Fault diagnosis and early warning method and system for wind power plant

Abstract

The invention belongs to the technical field of wind power generation, and particularly relates to a fault diagnosis and early warning method and system for a wind power plant, and the method comprises the steps: collecting the state data of a plurality of physical fields through a plurality of types of sensor arrays disposed at key parts of a fan; performing preprocessing and feature extraction on the data to obtain high-dimensional feature data; driving the field-level digital twinborn model to perform real-time simulation based on the high-dimensional feature data, and determining an initial fault recognition result and severity by comparing a simulation predicted value with actual sensing data and combining with an unsupervised anomaly detection algorithm; a diagnosis result is calibrated through multi-modal high-dimensional feature data fusion; predicting the remaining useful life of the key component based on the fault severity and the evolution trend thereof, and generating a collaborative operation and maintenance decision by using a multi-objective optimization algorithm; and an operation and maintenance decision is automatically executed through a closed-loop feedback execution mechanism. According to the invention, the accuracy of fault diagnosis and early warning of the wind power plant is improved.

Description

Technical Field This disclosure pertains to the field of wind power generation technology, and particularly relates to a method and system for fault diagnosis and early warning of wind power plants. Background Technology

[0001] As the global energy structure shifts towards green and low-carbon development, wind power, as one of the main forces, is becoming increasingly large-scale and complex, posing significant challenges to wind farm operation and maintenance. Wind turbines operate under harsh conditions year-round, resulting in high failure rates. Traditional periodic maintenance and reactive repair methods are costly and prone to causing substantial power generation losses. Therefore, developing intelligent fault diagnosis and early warning technologies capable of predictive maintenance is of great significance for ensuring the safe, stable, and economical operation of wind farms, and has become an urgent industry need and an inevitable trend in technological development. Currently, this field is evolving from analysis relying on single physiological models or data-driven approaches to a digital and intelligent approach that integrates physical mechanisms with big data and artificial intelligence. Simultaneously, the operation and maintenance model is also evolving from isolated single-unit management to collaborative optimization across the entire wind farm.

[0002] Existing technical solutions primarily rely on data provided by wind turbine data acquisition and monitoring systems, combined with traditional signal processing techniques or single machine learning algorithms, to construct fault diagnosis models. These solutions generally suffer from significant drawbacks: First, their data dimensions are limited, heavily dependent on data from data acquisition and monitoring systems, lacking deep fusion of multimodal data such as vibration, acoustics, and images, resulting in insensitivity to early, subtle fault characteristics and a high false negative rate. Second, the models have limited intelligence; most are supervised learning models, relying on large amounts of labeled historical fault data, exhibiting poor generalization ability to unknown new faults, and lagging model updates, failing to adapt to equipment performance degradation. Finally, the system functions are fragmented, with diagnostic, early warning, and maintenance decision-making processes disconnected. They typically only provide simple anomaly alarms, unable to accurately locate faults, assess severity, predict remaining lifespan, or generate optimized, comprehensive collaborative maintenance strategies, ultimately failing to support closed-loop automated management from perception to execution. Summary of the Invention

[0003] To address the aforementioned issues, this disclosure provides a fault diagnosis and early warning method and system for wind farms. It adopts a layered processing architecture with edge-cloud collaboration, deeply integrates multimodal sensing data and digital twin models, and introduces farm-level spatiotemporal correlation analysis and federated learning mechanisms. This enables highly sensitive identification and accurate location of early-stage minor faults, effectively distinguishes between inherent faults of individual units and cluster anomalies, and forms a complete closed loop from state perception to operation and maintenance execution through integrated remaining lifetime prediction and multi-objective optimization decision-making. Ultimately, this significantly improves the fault early warning capability and operational reliability of wind farms.

[0004] In a first aspect, this disclosure provides a method for fault diagnosis and early warning in wind power plants, the method comprising, By deploying multiple sensor arrays on key components of each wind turbine in the target wind farm, setting the preset range for each sensor and collecting several initial data corresponding to the multi-physical field state in real time based on the preset sampling frequency. Data preprocessing and feature extraction are performed on several initial data sets to obtain several high-dimensional feature data sets; Real-time simulation is performed based on a field-level digital twin model driven by several high-dimensional feature data to obtain simulation prediction values. The simulation prediction values ​​are compared with the actual sensing data, and combined with an unsupervised anomaly detection algorithm, the initial fault identification results and the initial fault severity are determined. Fault diagnosis is performed by fusing multimodal high-dimensional feature data to calibrate the initial fault identification results; Based on the initial fault severity and evolution trend of any faulty wind turbine, the remaining useful life of its key components is predicted. Based on the remaining useful life prediction results, a multi-objective optimization algorithm combined with multi-objective parameters is used to generate collaborative operation and maintenance decisions for the target wind farm. The collaborative operation and maintenance decisions are executed through a closed-loop feedback execution mechanism, which triggers corresponding control and scheduling instructions.

[0005] Furthermore, Based on a preset sampling frequency and preset range, several initial data corresponding to the state of multiple physics fields are collected in real time, specifically including: Multiple types of sensors are deployed on key components of each wind turbine to obtain multiple sensor arrays; The range is determined as the preset range based on the range of physical quantity changes of any wind turbine from standby to rated power operation and under extreme conditions. The potential fault characteristic frequency is determined based on the historical fault data of any key component of the wind turbine, and the sampling frequency range is determined based on the potential fault characteristic frequency as the preset sampling frequency. The array of multiple sensors is controlled to synchronously acquire data according to a preset sampling frequency to obtain several initial data. The key components include at least one of the following: gearbox, generator bearing housing, main shaft, blade root, blade surface, tower weld, tower base, and nacelle. The multiple types of sensors include at least two of the following: high-frequency acceleration sensor, acoustic emission sensor, strain gauge, inertial measurement unit, GPS displacement sensor, online oil monitoring sensor, magnetic flux sensor, and infrared thermal imager. The initial data includes at least two of the following: vibration signal, acoustic emission signal, structural strain signal, tower attitude signal, gearbox oil wear particle signal, generator air gap magnetic field signal, and infrared thermal imaging video stream.

[0006] Furthermore, Data preprocessing and feature extraction are performed on several sets of initial data, specifically including: The initial data is subjected to noise suppression processing based on wavelet packet transform to obtain noise-suppressed data. Anomaly removal data is obtained by identifying and replacing abnormal outliers in the noise suppression data based on an adaptive threshold algorithm. The anomaly removal data from the multi-channel system is time-synchronized and aligned to obtain preprocessed data. The optimal feature set is selected based on historical data using a feature selection algorithm. Based on the optimal feature set, feature data is extracted from the preprocessed data to obtain the high-dimensional feature data.

[0007] Furthermore, Real-time simulation is performed based on a field-level digital twin model driven by several of the aforementioned high-dimensional feature data, specifically including: A field-level digital twin model is constructed, which integrates a wind turbine mechanism dynamics model based on physical laws and a deep neural network surrogate model trained based on historical data. The high-dimensional feature data and operating parameters of each wind turbine are input into the field-level digital twin model, driving the model to perform real-time simulation according to a preset time step, and output simulation prediction values.

[0008] Furthermore, Determining the initial fault identification results and the initial fault severity includes: The simulation prediction values ​​and the actual sensing data are calculated dimension by dimension to generate a residual sequence; The residual sequence is subjected to moving average filtering. The processed residual sequence is input into an unsupervised anomaly detection algorithm based on a transfer learning framework. When the residual value is greater than the residual threshold, it is regarded as an abnormal residual sequence and an initial fault is determined to exist. When an initial fault exists, the time domain, frequency domain, and modal characteristics of the abnormal residual sequence are analyzed and matched with a preset component fault feature mapping library. When the matching degree exceeds a preset threshold, the fault source is determined to determine the initial fault identification result. Based on the statistical characteristic values ​​of the residual sequence and its trend characteristics over time, combined with the preset fault severity assessment model, the initial fault severity level is quantified.

[0009] Furthermore, Fault diagnosis is performed through the fusion of multimodal high-dimensional feature data, specifically including: Analyze the high-dimensional feature data corresponding to any sensor, and identify abnormal patterns based on the analysis results; Once the abnormal pattern is identified, the abnormal states of other sensor data within the same period are analyzed in conjunction with it to construct a multimodal chain of evidence. Based on a multimodal chain of evidence, multi-source information is fused through evidence theory algorithms to calculate the comprehensive diagnostic confidence. When the overall diagnostic confidence level exceeds the preset warning threshold, a calibrated fault diagnosis result is generated, and a warning signal of the corresponding level is triggered.

[0010] Furthermore, Predicting the remaining service life of its key components, specifically including: Obtain historical health data, current fault severity quantification, and evolution trend of key components of the faulty wind turbine; A sequence deep learning model that integrates physical degradation models is used for lifetime prediction; Output the probability distribution of the remaining useful life of the key component, including the mean life and confidence interval.

[0011] Furthermore, The collaborative operation and maintenance decision-making for the target wind farm is generated by combining multi-objective optimization algorithms with multi-objective parameters, specifically including: Establish a multi-objective optimization function with the objectives of minimizing total power generation loss and minimizing operation and maintenance costs; Set constraints including health constraints, environmental constraints, scheduling constraints, and resource constraints; A multi-objective optimization algorithm is used to solve the optimization function under constraints to generate a Pareto optimal solution set. Based on the wind farm's operational priority, the final solution is selected from the optimal solution set, and the output is a collaborative operation and maintenance decision that includes a list of wind turbines to be inspected, an inspection time window, a spare parts allocation plan, and a maintenance team assignment plan.

[0012] Furthermore, it also includes using graph neural networks to perform field-level spatiotemporal correlation analysis to distinguish between inherent faults of individual units and group abnormal faults caused by environmental factors, and to assess the risk of fault propagation in wind turbine clusters.

[0013] Secondly, based on the same inventive concept, this disclosure also provides a fault diagnosis and early warning system for wind power plants, the system comprising: The multimodal data sensing layer includes multiple sensor arrays deployed on key components of each wind turbine generator set, used to synchronously collect multi-physics state data; The edge intelligent processing layer is built into the data acquisition and edge computing units of each wind turbine generator set and is used to perform real-time preprocessing and feature extraction on the acquired raw data. The field-level digital twin analysis layer runs on a central server deployed in the wind farm control center and is used to drive the digital twin model for fault diagnosis and analysis based on the uploaded high-dimensional feature data. The proactive decision support layer, running on the central server, is used to predict lifespan and generate operation and maintenance decisions based on diagnostic results. The data acquisition and edge computing unit is interconnected with the central server through a field-level communication network.

[0014] Compared with the prior art, this disclosure has the following advantages: 1. By integrating multimodal sensor data with a high-fidelity digital twin model and combining it with an unsupervised anomaly detection mechanism based on residual analysis, the system can capture early fault characteristics that are difficult to detect using traditional single data sources, significantly reducing the false alarm rate and the false alarm rate, and improving the accuracy of fault identification and location.

[0015] 2. By deploying edge intelligent processing units on the wind turbine side, data preprocessing and feature extraction are achieved on-site, effectively reducing network transmission load and central server computing pressure, while ensuring the system's rapid response capability to single-machine failures and supporting collaborative diagnosis and maintenance of large-scale wind turbine clusters.

[0016] 3. By modeling the airflow coupling relationship of wind turbine clusters based on graph neural networks, it is possible to effectively distinguish between group anomalies caused by environmental factors such as wake effects and inherent equipment faults, and to assess the risk of fault propagation in the cluster, thereby improving the reliability and safety of the entire operation.

[0017] 4. By organically combining remaining lifetime prediction, multi-objective optimization decision-making and closed-loop execution mechanism, the system can dynamically generate and automatically execute the optimal operation and maintenance strategy based on the health status of the equipment, promote the transformation of operation and maintenance mode from passive response to proactive prediction and collaborative optimization, and significantly improve the operation efficiency and economy of wind farms.

[0018] Other features and advantages of this disclosure will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a fault diagnosis and early warning method for a wind power plant according to an embodiment of the present disclosure is shown. Figure 2A structural block diagram of a fault diagnosis and early warning system for a wind power plant according to an embodiment of the present disclosure is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0022] Figure 1 A flowchart illustrating a fault diagnosis and early warning method for a wind power plant according to an embodiment of the present disclosure is shown, such as... Figure 1 As shown in the figure, the fault diagnosis and early warning method for wind power plants according to this disclosure includes, S1, by deploying multiple sensor arrays on key components of each wind turbine in the target wind farm, setting the preset range of each sensor and collecting several initial data corresponding to the multi-physical field state in real time based on the preset sampling frequency. In this embodiment, by deploying various types of sensors on key components such as the gearbox, generator, main shaft, blades, and tower, the limitations of relying solely on SCADA data are overcome. These sensors continuously capture the equipment's status information under various operating conditions from multiple physical dimensions, including vibration, sound, deformation, temperature, and magnetic fields. This multimodal sensing approach can corroborate and complement each other, enabling the system to capture weaker and earlier signs of faults. For example, it can detect the initiation of tiny cracks through acoustic emission signals or detect early signs of wear through oil sensors, providing rich and high-quality data for accurate diagnosis.

[0023] In this embodiment of the disclosure, step S1 specifically includes: S11, deploy multiple types of sensors on key components of each wind turbine generator set to obtain multiple sensor arrays; S12, determine the range as the preset range based on the range of physical quantity changes of any fan from standby to rated power operation and under extreme conditions; S13, determine the potential fault characteristic frequency based on the historical fault data of any key component of the wind turbine, and use the wide-range sampling frequency determined based on the potential fault characteristic frequency as the preset sampling frequency; S14, control the multi-sensor array to synchronously collect data according to a preset sampling frequency to obtain a number of initial data; The key components include at least one of the following: gearbox, generator bearing housing, main shaft, blade root, blade surface, tower weld, tower base, and nacelle. The multiple types of sensors include at least two of the following: high-frequency acceleration sensor, acoustic emission sensor, strain gauge, inertial measurement unit, GPS displacement sensor, online oil monitoring sensor, magnetic flux sensor, and infrared thermal imager. The initial data includes at least two of the following: vibration signal, acoustic emission signal, structural strain signal, tower attitude signal, gearbox oil wear particle signal, generator air gap magnetic field signal, and infrared thermal imaging video stream.

[0024] In this embodiment of the disclosure, the multi-sensor array is used to synchronously acquire multi-physical field state data, including vibration signals, acoustic emission signals, structural strain signals, tower attitude signals, gearbox oil wear particle signals, generator air gap magnetic field signals, and infrared thermal imaging video streams; the sampling frequency of the sensors covers a wide range from low-frequency quasi-static to high-frequency transient impact, and its range design covers the entire range of physical quantity changes of the wind turbine from standby to rated power and extreme operating conditions.

[0025] Specifically, the initial data corresponding to the aforementioned multiphysics states are acquired through dedicated sensor arrays deployed on various key components: vibration signals are acquired by high-frequency accelerometers installed on the gearbox, main shaft, and generator bearing housing; acoustic emission signals are captured by acoustic emission sensors arranged in stress concentration areas such as blade roots and tower welds; structural strain signals are measured by strain gauges attached to the blade surface and tower base; tower attitude signals (such as tilting and swaying) are provided by inertial measurement units (IMUs) or GPS displacement sensors installed on the nacelle or tower top; gearbox oil wear particle signals are analyzed in real time by online oil monitoring sensors (such as inductive particle counters) to analyze metal debris in the lubricating oil; generator air gap magnetic field signals are detected using flux sensors embedded in the stator or rotor; and infrared thermal imaging video streams are continuously generated by infrared thermal imagers installed on the bottom of the nacelle or on inspection robots.

[0026] In this embodiment of the disclosure, determining the wide-range sampling frequency based on the potential fault characteristic frequencies of each component specifically includes: Collect and analyze historical fault data of any key component of a wind turbine in the target wind farm within a preset historical period. The historical fault data records specific fault modes and their corresponding characteristic frequencies. Based on the historical fault data, potential fault characteristic frequencies of each key component are identified and extracted, wherein: For gearboxes, potential fault characteristic frequencies include gear meshing frequencies and their sidebands; For spindle and generator bearings, potential failure characteristic frequencies include failure characteristic frequencies of the bearing inner ring, outer ring, rolling elements and cage; For blades, potential fault characteristic frequencies include the blade passage frequency and its harmonics; Based on the Nyquist sampling theorem, for each potential fault characteristic frequency, its sampling frequency is set to be at least twice that of the highest frequency component in the fault characteristic frequency, with a preset margin reserved, thereby comprehensively determining the wide range of sampling frequencies covering the potential fault characteristics of all key components.

[0027] In this embodiment, the "wide-range" sampling frequency and measurement range of the sensor are strictly determined based on the physical characteristics and operating conditions of potential faults in various components of the wind turbine. The sampling frequency range (e.g., from low-frequency quasi-static frequencies of several Hz to high-frequency transient frequencies of tens of kHz) must cover all key fault characteristic frequencies. For example, high-frequency components such as gear meshing frequencies, bearing fault characteristic frequencies, and blade passage frequencies must be fully captured, while low-frequency phenomena such as slow tower oscillations must also be effectively monitored. The measurement range design must cover the entire process of the wind turbine from standby, startup, rated power operation to emergency shutdown and encountering extreme wind conditions, ensuring that physical quantities (such as vibration acceleration and strain amplitude) do not exceed the sensor's measurement range under any operating condition, thus preventing signal distortion.

[0028] S2, perform data preprocessing and feature extraction on the initial data to obtain several high-dimensional feature data; In this embodiment of the disclosure, step S2 specifically includes: S21, noise suppression processing is performed on the initial data based on wavelet packet transform to obtain noise-suppressed data; S22, Based on the adaptive threshold algorithm, identify and replace abnormal outliers in the noise suppression data to obtain abnormal removal data; S23, perform time synchronization and alignment processing on the anomaly removal data from the multi-channel to obtain preprocessed data; S24, Based on historical data, the optimal feature set is selected using a feature selection algorithm; S25, extract feature data from the preprocessed data based on the optimal feature set to obtain the high-dimensional feature data.

[0029] In this embodiment, the preprocessing operations include noise suppression based on wavelet packet transform, outlier removal based on adaptive threshold, time synchronization alignment of multi-channel data, and preliminary extraction and compression of high-dimensional features. The processed feature data stream and compressed original data segments are then uploaded through the field-level communication network. First, the data is cleaned and standardized to filter out irrelevant noise interference and ensure that data from different sensors are precisely synchronized in time. Subsequently, the model extracts the feature information that best represents the device status from the original data, greatly compressing the amount of data that needs to be transmitted. This processing method not only significantly saves communication bandwidth and reduces the processing pressure on the central server, but more importantly, it provides an immediate and high-quality feature data stream for subsequent real-time diagnosis and analysis.

[0030] In this embodiment of the disclosure, noise suppression processing of the initial data based on wavelet packet transform specifically includes: The initial data is classified according to the type of sensor signal corresponding to it, and the format of the classified initial data is standardized to determine the source of noise. Based on the characteristics of the initial multimodal data, select wavelet basis functions suitable for preserving fault features, such as prioritizing the use of db series wavelets (e.g., db4, db6) or sym series wavelets (e.g., sym5). The number of wavelet packet decomposition layers is dynamically determined by combining the preset sampling frequency and noise frequency range of the initial data; The noise dominance of the detail coefficients and approximation coefficients obtained from the decomposition is determined and then processed. The processed noise-dominant coefficients are recombined with the unprocessed effective coefficients (including detail coefficients of fault features and noise-free approximation coefficients), and the same wavelet basis functions and decomposition levels as during decomposition are substituted. The inverse wavelet packet transform is then performed to reconstruct the denoised time-domain signal. Perform preliminary spectrum analysis (such as Fourier transform) on the denoised signal to confirm that the peak values ​​of key fault characteristic frequencies (such as gear meshing frequency and bearing fault frequency) have not been weakened (peak value reduction < 10%), and the energy proportion of the noise frequency band has significantly decreased (such as from 40% before denoising to below 10%), then the denoising process is complete.

[0031] In embodiments of this disclosure, the adaptive threshold algorithm includes, for example, the 3σ criterion.

[0032] In this embodiment, the high-dimensional feature data extracted from the original data refers to a set of feature parameters obtained by transforming the original time-domain waveform data into a higher-dimensional feature space that better reveals the essence of the fault through signal processing algorithms (such as wavelet packet transform, empirical mode decomposition, etc.). These features may include: time-domain features (such as root mean square value, peak value, kurtosis, waveform factor), frequency-domain features (such as spectral centroid, mean square frequency, energy proportion of a specific frequency band), and time-frequency-domain features (such as energy of each node of the wavelet packet, Hilbert spectral entropy), etc., with dimensions reaching tens or even hundreds of dimensions. The determination of these features is not arbitrary, but based on in-depth feature mining and analysis of a large amount of historical data and fault cases. Through feature selection algorithms such as principal component analysis (PCA) and maximum correlation minimum redundancy (mRMR), the optimal set of features with the strongest correlation to the degradation of equipment health status and the lowest redundancy is selected, and finally solidified into the lightweight model of the edge computing unit to achieve efficient and accurate feature extraction.

[0033] S3, based on several high-dimensional feature data, drive the field-level digital twin model to perform real-time simulation, obtain simulation prediction values, compare the simulation prediction values ​​with the actual sensing data, and combine with the unsupervised anomaly detection algorithm to determine the initial fault identification result and the initial fault severity. In this embodiment of the disclosure, step S3 specifically includes: S31, Construct a field-level digital twin model, wherein the field-level digital twin model integrates a wind turbine mechanism dynamics model based on physical laws and a deep neural network surrogate model trained based on historical data; S32, input the high-dimensional feature data and operating condition parameters of each wind turbine into the field-level digital twin model, drive the model to perform real-time simulation according to a preset time step, and output the simulation prediction value.

[0034] In this embodiment of the disclosure, the wind turbine mechanism dynamics model includes the gearbox transmission dynamics equation, the generator electromagnetic equation, the blade aeroelasticity equation, and the tower structure dynamics equation; The deep neural network proxy model is a CNN-LSTM hybrid network trained with historical health status data and corresponding working condition parameters; The mechanistic dynamics model and the deep neural network proxy model are integrated into a high-fidelity digital twin model of the wind turbine cluster using a weighted fusion algorithm.

[0035] In this embodiment of the disclosure, the simulation prediction values ​​include vibration amplitude prediction sequence, oil abrasive particle concentration prediction value, structural strain prediction distribution, and air gap magnetic field harmonic prediction component.

[0036] In this embodiment of the disclosure, the calculation of the residual comparison result specifically includes: The simulated predicted values ​​are compared with the preprocessed actual sensor data uploaded by the corresponding wind turbines point by point or dimension by dimension to generate a residual sequence. The residual sequence is subjected to moving average filtering to suppress instantaneous random fluctuations, and the steady-state residual signal characterizing the deviation of the system state is extracted to obtain residual comparison results for fault identification.

[0037] Specifically, this disclosure receives feature data streams from all wind turbines and drives a digital twin model for real-time simulation. By comparing the residuals between the predicted values ​​of the digital twin model and the actual sensor data, and combining an unsupervised anomaly detection algorithm under a transfer learning framework, it achieves sensitive identification and diagnosis of early, subtle faults and quantifies the severity of the fault. The digital model combines the advantages of a mechanistic model based on physical laws and a data-driven AI model, enabling high-fidelity simulation of the wind turbine's normal operating state under current conditions. The system continuously compares the predicted values ​​of the digital twin model with the actual sensor readings transmitted from the field; any significant and persistent deviation (residual) indicates the occurrence of an anomaly. Through advanced anomaly detection algorithms, the system analyzes these residuals, enabling it not only to determine "whether there is an anomaly," but also to accurately locate the source of the fault and assess its severity, achieving a leap from perception to diagnosis.

[0038] In this embodiment of the disclosure, step S3 further includes: S33, calculate the dimension-by-dimensional difference between the simulation prediction value and the actual sensing data to generate a residual sequence; S34, Perform moving average filtering on the residual sequence; S35, the processed residual sequence is input into an unsupervised anomaly detection algorithm based on the transfer learning framework. When the residual value is greater than the residual threshold, it is regarded as an abnormal residual sequence and an initial fault is determined. S36. When an initial fault exists, analyze the time domain, frequency domain and modal characteristics of the abnormal residual sequence, perform pattern matching with the preset component fault feature mapping library, and determine the fault source when the matching degree exceeds the preset threshold in order to determine the initial fault identification result. S37. Based on the statistical characteristic values ​​of the residual sequence and its trend characteristics over time, combined with the preset fault severity assessment model, the initial fault severity level is quantified.

[0039] In this embodiment of the disclosure, the residual threshold is a dynamic threshold range set based on historical health data.

[0040] In this embodiment, the core of the precise fault location mechanism lies in the high fidelity and refinement of the digital twin model. This model is not a black box, but rather organically coupled with sub-models corresponding to various subsystems and key components of the wind turbine (such as gearboxes, generators, blades, and main shafts). When the system detects a significant residual signal, it further analyzes the specific characteristics of the residual in the time domain, frequency domain, and modal domain, and performs pattern matching with the simulation results of different component faults in the twin model. For example, if a high-frequency vibration component significantly matches the meshing frequency of a specific gear in the gearbox appears in the residual sequence, the system can be highly certain that the fault source is located on that pair of gears in the gearbox. Similarly, if the residual is mainly reflected in the air gap magnetic field harmonics of the generator, the fault can be located to the generator. This combination of residual feature pattern recognition and reverse tracing based on component sub-models achieves precise location from "system anomaly" to "component-level fault."

[0041] In this embodiment, the severity assessment of the fault is based on the magnitude, persistence, development trend, and quantitative analysis of the residual's impact on system performance. First, the residual's amplitude directly reflects the deviation between the actual state and the ideal healthy state, making it the most direct indicator of severity. The system calculates the statistical characteristics of the residual (such as root mean square value, peak value, kurtosis, etc.) and compares them with multi-level warning thresholds preset based on historical data and expert experience. Second, the temporal evolution trend of the residual is crucial: a residual with a small amplitude but continuously increasing linearly may indicate a more pressing fault risk than a residual with a large amplitude but stable value, as it indicates that the defect is actively expanding. Finally, the system integrates all the residual characteristics and combines them with simulations of physical models (such as crack propagation models and wear models) to map the abstract residual sequence to the degradation of one or more key performance indicators (KPIs) (e.g., percentage decrease in transmission efficiency, vibration intensity level), thereby providing an intuitive and quantitative graded assessment of the fault severity and offering precise basis for subsequent operation and maintenance decisions.

[0042] S4, perform fault diagnosis by fusing multimodal high-dimensional feature data, and calibrate the initial fault identification result; In this embodiment of the disclosure, step S4 specifically includes: S41, Analyze the high-dimensional feature data corresponding to any sensor, and identify abnormal patterns based on the analysis results; S42, After the abnormal pattern is identified, the abnormal state of other sensor data within the same period is analyzed in relation to construct a multimodal evidence chain; S43, based on a multimodal evidence chain, uses evidence theory algorithms to fuse multi-source information and calculate the comprehensive diagnostic confidence. S44, when the comprehensive diagnostic confidence exceeds the preset warning threshold, a calibrated fault diagnosis result is generated, and a warning signal of the corresponding level is triggered.

[0043] In this embodiment of the disclosure, taking the acoustic emission signal as an example to determine crack initiation, the anomaly is usually not reflected in a simple exceedance of the signal amplitude, but rather in a significant increase in the frequency of sudden high-energy acoustic emission events, an energy distribution of events crossing a specific threshold, or a sustained abnormal surge in the energy of specific frequency bands of the acoustic emission signal (such as the energy of detail coefficients obtained through wavelet packet decomposition). These characteristics indicate that rapid energy release is occurring inside the material, which is a typical sign of microcrack propagation.

[0044] However, a single acoustic emission signal anomaly can be affected by environmental noise or electromagnetic interference. Therefore, it is necessary to fuse it with other signals to greatly improve accuracy. For example, if an acoustic emission signal shows a sudden event resembling a crack, and simultaneously the high-frequency (above 5kHz) spectral energy of the vibration signal also increases synchronously (possibly due to the impact caused by the opening and closing of the crack), and infrared thermal imaging shows a slight local temperature rise in that area (due to frictional heat generated by the crack), then the multimodal evidence chain can form a strong mutual corroboration, thereby raising the diagnostic confidence of crack initiation to a level that can trigger an early warning. This cross-modal correlation analysis is the core of the initial judgment of edge-layer intelligence.

[0045] In this embodiment of the disclosure, the warning threshold is dynamically set based on the wind turbine's historical operating data, fault case library, and real-time operating conditions.

[0046] S5. Based on the severity of the initial fault and its evolution trend corresponding to any faulty wind turbine, predict the remaining useful life of its key components. Based on the prediction results of the remaining useful life, use a multi-objective optimization algorithm combined with multi-objective parameters to generate collaborative operation and maintenance decisions for the target wind farm. In this embodiment of the disclosure, step S5 specifically includes: S51, obtain historical health data, current fault severity quantification value and evolution trend of key components of the faulty wind turbine; S52 uses a sequence deep learning model that integrates physical degradation models for lifetime prediction; S53, output the probability distribution of the remaining useful life of the key component, including the mean life and confidence interval.

[0047] In this embodiment of the disclosure, a sequence deep learning model combined with a physical degradation model is used to predict the probability distribution of the remaining service life of key components based on the current severity of the fault and its evolution trend.

[0048] In this embodiment of the disclosure, the acquisition of historical health data, current fault severity quantification, and evolution trend of key components of the faulty wind turbine specifically includes: Acquire complete health status time-series data of key components of the faulty wind turbine within a preset historical period, including but not limited to vibration characteristic values, temperature readings, oil abrasive concentration, and structural strain measurements. Collect quantitative indicators of the current fault, including the fault severity level, the magnitude of deviation of characteristic parameters from the normal baseline, and the slope of their recent evolution trend; Multidimensional feature vectors strongly correlated with component degradation are extracted to form the input dataset for lifetime prediction.

[0049] In this embodiment of the disclosure, a hybrid prediction model integrating physical degradation mechanisms and data-driven algorithms is established, wherein: Physical degradation models are established based on the fatigue characteristics, crack propagation laws, or wear mechanisms of component materials, such as using Paris's law to describe crack propagation or Arcard's formula to quantify the wear process. Sequence deep learning models employ Transformer networks or bidirectional LSTM networks with encoder-decoder architectures to learn long-term dependencies and degradation trajectory patterns in health status data. The hybrid prediction model is trained end-to-end using historical complete lifespan data to optimize model parameters.

[0050] In this embodiment of the disclosure, outputting the probability distribution of the remaining service life of the key component specifically includes: Input the current health status feature vector into the trained prediction model; The model generates the evolution trajectory of the future health status of components through multi-step forward recursive prediction; When the predicted health indicator first exceeds the preset failure threshold, record that time point as the predicted failure time. Calculate the time interval from the current moment to the predicted failure moment, and use it as the baseline value for the remaining useful life; The uncertainty of the prediction process is assessed using Monte Carlo simulation or Bootstrap sampling methods to generate a probability distribution of the remaining useful life.

[0051] In this embodiment of the disclosure, step S5 further includes: S54, establish a multi-objective optimization function with the objectives of minimizing the total power generation loss and minimizing the operation and maintenance cost; S55 sets constraints including health constraints, environmental constraints, scheduling constraints, and resource constraints; S56, A multi-objective optimization algorithm is used to solve the optimization function under constraints to generate a Pareto optimal solution set; S57. Select the final solution from the set of optimal solutions based on the wind farm operation priority, and output a collaborative operation and maintenance decision including a list of wind turbines to be inspected, an inspection time window, a spare parts allocation plan, and a maintenance team assignment plan.

[0052] In this embodiment, based on the health status prediction results of all wind turbines, weather forecast information, grid dispatch instructions, and spare parts inventory information, the optimal collaborative maintenance plan and resource scheduling scheme are dynamically generated using the minimization of overall power generation loss and the minimization of operation and maintenance costs as multi-objective optimization functions. After accurate diagnosis and quantification of faults, the remaining useful life prediction module predicts the future health trend of components based on the current state and historical degradation trajectory, providing a probabilistic lifespan estimate. Subsequently, the operation and maintenance decision optimization engine comprehensively considers the health status prediction of all wind turbines, future weather windows, grid dispatch requirements, and existing spare parts and human resources, performing multi-objective optimization calculations with the goal of maximizing power generation revenue and minimizing operation and maintenance costs. The final output is a globally optimal, dynamically adjusted maintenance plan, guiding the operation and maintenance team to take the right maintenance actions at the right time for the right wind turbines.

[0053] S6 executes the collaborative operation and maintenance decision through a closed-loop feedback execution mechanism and triggers corresponding control and scheduling instructions.

[0054] In this embodiment, the closed-loop feedback execution mechanism includes the automatic distribution of operation and maintenance decision-making plans to the wind farm monitoring system and asset management system. This triggers control commands to limit the power of specific wind turbines or provide shutdown protection, and automatically generates work orders, allocates spare parts, and connects maintenance teams, forming a complete closed loop from status perception, intelligent diagnosis, predictive warning, to decision execution. This mechanism ensures that the system's decisions are automatically implemented, forming a closed loop. Once an optimized operation and maintenance plan is generated, the system automatically converts it into executable commands. These commands can be directly distributed to the wind farm control system to automatically and flexibly adjust the power of wind turbines that have received warnings or to safely shut them down, preventing the fault from escalating. Simultaneously, the system automatically connects to the asset management system, creates maintenance work orders, notifies relevant personnel, and triggers spare parts procurement or allocation processes.

[0055] In this embodiment of the disclosure, the method further includes using graph neural networks to perform field-level spatiotemporal correlation analysis to distinguish between inherent faults of a single unit and group abnormal faults caused by environmental factors, and to assess the risk of fault propagation in the wind turbine cluster.

[0056] In this embodiment, graph neural network technology is used to model the layout and airflow field of the wind turbine cluster, analyze the state coupling relationship between wind turbines caused by wake effects and gust propagation, distinguish between inherent faults of individual turbines and group anomalies caused by environmental factors, and predict the propagation risk of specific faults in the wind turbine cluster. Graph neural networks and other technologies are used to model the geographical layout of the wind farm, wind direction and speed, and the mutual aerodynamic influence between wind turbines. When multiple wind turbines exhibit similar anomalies, this module can analyze and determine whether this is caused by common external environmental factors (such as consistent turbulence or wake effects), thereby avoiding false alarms of group equipment failures. Simultaneously, it can also predict the potential risk of a specific fault (such as vibration at a specific frequency caused by blade imbalance) propagating downwind turbines through aerodynamic coupling, achieving true farm-level collaborative early warning.

[0057] In this embodiment of the disclosure, assessing the risk of fault propagation in a wind turbine cluster specifically includes: A graph structure model of a wind turbine cluster is constructed, with each wind turbine generator as a node and the coupling relationship between the generators due to the interaction of the wind flow field as an edge, to establish a graph structure that reflects the spatiotemporal relationship between the wind turbines. The operating status characteristics, spatiotemporal location information and environmental parameters of each wind turbine are input into the graph neural network model, and the state propagation law in the wind turbine cluster is learned through graph convolution operation. Based on the learned state propagation rules, the distribution pattern of abnormal states in the wind turbine cluster is identified, and the inherent faults of individual units are distinguished from group anomalies caused by environmental factors. For the identified faulty wind turbines, the propagation path and propagation risk level of the fault in the wind turbine cluster are evaluated based on their positional relationship and edge weights in the graph structure. The output includes a field-level spatiotemporal correlation analysis report that includes fault type differentiation results and propagation risk assessment.

[0058] Based on the above method, this disclosure also provides a fault diagnosis and early warning system for wind power plants corresponding to the above method. Figure 2 A structural block diagram of a fault diagnosis and early warning system for a wind power plant according to an embodiment of the present disclosure is shown, such as... Figure 2 As shown, the system includes: The multimodal data sensing layer includes multiple sensor arrays deployed on key components of each wind turbine generator set, used to synchronously collect multi-physics state data; The edge intelligent processing layer is built into the data acquisition and edge computing units of each wind turbine generator set and is used to perform real-time preprocessing and feature extraction on the acquired raw data. The field-level digital twin analysis layer runs on a central server deployed in the wind farm control center and is used to drive the digital twin model for fault diagnosis and analysis based on the uploaded high-dimensional feature data. The proactive decision support layer, running on the central server, is used to predict lifespan and generate operation and maintenance decisions based on diagnostic results. The data acquisition and edge computing unit is interconnected with the central server through a field-level communication network. The edge intelligent processing layer is configured to perform the following operations: noise suppression based on wavelet packet transform, outlier removal based on an adaptive threshold algorithm, time synchronization alignment of multi-channel data, and extraction of high-dimensional features from preprocessed data. The field-level digital twin analysis layer includes a high-fidelity wind turbine cluster digital twin model and an unsupervised anomaly detection module, which is configured to identify faults and quantify their severity by comparing the residuals of the simulation predictions output by the digital twin model with the actual sensing data. The proactive decision support layer includes a remaining useful life prediction module and an operation and maintenance decision optimization engine, which are configured to predict the remaining life of components based on the severity of the fault and generate a collaborative maintenance plan based on a multi-objective optimization algorithm. The system also includes a closed-loop feedback execution mechanism, which automatically distributes the operation and maintenance decision plan to the wind farm monitoring system and asset management system for execution.

[0059] Based on the same inventive concept as the above disclosure, this disclosure also provides an electronic device. The electronic device of this disclosure includes at least one processor and at least one memory electrically connected to the processor. The memory is electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0060] It should be noted that the electrical connection between the above-mentioned units does not necessarily mean the connection between lines. The indirect connection method can be applied to the embodiments of this disclosure as long as it achieves the purpose of this disclosure.

[0061] Based on the same inventive concept, this disclosure also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the above method.

[0062] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for fault diagnosis and early warning in wind power plants, characterized in that, The method includes, By deploying multiple sensor arrays on key components of each wind turbine in the target wind farm, setting the preset range for each sensor and collecting several initial data corresponding to the multi-physical field state in real time based on the preset sampling frequency. Data preprocessing and feature extraction are performed on several initial data sets to obtain several high-dimensional feature data sets; Real-time simulation is performed based on a field-level digital twin model driven by several high-dimensional feature data to obtain simulation prediction values. The simulation prediction values ​​are compared with the actual sensing data, and combined with an unsupervised anomaly detection algorithm, the initial fault identification results and the initial fault severity are determined. Fault diagnosis is performed by fusing multimodal high-dimensional feature data to calibrate the initial fault identification results; Based on the initial fault severity and evolution trend of any faulty wind turbine, the remaining useful life of its key components is predicted. Based on the remaining useful life prediction results, a multi-objective optimization algorithm combined with multi-objective parameters is used to generate collaborative operation and maintenance decisions for the target wind farm. The collaborative operation and maintenance decisions are executed through a closed-loop feedback execution mechanism, which triggers corresponding control and scheduling instructions.

2. The method according to claim 1, characterized in that, Based on a preset sampling frequency and preset range, several initial data corresponding to the state of multiple physics fields are collected in real time, specifically including: Multiple types of sensors are deployed on key components of each wind turbine to obtain multiple sensor arrays; The range is determined as the preset range based on the range of physical quantity changes of any wind turbine from standby to rated power operation and under extreme conditions. The potential fault characteristic frequency is determined based on the historical fault data of any key component of the wind turbine, and the sampling frequency range is determined based on the potential fault characteristic frequency as the preset sampling frequency. The array of multiple sensors is controlled to synchronously acquire data according to a preset sampling frequency to obtain several initial data. The key components include at least one of the following: gearbox, generator bearing housing, main shaft, blade root, blade surface, tower weld, tower base, and nacelle. The multiple types of sensors include at least two of the following: high-frequency acceleration sensor, acoustic emission sensor, strain gauge, inertial measurement unit, GPS displacement sensor, online oil monitoring sensor, magnetic flux sensor, and infrared thermal imager. The initial data includes at least two of the following: vibration signal, acoustic emission signal, structural strain signal, tower attitude signal, gearbox oil wear particle signal, generator air gap magnetic field signal, and infrared thermal imaging video stream.

3. The method according to claim 2, characterized in that, Data preprocessing and feature extraction are performed on several sets of initial data, specifically including: The initial data is subjected to noise suppression processing based on wavelet packet transform to obtain noise-suppressed data. Anomaly removal data is obtained by identifying and replacing abnormal outliers in the noise suppression data based on an adaptive threshold algorithm. The anomaly removal data from the multi-channel system is time-synchronized and aligned to obtain preprocessed data. The optimal feature set is selected based on historical data using a feature selection algorithm. Based on the optimal feature set, feature data is extracted from the preprocessed data to obtain the high-dimensional feature data.

4. The method according to claim 3, characterized in that, Real-time simulation is performed based on a field-level digital twin model driven by several of the aforementioned high-dimensional feature data, specifically including: A field-level digital twin model is constructed, which integrates a wind turbine mechanism dynamics model based on physical laws and a deep neural network surrogate model trained based on historical data. The high-dimensional feature data and operating parameters of each wind turbine are input into the field-level digital twin model, driving the model to perform real-time simulation according to a preset time step, and output simulation prediction values.

5. The method according to claim 4, characterized in that, Determining the initial fault identification results and the initial fault severity includes: The simulation prediction values ​​and the actual sensing data are calculated dimension by dimension to generate a residual sequence; The residual sequence is subjected to moving average filtering. The processed residual sequence is input into an unsupervised anomaly detection algorithm based on a transfer learning framework. When the residual value is greater than the residual threshold, it is regarded as an abnormal residual sequence and an initial fault is determined to exist. When an initial fault exists, the time domain, frequency domain, and modal characteristics of the abnormal residual sequence are analyzed and matched with a preset component fault feature mapping library. When the matching degree exceeds a preset threshold, the fault source is determined to determine the initial fault identification result. Based on the statistical characteristic values ​​of the residual sequence and its trend characteristics over time, combined with the preset fault severity assessment model, the initial fault severity level is quantified.

6. The method according to claim 5, characterized in that, Fault diagnosis is performed through the fusion of multimodal high-dimensional feature data, specifically including: Analyze the high-dimensional feature data corresponding to any sensor, and identify abnormal patterns based on the analysis results; Once the abnormal pattern is identified, the abnormal states of other sensor data within the same period are analyzed in conjunction with it to construct a multimodal chain of evidence. Based on a multimodal chain of evidence, multi-source information is fused through evidence theory algorithms to calculate the comprehensive diagnostic confidence. When the overall diagnostic confidence level exceeds the preset warning threshold, a calibrated fault diagnosis result is generated, and a warning signal of the corresponding level is triggered.

7. The method according to claim 6, characterized in that, Predicting the remaining service life of its key components, specifically including: Obtain historical health data, current fault severity quantification, and evolution trend of key components of the faulty wind turbine; A sequence deep learning model that integrates physical degradation models is used for lifetime prediction; Output the probability distribution of the remaining useful life of the key component, including the mean life and confidence interval.

8. The method according to claim 7, characterized in that, The collaborative operation and maintenance decision-making for the target wind farm is generated by combining multi-objective optimization algorithms with multi-objective parameters, specifically including: Establish a multi-objective optimization function with the objectives of minimizing total power generation loss and minimizing operation and maintenance costs; Set constraints including health constraints, environmental constraints, scheduling constraints, and resource constraints; A multi-objective optimization algorithm is used to solve the optimization function under constraints to generate a Pareto optimal solution set. Based on the wind farm's operational priority, the final solution is selected from the optimal solution set, and the output is a collaborative operation and maintenance decision that includes a list of wind turbines to be inspected, an inspection time window, a spare parts allocation plan, and a maintenance team assignment plan.

9. The method according to claim 8, characterized in that, It also includes using graph neural networks to perform field-level spatiotemporal correlation analysis to distinguish between inherent faults of individual units and group abnormal faults caused by environmental factors, and to assess the risk of fault propagation in wind turbine clusters.

10. A system based on the fault diagnosis and early warning method for wind power plants according to any one of claims 1-9, characterized in that, The system includes: The multimodal data sensing layer includes multiple sensor arrays deployed on key components of each wind turbine generator set, used to synchronously collect multi-physics state data; The edge intelligent processing layer is built into the data acquisition and edge computing units deployed in each wind turbine generator set, and is used to perform real-time preprocessing and feature extraction on the acquired raw data. The field-level digital twin analysis layer runs on a central server deployed in the wind farm control center and is used to drive the digital twin model for fault diagnosis and analysis based on the uploaded high-dimensional feature data. The proactive decision support layer, running on the central server, is used to predict lifespan and generate operation and maintenance decisions based on diagnostic results. The data acquisition and edge computing unit is interconnected with the central server through a field-level communication network.

Images