. The goal of ICA is to find a separation matrix that satisfies the condition that the independent component matrix is ​​equal to the product of the principal component score matrix and the separation matrix.

[0106] The FastICA algorithm is used to iteratively solve for the separation matrix. The FastICA algorithm first initializes the separation matrix and then iteratively updates it until convergence, thus obtaining the independent component matrix. The expression for the independent component matrix is: In the formula, Represented as an independent component matrix, Represented as a separation matrix;

[0107] The independent component matrix is ​​used as the eigenvector. This forms the extracted feature data.

[0108] S2. Establish a scene feature template library for complex marine environments, including strong wind-tide combined effect feature mode, blade moisture absorption feature mode, and composite scene feature mode. Compare the extracted feature data with the scene feature template library, use the cosine similarity algorithm to identify the scene where the wind turbine is located, define the feature threshold of the abnormal early warning mechanism, and trigger the abnormal early warning when the extracted feature data exceeds the feature threshold.

[0109] Specifically, the steps for establishing the scene feature template library are as follows:

[0110] Based on long-term historical data of wind turbines in a healthy state, and covering different scenarios such as combined strong wind and tidal effects, blade moisture absorption, and complex scenarios, multi-sensor data were obtained during the actual operation of wind turbines.

[0111] We still use the principal component analysis (PCA) combined with independent component analysis (ICA) to extract feature vectors from the data collected for each scenario.

[0112] The extracted feature vectors are clustered using the unsupervised clustering method K-Means to match feature patterns.

[0113] The feature vectors are divided into three clusters, and the centroid of each cluster constitutes a scene feature template, including the strong wind-tidal combined effect feature pattern, the leaf moisture absorption feature pattern, and the composite scene feature pattern.

[0114] For each scene feature template, calculate the Euclidean distance distribution from all feature vectors within the cluster to the centroid, and set a dynamic threshold to determine whether the data points of the new feature vector are abnormal with respect to the scene feature template.

[0115] Analyze the statistical distribution of all feature vector data points falling into each cluster with their corresponding scene feature templates, and mark the data points of abnormal scenes.

[0116] Specifically, the steps of the cosine similarity algorithm for scene recognition are as follows:

[0117] Calculate the cosine similarity between the data points of the new feature vector and each template in the scene feature template library;

[0118] Select the scene corresponding to the template with the highest similarity as the scene where the wind turbine is currently located;

[0119] After identifying the scene, calculate the Euclidean distance between the data points of the new feature vector and the centroid of the scene feature template;

[0120] The Euclidean distance is compared with a dynamic threshold. When the Euclidean distance exceeds the dynamic threshold, the scene is marked as abnormal and an abnormality warning is triggered.

[0121] S3. Based on the scene identified by the extracted feature data, a precise scene model is established using physical modeling combined with data-driven modeling. This includes establishing a wind turbine-ocean environment coupling model for the strong wind-tide combined action scenario and a moisture absorption-mechanical performance change model for the blade moisture absorption scenario. At the same time, the XGBoost algorithm is used to establish a data-driven compensation model to correct model errors. The extracted feature data is input into the corresponding scene model to simulate normal operation. The abnormal feature indicators are compared with the actual data to quantify abnormal feature indicators, including the low frequency modulation index LMI, moisture absorption influence factor MIF, and composite anomaly index CAI, to refine the scene classification, set abnormal feature thresholds, and determine the anomaly type and severity based on the comparison and quantification of abnormal feature indicators.

[0122] Specifically, the steps for building an accurate scene model are as follows:

[0123] A coupled model of wind turbine and marine environment and a model of moisture absorption and mechanical performance change were established by using physical modeling combined with data-driven modeling.

[0124] Among them, for establishing a coupled wind turbine-ocean environment model:

[0125] The dynamic behavior of the wind turbine under the combined effects of strong winds and tides is collected, including environmental information vectors such as wind speed, wind direction, tidal height, and significant wave height, as well as the wind turbine operating state vectors including tower top displacement / velocity / acceleration, nacelle acceleration, blade pitch angle, and blade tip displacement. The environmental information vectors and the wind turbine operating state vectors are respectively calibrated as follows: and ,and In the formula, Represented as an environmental information vector, Expressed as wind speed, Indicated as wind direction, Expressed as tidal height, Represented as significant wave height, This is represented as transpose. In the formula, This is represented as the state vector of the wind turbine operation. This is expressed as the displacement at the top of the tower. Expressed as the velocity at the top of the tower, This is expressed as the acceleration at the top of the tower. Expressed as cabin acceleration, This is expressed as the blade pitch angle. Expressed as tip displacement;

[0126] Based on environmental information vectors and state vectors, a force vector function generated by strong winds is constructed to calculate the aerodynamic load vector acting on the wind turbine. Simultaneously, a force vector function generated by tidal forces is constructed to calculate the hydrodynamic load vector of the forces exerted by waves and ocean currents on the wind turbine foundation structure. The formula for calculating the aerodynamic load vector is as follows: In the formula, It is represented as an aerodynamic load vector. Represented as a force vector function generated by strong winds;

[0127] The formula for calculating the hydrodynamic load vector is as follows: In the formula, Represented as a hydrodynamic load vector, Represented as a force vector function generated by tidal forces;

[0128] Based on aerodynamic and hydrodynamic load vectors, and combined with data-driven inertial, damping, and elastic forces during wind turbine operation, a coupled dynamic behavior model is obtained to describe the system's trajectory under the influence of internal and external forces. The expression for this dynamic behavior model is as follows: In the formula, This is represented as the mass matrix describing the inertial forces during wind turbine operation. The damping matrix represents the damping force during wind turbine operation. This represents the stiffness matrix of the elastic force during wind turbine operation. It is represented as the second derivative of the state vector of the wind turbine operation. The first derivative of the wind turbine's operating state vector;

[0129] For the hygroscopic-mechanical property change model:

[0130] Collection of the location within the composite material of the wind turbine blades In The data includes the constantly absorbed moisture concentration, ambient temperature, and humidity, among which the location of the wind turbine blade composite material is considered. In The concentration of water absorbed at any given time is calibrated as ;

[0131] Based on the theoretical water diffusion coefficient of water diffusion kinetics, Fick's second law is used to construct a diffusion equation for water within the composite material from a high-concentration region to a low-concentration region. The water concentration field is calculated, and boundary conditions are defined based on ambient humidity. The expression for the diffusion equation constructed using Fick's second law is as follows: In the formula, This represents the diffusion pattern of water concentration over time. Represented as the divergence operator, Expressed as the theoretical moisture diffusion coefficient, This is represented as a gradient operator for water concentration;

[0132] Based on the moisture concentration field and combined with the equivalent elastic modulus and damping coefficient of the blade material, a moisture absorption-mechanical performance change model is established to obtain the mechanical performance degradation relationship between mechanical properties and moisture absorption, thus describing the vibration response of the fan after moisture absorption in humid coastal air. The expression for the moisture absorption-mechanical performance change model is as follows: In the formula, Expressed as the equivalent elastic modulus, Expressed as the elastic modulus of a material in its dry state. It is expressed as the degradation coefficient of the elastic modulus. Expressed as the equivalent elastic modulus, It represents the damping coefficient of the material in a dry state. It is expressed as the degradation coefficient of damping;

[0133] The established wind turbine-ocean environment coupling model and moisture absorption-mechanical performance change model will simulate the operating state of the wind turbine in its environment and output the feature vector predicted by the scenario model. The expression for the feature vector predicted by the scenario model is as follows: In the formula, This is represented as the feature vector predicted by the scene model.

[0134] To further clarify, the steps for establishing a data-driven compensation model are as follows:

[0135] The feature vectors predicted by the scenario model and the feature vectors actually extracted from the historical data of the wind turbine's operating health status are used as training sample sets.

[0136] Based on the training sample set, a data-driven compensation model is trained using the XGBoost algorithm, and the prediction error value of the data-driven compensation model is output. The expression for the data-driven compensation model is: ,and In the formula, This is represented as the model's prediction error value. This is represented as a feature vector actually extracted from the historical data of the wind turbine's operational health status. This is represented as the feature vector predicted by the scene model. This is represented as the data-driven compensation result for the prediction error value. This is represented as the XGBoost regression algorithm. This is represented as the tree structure and weight parameters in the XGBoost algorithm;

[0137] By defining the number of decision trees and minimizing the objective function, the tree structure and weight parameters of the data-driven compensation model are learned and optimized, resulting in a trained data-driven compensation model. The expression for the minimization objective function is: In the formula, Represented as minimizing the loss function, This is represented by the number of decision trees. Represented as the first The data output of each decision tree drives the compensation results. This is represented as a regularization term to prevent overfitting in the model;

[0138] The feature vectors predicted by the scene model are input into the trained data to drive the compensation model, which outputs the predicted feature vectors that compensate for the error values ​​of the scene model, thus further refining the scene model. The expression for refining the scene model is as follows: In the formula, This represents the accurate scene model after compensation.

[0139] Specifically, the quantification steps for abnormal feature indicators are as follows:

[0140] Real-time online monitoring of coastal / offshore wind turbine operation, using synchronously denoised characteristic data from actual measurements. The predicted feature vector output after simulating the normal operation of the wind turbine with an accurate scene model. Comparative analysis was conducted to quantitatively calculate the abnormal characteristic indicators of the low-frequency modulation index (LMI), the moisture absorption effect factor (MIF), and the composite anomaly index (CAI).

[0141] The Low Frequency Modulation Index (LMI) describes whether the low-frequency modulation intensity caused by tides / waves in the actual signal exceeds the normal range predicted by the model. It is calculated by performing a Hilbert transform on the residual signal (the difference between the real-time monitored feature data and the predicted feature vector output by the accurate scene model) to obtain the envelope. Spectral analysis is then performed on the envelope to obtain the spectrum of the actual monitored signal and the spectrum of the simulated signal from the accurate scene model. The energy at which the tidal / wave characteristic frequencies affect the signal is then identified, and the LMI is calculated. The formula for calculating the LMI is as follows: ,and In the formula, It is represented as the low-frequency modulation index. Represented as the characteristic frequency of tides / waves, Expressed as frequency bandwidth, The power spectral density is expressed as the envelope. It is represented as a weighting function of the modulation frequency. Represented as the envelope of the residual signal. Represented as residual signal;

[0142] The Moisture Absorption Influence Factor (MIF) quantifies the degree to which blade vibration characteristics deviate from the expected baseline under current humidity conditions. A weight matrix is ​​set based on prior knowledge of humidity sensitivity for different scenario characteristics. The MIF is calculated using Mahalanobis distance based on the residual signal. The formula for calculating the MIF is as follows: In the formula, This is expressed as the moisture absorption influencing factor. This is represented by setting a weight matrix based on prior knowledge of humidity sensitivity to different scene characteristics. Represented as the symbol for Mahalanobis distance calculation;

[0143] The Composite Anomaly Index (CAI) integrates multiple anomaly indices to calculate an overall composite anomaly score. Based on the Low Frequency Modulation Index (LMI) and the Moisture Absorption Factor (MIF), a weighted average method is used to calculate the CAI. The formula for calculating the CAI is as follows: ,and In the formula, This is represented as a composite abnormality index. This represents the weighting coefficient corresponding to the low-frequency modulation index (LMI). This represents the weighting coefficient of the corresponding moisture absorption factor (MIF).

[0144] To further clarify, the steps for determining the anomaly type and severity are as follows:

[0145] The thresholds for the low-frequency modulation index, the hygroscopic influence factor, and the composite anomaly index are defined and calibrated as follows: , , ;

[0146] By comparing the abnormal characteristic indicators of the quantitatively calculated low-frequency modulation index (LMI), moisture absorption effect factor (MIF), and composite anomaly index (CAI) with the threshold values ​​defined for the LMI, MIF, and CAI, the anomaly types are determined, including when... and At that time, the anomaly type was determined to be structural loosening or aerodynamic asymmetry caused by strong wind-tidal modulation;

[0147] when and When the abnormality is determined to be caused by moisture absorption by the blades, the material properties are degraded or damaged.

[0148] when Complex abnormalities;

[0149] The severity of an anomaly is determined by the ratio of abnormal characteristic indicators to a threshold, and then classified into severity levels: mild, moderate, and severe. The formula for calculating the severity of an anomaly is as follows: In the formula, This indicates the severity of the abnormality. This is represented by the maximum value sign, when... At that time, it was classified as mildly abnormal;

[0150] when At that time, it was classified as moderately abnormal;

[0151] when At that time, it was classified as a severe abnormality.

[0152] S4. By comprehensively utilizing vibration analysis technology, acoustic detection technology, and machine learning diagnostic algorithms, fault diagnosis is performed on quantified abnormal features to generate diagnostic results that accurately locate the fault. A scene adaptive mechanism is developed to adaptively adjust diagnostic parameters according to scene characteristics and automatically trigger corresponding processing decisions.

[0153] Specifically, the steps for generating the diagnostic results are as follows:

[0154] Receive quantified abnormal feature indicators, combine them with extracted feature data, and list and statistically analyze the types of faults and normal states that may occur in the wind turbine operation status to build a status identification framework.

[0155] Vibration analysis technology, acoustic detection technology, and machine learning diagnostic algorithms are used as evidence sources (BPA) for fault diagnosis. Fault diagnosis probabilities are assigned to different state types in the state recognition framework, including generating vibration analysis evidence sources by analyzing the spectrum, envelope spectrum, and modal parameters of vibration signals.

[0156] By analyzing the amplitude, count, energy, and location information of acoustic emission signals, acoustic detection evidence sources are generated.

[0157] By inputting the feature data vector and abnormal feature indicators extracted in real time during wind turbine operation into the probability distribution output by a pre-trained multi-class XGBooster, a source of machine learning evidence is generated.

[0158] The decision-level fusion combination rule using DS evidence theory integrates vibration analysis evidence sources, acoustic detection evidence sources, and machine learning evidence sources to obtain a comprehensive quality function, and then formulates processing decisions based on the comprehensive quality function.

[0159] The diagnostic parameters for identifying the scene are dynamically adjusted using a scene adaptation mechanism. The scene features are mapped to the diagnostic parameters, and the final diagnostic results, including the fault type, diagnostic confidence, and fault location, are generated.

[0160] S5. Based on the system feedback of diagnostic results and processing decisions, establish a dual-loop feedback mechanism to continuously optimize system performance. The inner loop feedback is to monitor the wind turbine operation results in real time and adjust sensor data acquisition and feature data extraction. The outer loop feedback is to update model parameters and knowledge base regularly. At the same time, a knowledge base including case library, model library and rule library is constructed, and the knowledge base is analyzed and updated regularly to optimize the system's monitoring and diagnostic capabilities.

[0161] Specifically, the steps for establishing the dual-loop feedback mechanism are as follows:

[0162] Inner loop feedback: Based on the diagnostic confidence in the diagnostic results, the data acquisition and feature extraction strategies of the system front end are optimized in real time, including sensor sampling rate, data sampling time window length, feature selection weight vector and filtering parameters;

[0163] Outer loop feedback: The system model is periodically retrained using historical fault data accumulated from wind turbine operation. In each update cycle, the parameters of each model are updated using gradient descent optimization algorithm based on the characteristic data of real-time wind turbine operation. At the same time, the knowledge base is updated based on new fault diagnosis cases and operation data.

[0164] Specifically, the knowledge base is constructed as follows: The knowledge base includes a case library, a model library, and a rule library. For the case library: Fault cases that occur during wind turbine operation are stored in the case library in a structured form. Each fault case is a tuple, which includes the environment, characteristics, indicator data, scenario, diagnostic results with confidence level, the processing decision taken, and the operational results after the decision processing. Furthermore, the fault cases are automatically added to the library based on the verified data during the outer loop feedback process.

[0165] For the model library: it stores historical models of different scenarios and versions, including but not limited to accurate scenario models, data-driven compensation models, K-Means clustering analysis models, principal component analysis (PCA) combined with independent component analysis (ICA) models, XGBoost models, and machine learning models. After each retraining of the model through outer loop feedback, the new model and its performance are stored in the model library as the new version. Otherwise, the system can roll back to a historical version with better performance.

[0166] For the rule base: it stores the logical rules for diagnosis and decision-making. Using decision tree or association rule algorithms, it learns frequently occurring "condition-outcome" patterns from fault cases and feature data in the case library to generate candidate rules. After the candidate rules are reviewed, corrected and confirmed by experts in the field, they are officially added to the rule base. The rule base is also regularly reviewed to merge redundant rules and resolve conflicting rules.

[0167] Example 2

[0168] This invention provides, for example Figure 2 The wind turbine operation abnormal vibration monitoring system shown includes a multi-sensor fusion-based sensor module for collecting multi-source data: multiple types of sensors, such as acceleration sensors, displacement sensors, acoustic sensors, anemometers, tidal sensors, and temperature and humidity sensors, are deployed on the nacelle, tower, and blades of the wind turbine. The system uses the IEEE 1588 protocol to synchronize the microsecond-level time of all sensors and collects sensor data on vibration, acoustics, and environment in real time when the wind turbine is operating in coastal / offshore conditions.

[0169] Data preprocessing and feature extraction module: Receives sensor data and extracts feature data using wavelet denoising combined with principal component analysis and PCA-ICA algorithm;

[0170] Scene modeling and anomaly feature quantification module: Construct a scene feature template library, identify scenes by comparing feature data, establish scene models by combining physical modeling with data-driven modeling, quantify anomaly feature indicators, further refine scene classification, and set anomaly feature thresholds to assess anomaly type and severity;

[0171] Multi-technology integrated fault diagnosis module: It comprehensively utilizes vibration analysis technology, acoustic detection technology and machine learning diagnostic algorithms to diagnose faults based on abnormal characteristic indicators, develops a scene adaptive mechanism, adjusts diagnostic parameters according to scene characteristics, and triggers corresponding processing decisions.

[0172] Feedback optimization knowledge base management module: Establish a dual-loop feedback mechanism to feed back diagnostic results and processing decisions to various modules of the system. After building and updating the knowledge base, adjust sensor data acquisition strategies, feature data extraction methods, scene model construction, and anomaly threshold settings based on the feedback information.

[0173] To further clarify, multiple types of sensors are installed on the wind turbine and networked together to ensure that the sensors can collect data normally and transmit the data to the monitoring system.

[0174] Deploying edge computing devices enables data acquisition and preliminary processing. Edge computing devices can preprocess the acquired data locally, reducing data transmission volume and improving system response speed.

[0175] Establish a cloud computing platform for in-depth data analysis and intelligent decision-making. The cloud computing platform has powerful computing and storage capabilities, which can analyze and process large amounts of data to support the system's decision-making.

[0176] Develop a human-machine interface to display monitoring results and alarm information. The human-machine interface can help operation and maintenance personnel understand the operating status of the wind turbine in real time and handle abnormal situations in a timely manner.

[0177] Regular system calibration and model updates are necessary to ensure the accuracy and reliability of the system. System calibration guarantees the measurement accuracy of the sensors, while model updates enable the system to adapt to changes in the marine environment.

[0178] The wind turbine abnormal vibration monitoring system based on multi-sensor fusion provided in this embodiment of the invention is implemented by the above-mentioned wind turbine abnormal vibration monitoring method based on multi-sensor fusion. For details of the specific method and process of the wind turbine abnormal vibration monitoring system based on multi-sensor fusion, please refer to the above-mentioned embodiment of the wind turbine abnormal vibration monitoring method based on multi-sensor fusion, which will not be repeated here.

[0179] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0180] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0181] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0182] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0183] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring abnormal vibration during wind turbine operation based on multi-sensor fusion, characterized in that, Includes the following steps: S1. Deploy multiple types of sensors on the wind turbine to collect data from each sensor in real time. Use the IEEE1588 protocol to synchronize and process the sensor data. Combine wavelet denoising for preprocessing and then use principal component analysis (PCA) and independent component analysis (ICA) to extract feature data. S2. Establish a scene feature template library for complex marine environments, compare the extracted feature data with the scene feature template library, use the cosine similarity algorithm to identify the scene where the wind turbine is located, define the feature threshold of the abnormal early warning mechanism, and trigger the abnormal early warning when the extracted feature data exceeds the feature threshold. S3. Based on the scene identified by the extracted feature data, a precise scene model is established using physical modeling combined with data-driven modeling. The extracted feature data is input into the corresponding scene model to simulate normal operation. The abnormal feature indicators are compared with the actual data to quantify the abnormal features, refine the scene classification, set abnormal feature thresholds, and determine the abnormal type and severity based on the comparison and quantification of abnormal feature indicators. S4. By comprehensively utilizing vibration analysis technology, acoustic detection technology, and machine learning diagnostic algorithms, fault diagnosis is performed on quantified abnormal features to generate diagnostic results that accurately locate the fault. A scene adaptive mechanism is developed to adaptively adjust diagnostic parameters according to scene characteristics and automatically trigger corresponding processing decisions. S5. Based on the system feedback of diagnostic results and processing decisions, establish a dual-loop feedback mechanism to continuously optimize system performance. At the same time, build a knowledge base, analyze and update the knowledge base regularly, and optimize the system's monitoring and diagnostic capabilities.

2. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 1, characterized in that, The steps for collecting data from each sensor are as follows: For the various types of sensors deployed on the wind turbine, the sensor types and the corresponding number of sensor types are counted. Among them, the sensor types are divided into acceleration sensors, displacement sensors, acoustic sensors, anemometers, tidal sensors and temperature and humidity sensors. The various types of sensors were statistically analyzed and assigned type and quantity numbers to identify each sensor deployed on the wind turbine. Multiple types of sensors are activated to collect data over a continuous period of time, forming a time series of sensor data. Data from various sensors is integrated into a multi-sensor dataset.

3. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 2, characterized in that, The steps of data synchronization processing and wavelet denoising preprocessing are as follows: For each sensor, there is a time offset and network transmission delay between the local timestamp and the master clock timestamp; The time offset and network transmission delay are calculated by using the interleaved signaling mechanism of the IEEE 1588PTP protocol on the data from each sensor. Based on time offset and network transmission delay correction of time series data from each sensor, the timestamps of all sensor data are unified to a unified timestamp on a master clock to generate a time-aligned multi-sensor dataset. Wavelet denoising preprocessing is performed on time-aligned multi-sensor datasets. Wavelet basis functions are selected and decomposed into wavelet coefficients of different scales through wavelet transform. Apply a threshold function to the wavelet coefficients to remove wavelet coefficients with small amplitudes, and obtain the processed detail coefficients. Wavelet reconstruction is performed using the processed detail coefficients to obtain the denoised multi-sensor dataset.

4. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 3, characterized in that, The steps for extracting the feature data are as follows: The multi-sensor dataset, after synchronization and denoising, is standardized and organized into an observation matrix; Based on the observation matrix, the covariance matrix is ​​calculated using principal component analysis (PCA). Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues; Select the eigenvectors corresponding to the top N largest eigenvalues ​​as principal components, and project the standardized observation matrix onto the principal components to obtain the principal component score matrix. The principal component score matrix is ​​input into the independent component ICA analysis method, and the FastICA algorithm is used to iteratively solve the separation matrix. The FastICA algorithm requires initializing the separation matrix first, and then iteratively updating the initial separation matrix until convergence, thus obtaining the independent component matrix. The independent component matrix is ​​used as the feature vector to form the extracted feature data.

5. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 1, characterized in that, The steps for establishing the scene feature template library are as follows: Based on long-term historical data of wind turbines in a healthy state, and covering different scenarios such as combined strong wind and tidal effects, blade moisture absorption, and complex scenarios, multi-sensor data were obtained during the actual operation of wind turbines. We still use the principal component analysis (PCA) combined with independent component analysis (ICA) to extract feature vectors from the data collected for each scenario. The extracted feature vectors are clustered using the unsupervised clustering method K-Means to match feature patterns; The feature vectors are divided into three clusters, and the centroid of each cluster constitutes a scene feature template, including the strong wind-tidal combined effect feature pattern, the leaf moisture absorption feature pattern, and the composite scene feature pattern. For each scene feature template, calculate the Euclidean distance distribution from all feature vectors within the cluster to the centroid, and set a dynamic threshold to determine whether the data points of the new feature vector are abnormal with respect to the scene feature template. Analyze the statistical distribution of all feature vector data points falling into each cluster with their corresponding scene feature templates, and mark the data points of abnormal scenes.

6. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 1, characterized in that, The steps of the cosine similarity algorithm for scene recognition are as follows: Calculate the cosine similarity between the data points of the new feature vector and each template in the scene feature template library; Select the scene corresponding to the template with the highest similarity as the scene where the wind turbine is currently located; After identifying the scene, calculate the Euclidean distance between the data points of the new feature vector and the centroid of the scene feature template; The Euclidean distance is compared with a dynamic threshold. When the Euclidean distance exceeds the dynamic threshold, the scene is marked as abnormal and an abnormality warning is triggered.

7. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 1, characterized in that, The steps for establishing the accurate scene model are as follows: A coupled model of wind turbine and marine environment and a model of moisture absorption and mechanical performance change were established by using physical modeling combined with data-driven modeling. Among them, for establishing a coupled wind turbine-ocean environment model: Collect the dynamic behavior of the wind turbine under the combined action of strong wind and tide, including environmental information vectors and the state vector of wind turbine operation; Based on environmental information vectors and state vectors, a force vector function generated by strong winds is constructed to calculate the aerodynamic load vector of the aerodynamic forces acting on the wind turbine. At the same time, a force vector function generated by tidal forces is constructed to calculate the hydrodynamic load vector of the forces of waves and ocean currents acting on the wind turbine foundation structure. Based on aerodynamic load vectors and hydrodynamic load vectors, and combined with data-driven inertial forces, damping forces, and elastic forces during wind turbine operation, a coupled dynamic behavior model is obtained. For the hygroscopic-mechanical property change model: Collect data on the moisture concentration, ambient temperature, and humidity absorbed by the composite material of the wind turbine blades; Based on the theoretical water diffusion coefficient of water diffusion dynamics, Fick's second law is used to construct the diffusion equation of water in the leaf from the high concentration region to the low concentration region inside the composite material, calculate the water concentration field, and define the boundary conditions according to the ambient humidity. Based on the moisture concentration field, combined with the equivalent elastic modulus and damping coefficient of the blade material, a moisture absorption-mechanical property change model is established to obtain the mechanical property degradation relationship between mechanical properties and moisture absorption. The established wind turbine-ocean environment coupling model and moisture absorption-mechanical performance change model will simulate the operating state of the wind turbine in the scenario and output the feature vector predicted by the scenario model.

8. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 1, characterized in that, The steps for generating the diagnostic results are as follows: Receive quantified abnormal feature indicators, combine them with extracted feature data, and list and statistically analyze the types of faults and normal states that may occur in the operation of the wind turbine to build a state recognition framework. Vibration analysis technology, acoustic detection technology, and machine learning diagnostic algorithms are used as evidence sources (BPA) for fault diagnosis. Fault diagnosis probabilities are assigned to different state types in the state recognition framework, including generating vibration analysis evidence sources by analyzing the spectrum, envelope spectrum, and modal parameters of vibration signals. By analyzing the amplitude, count, energy, and location information of acoustic emission signals, acoustic detection evidence sources are generated. By inputting the feature data vector and abnormal feature indicators extracted in real time during wind turbine operation into the probability distribution output by a pre-trained multi-class XGBooster, a source of machine learning evidence is generated. The decision-level fusion combination rule using DS evidence theory integrates vibration analysis evidence sources, acoustic detection evidence sources, and machine learning evidence sources to obtain a comprehensive quality function, and then formulates processing decisions based on the comprehensive quality function. The diagnostic parameters for scene recognition are dynamically adjusted using a scene adaptation mechanism, the relationship between scene features and diagnostic parameters is mapped, and diagnostic results are generated.

9. The method for monitoring abnormal vibration of wind turbine operation based on multi-sensor fusion according to claim 1, characterized in that, The quantification steps for the abnormal feature indicators are as follows: Real-time online monitoring of coastal / offshore wind turbine operation is conducted. The actual measured synchronously denoised feature data is compared and analyzed with the predicted feature vector output after simulating the normal operation of the wind turbine by a precise scenario model. Abnormal feature indicators of low frequency modulation index (LMI), moisture absorption influence factor (MIF), and composite anomaly index (CAI) are quantitatively calculated. For the low-frequency modulation index (LMI), the residual signal of the difference between the real-time monitored feature data and the predicted feature vector output by the accurate scene model is subjected to Hilbert transform to obtain the envelope. The envelope is then subjected to spectral analysis to obtain the spectrum of the actual monitored signal and the spectrum of the accurate scene model simulated signal. The energy affected by the tidal / wave characteristic frequency is found, and the low-frequency modulation index (LMI) is calculated. For the moisture absorption impact factor (MIF), a weight matrix is ​​set based on prior knowledge of the humidity sensitivity of different scene characteristics, and the moisture absorption impact factor (MIF) is calculated based on the residual signal using Mahalanobis distance. For the composite anomaly index CAI, the weighted average method is used to calculate the composite anomaly index CAI based on the low-frequency modulation index LMI and the moisture absorption effect factor MIF.

10. A wind turbine abnormal operation vibration monitoring system based on multi-sensor fusion, implemented by the wind turbine abnormal operation vibration monitoring method based on multi-sensor fusion as described in any one of claims 1-9, characterized in that, This includes a multi-source data acquisition module that deploys sensors: various types of sensors, such as acceleration sensors, displacement sensors, acoustic sensors, anemometers, tidal sensors, and temperature and humidity sensors, are deployed on the nacelle, tower, and blades of the wind turbine. The microsecond-level time of all sensors is synchronized using the IEEE1588 protocol, and sensor data on vibration, acoustics, and environment are collected in real time when the wind turbine is operating in coastal / offshore conditions. Data preprocessing and feature extraction module: Receives sensor data and extracts feature data using wavelet denoising combined with principal component analysis and PCA-ICA algorithm; Scene modeling and anomaly feature quantification module: Construct a scene feature template library, identify scenes by comparing feature data, establish scene models using physical modeling combined with data-driven modeling, quantify anomaly feature indicators, further refine scene classification, and set anomaly feature thresholds to assess anomaly type and severity; Multi-technology integrated fault diagnosis module: It comprehensively utilizes vibration analysis technology, acoustic detection technology and machine learning diagnostic algorithms to diagnose faults based on abnormal characteristic indicators, develops a scene adaptive mechanism, adjusts diagnostic parameters according to scene characteristics, and triggers corresponding processing decisions. Feedback optimization knowledge base management module: Establish a dual-loop feedback mechanism to feed back diagnostic results and processing decisions to various modules of the system. After building and updating the knowledge base, adjust sensor data acquisition strategies, feature data extraction methods, scene model construction, and anomaly threshold settings based on the feedback information.

Images