An online condition monitoring system for a wind turbine generator

Abstract

The present application relates to the technical field of wind turbine state monitoring, in particular to a wind turbine online state monitoring system. According to its function and data level, it is divided from bottom to top into: physical sensing layer, data acquisition and transmission layer, data platform layer, algorithm model layer; the physical sensing layer senses the physical state quantity analog signal data of the wind turbine operation state; the data acquisition and transmission layer converts the analog signal into digital signal through data acquisition hardware; the data platform layer preprocesses data; the algorithm model layer analyzes and warns the operation state and fault of the wind turbine; based on the present application, abnormal operation characteristics can be found through analysis of real-time operation state data, abnormal information can be captured at the budding period of fault, a pre-stage window is created for maintenance and diagnosis of the wind turbine, and occurrence of catastrophic failure and secondary damage can be effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of wind turbine condition monitoring technology, specifically to an online condition monitoring system for wind turbines. Background Technology

[0002] With the continuous growth of wind power installed capacity, higher requirements are being placed on the operational reliability, maintenance efficiency, and life-cycle cost control of wind turbine generators (hereinafter referred to as "wind turbines"). However, wind turbines are usually installed in harsh environments such as remote areas, high altitudes, or offshore, and are subjected to randomly alternating wind loads, extreme temperature differences, and salt spray corrosion. Their core components, such as gearboxes, generators, main bearings, and blades, are highly susceptible to failure. The traditional "periodic maintenance" model carries the risk of over-maintenance or neglect, and "post-failure repair" results in huge downtime losses and high spare parts costs. Summary of the Invention

[0003] The purpose of this invention is to establish an online condition monitoring system with condition-based maintenance as its core. By real-time sensing, accurate diagnosis, and trend prediction of the wind turbine's operating status, and by analyzing the deviation between the wind turbine's operating status and its normal operating status, the system can issue early warnings before potential faults occur, providing a technical means to establish an active early warning and intervention mechanism and avoid post-event maintenance.

[0004] To achieve the above objectives, the present invention adopts the following technical solution.

[0005] An online condition monitoring system for wind turbine generator sets includes an algorithm model layer. The algorithm model layer establishes a condition warning model to describe the wind turbine generator set during normal operation, and analyzes and warns of the operating status and faults of the wind turbine generator set by the degree of abnormal deviation between the operating process and the normal operation. The algorithm model layer specifically includes a data extraction unit, a model library, and a dynamic monitoring and early warning unit.

[0006] The data extraction unit establishes a normal state database based on historical data acquired during the healthy and fault-free operation cycle of the wind turbine generator set by the wind turbine generator set condition monitoring hardware, and allocates the normal state database to the training set, validation set, and test set; the historical data includes vibration, speed, and temperature data;

[0007] The model library establishes normal behavior models based on historical data and trains the models using a normal state database; the normal behavior model includes:

[0008] a. A support vector data description model for vibration and rotational speed. The goal of the support vector data description model is to find the smallest volume three-dimensional bounding volume in the high-dimensional space composed of the feature values ​​of vibration and rotational speed, so that it contains as many vibration and rotational speed feature samples as possible under normal conditions, and to determine the optimal state feature set through the surface of the three-dimensional bounding volume.

[0009] b. For temperature change data, the TCN-BiLSTM-Attention model performs correlation analysis on temperature data and temperature-related operating condition data, selects a set of feature variables that are strongly correlated with temperature data, and identifies anomalies and judges the degree of temperature anomalies by the residual change trend of real-time temperature monitoring results and temperature prediction output.

[0010] The dynamic monitoring and early warning unit establishes a model for normal and abnormal states, and outputs an early warning threshold setting and dynamic adjustment process, including:

[0011] a. Initial threshold setting based on offline data: Input historical health data into the trained model, calculate the reconstruction error between vibration and rotation speed samples and the surface distance of the 3D bounding body, as well as temperature change data samples and the predictions of the TCN-BiLSTM-Attention model; use statistical methods to set the initial threshold as follows: ,in and These are the mean and standard deviation of the difference sequence of samples under normal conditions. This is the sensitivity coefficient;

[0012] b. Adaptive dynamic threshold adjustment based on online data: Construct a dual sliding window mechanism, defining a fast observation window and a slow baseline window; when the system is running online, the data stream enters the observation window, and the dynamic threshold is adjusted accordingly. Determined by the statistical characteristics of the data within the baseline window. ;in, and These are the mean and standard deviation of the difference series within the current reference window. It is an adjustment factor for controlling sensitivity; when the average level of the difference in the observation window continues to be higher than the baseline window, the baseline window is gradually slid forward, and the statistics are updated with new normal state data, thereby realizing the self-learning and self-adaptation of the threshold to slow normal drift, avoiding false alarms. When the difference suddenly and drastically exceeds the dynamic threshold, the wind turbine state abnormality alarm is triggered.

[0013] Further improvements or preferred implementations of the aforementioned online status monitoring system for wind turbine generators can be divided from bottom to top according to their functions and data levels as follows: physical sensing layer, data acquisition and transmission layer, and data platform layer.

[0014] The physical sensing layer is responsible for sensing the physical state quantity analog signal data of the wind turbine's operating status through sensors deployed on various key structures of the wind turbine that are to be monitored and analyzed.

[0015] The data acquisition and transmission layer uses data acquisition hardware and network communication protocols to convert the analog signals of the physical sensing layer into digital signal data that the system can directly process and use, and establishes a communication network for transmission to local or remote data centers.

[0016] The data platform layer establishes a preprocessing flow for digital signal data uploaded from the data acquisition and transmission layer, and stores it in a database with a data structure that matches and adapts to the source data. Based on the preset wind turbine structure and equipment and the measurement point coding rules, the labeling of the digital signal data is completed.

[0017] A further improvement or preferred implementation of the aforementioned online condition monitoring system for wind turbine generators, wherein the TCN-BiLSTM-Attention model includes a front-end multi-scale local spatial feature extractor, a BiLSTM module, and a multi-head attention module;

[0018] The front-end local feature extractor consists of three parts: causal convolution, dilated convolution, and residual connections. Causal convolution performs weighted calculations on the historical input data up to the current time step. Dilated convolution introduces a dilation rate to allow the convolution kernel to perform skip sampling of the input features. Residual connections directly pass the input features of the TCN module across layers to the output of the module, and add them element-wise with the features extracted by the convolutional layer. The front-end multi-scale local spatial feature extractor transforms the high-dimensional raw monitoring data into a set of high-dimensional feature sequences containing sufficient local detail information, thereby obtaining multi-scale features of the monitoring data spatial dimension.

[0019] The BiLSTM module consists of two unidirectional LSTM networks with completely independent structures but opposite processing directions. At each time step, the unidirectional LSTM network outputs its own hidden state. Finally, the hidden state vectors of the two directions are concatenated according to the feature dimension to output bidirectional temporal features. The BiLSTM module performs bidirectional temporal feature extraction on the multi-scale local features extracted by the front-end multi-scale local spatial feature extractor to obtain the feature hidden relationship at mid-to-long-distance time steps.

[0020] The multi-head attention module learns a feature weight allocation strategy through training. It strengthens or weakens the weight of each time step feature or each feature dimension in the bidirectional time series features output by the BiLSTM layer, and assigns higher weights to key time step features that are strongly correlated with the temperature data changes of the wind turbine, thereby obtaining a subset of core features.

[0021] In a further improvement or preferred embodiment of the aforementioned online condition monitoring system for wind turbine generators, the support vector data description model consists of a feature extraction and mapping layer, a high-dimensional decision boundary construction layer, and a three-dimensional bounding volume generation and output layer.

[0022] The feature extraction and mapping layer extracts features by splicing and fusing multidimensional inputs through a one-dimensional convolutional neural network or a lightweight autoencoder. It uses multi-layer feature extraction and nonlinear mapping operations to transform the linearly inseparable unit operation data in the original low-dimensional space into approximately linearly separable features in the high-dimensional feature space.

[0023] The high-dimensional decision boundary construction layer maps feature samples through radial basis functions, mapping the original features obtained by the feature extraction mapping layer from the low-dimensional feature space to the high-dimensional regenerating kernel Hilbert space. It constructs the objective function that minimizes the hypersphere and the constraint condition that most normal sample points are contained inside or on the surface of the hypersphere. Through quadratic programming constraint optimization, it solves for the high-dimensional feature vector composed of normal sample points located on the surface of the minimum hypersphere. Then, it reduces the high-dimensional feature vector to three dimensions through a linear dimensionality reduction algorithm. Based on the coordinate distribution of all normal samples in the three-dimensional space after dimensionality reduction, it calculates a minimum volume bounding body that contains all or most of the normal samples and outputs the complete surface feature set of the minimum volume bounding body. The newly input real-time data points are then mapped to three-dimensional space after the same feature extraction, dimensionality reduction and spatial transformation processing to obtain the distance between the input data and the center of the bounding body.

[0024] A further improvement or preferred embodiment of the aforementioned online condition monitoring system for wind turbine generators involves establishing a multi-channel synchronous data acquisition mechanism in the data acquisition and transmission layer. This mechanism employs a PXI or CompactDAQ-based data acquisition and transmission module to collect data at different frequencies based on the characteristics of the monitoring data from multiple sensors within the physical sensing layer. Specifically:

[0025] For vibration and speed signals, the set acquisition frequency range covers twice the expected maximum frequency of the wind turbine; for temperature signals, a frequency of 5~80Hz is used or the average temperature value of a fixed period is obtained according to the statistical averaging unit built into the module.

[0026] The data acquisition and transmission layer establishes a Modbus TCP network between sensors and data acquisition and transmission modules located at wind turbine towers and other locations, as well as an OPC UA fiber optic network between the wind turbines and the data platform layer and algorithm model layer located in the regional monitoring center. This OPC UA fiber optic network, through the establishment of bidirectional handshake, disconnection reconnection, data buffering, and security encryption and authentication mechanisms, ensures and controls the effectiveness and security of data transmission between layers while transmitting basic sensor monitoring data values. Simultaneously, it controls and transmits semantic data such as data source stamps, data quality stamps, timestamps, units, ranges, and engineering units established by the data transmission layer or data platform layer through a data protocol library.

[0027] For further improvements or preferred implementations of the aforementioned online condition monitoring system for wind turbine generators, the data platform layer needs to complete the centralized storage, cleaning, management, and governance of this data; specifically including:

[0028] a. Data preprocessing, including outlier removal, missing value handling, data smoothing and denoising, and data normalization and standardization;

[0029] b. Data storage and indexing: Establish a hybrid storage architecture that combines time-series databases and relational databases. Use time-series databases to store and maintain high-frequency, time-series vibration, speed, and temperature raw and characteristic data. Use relational databases to store wind turbine basic information, measurement point configuration, rule base, and static or structured relationships of maintenance records, and establish a unified asset hierarchy and equipment measurement point coding system.

[0030] In a further improvement or preferred embodiment of the aforementioned online condition monitoring system for wind turbine generators, the model library further includes a feature extraction unit, which is used to extract features from vibration, rotational speed and temperature signals, and to complete feature selection and dimensionality reduction processing.

[0031] The vibration and rotation speed signal analysis and feature extraction process is divided into time domain feature extraction, frequency domain feature extraction and time-frequency domain feature extraction. The vibration and rotation speed features are formed by time domain features, frequency domain features and time-frequency domain features.

[0032] The time-domain features are directly calculated and statistically derived from the preprocessed time-domain waveform, including dimensional and dimensionless indices for vibration and rotational speed. Dimensional indices for vibration and rotational speed include peak value, peak-to-peak value, mean, RMS value, variance, standard deviation, and root mean square value. Dimensionless indices include peak factor, kurtosis, skewness, waveform factor, impulse factor, and margin factor. The frequency-domain features are obtained by converting the time-domain signal to the frequency domain using Fast Fourier Transform (FFT) and performing spectrogram analysis, envelope spectrum analysis, power spectral density analysis, and cepstral analysis. The time-frequency domain features are obtained through wavelet packet decomposition of energy entropy and parameter-adaptive eigenmode decomposition.

[0033] The feature selection and dimensionality reduction process includes: calculating the correlation between each feature and the device status label, and removing features with low correlation; using the feature importance assessment capability of the machine learning model itself, and combining the average impact value algorithm to set an importance threshold, only retaining features with scores exceeding the threshold to form the final feature set.

[0034] In a further improvement or preferred embodiment of the aforementioned online condition monitoring system for wind turbine generators, the preprocessing further includes: for the associated vibration and speed sensor data, adaptively decomposing the signal into a series of intrinsic mode functions through complete set empirical mode decomposition, separating mode aliasing, and then using variational mode decomposition to separate signal components with similar frequency components, thereby jointly filtering out background noise and electromagnetic interference.

[0035] In a further improvement or preferred embodiment of the aforementioned online condition monitoring system for wind turbine generator sets, the physical sensing layer includes at least: an integrated circuit piezoelectric accelerometer for monitoring the vibration state of the transmission system; a precise speed sensor for acquiring the low-speed end of the main shaft and the high-speed end of the generator, which may employ sensors such as eddy current or Hall effect proximity switches; a temperature sensor for monitoring the gearbox oil sump, generator windings, and main bearing area; and an oil sensor for monitoring the health status parameters of the wind turbine generator set gearbox.

[0036] The operating data includes wind speed, ambient temperature, current power, and oil temperature at the previous moment.

[0037] Its beneficial effects are as follows:

[0038] This application's online condition monitoring system for wind turbine generators, based on existing technologies, combines sensing technology, industrial IoT communication, big data processing, and artificial intelligence to establish and improve a closed-loop management system for key components of wind turbine generators, encompassing the entire process from condition perception and feature extraction to intelligent early warning, thereby enhancing the intelligent operation and maintenance level of wind farms. Based on this application, the system can detect abnormal operating characteristics through analysis of real-time operating status data, capturing abnormal information in the early stages of faults, creating an early window for wind turbine generator maintenance and diagnosis, and effectively preventing catastrophic failures and secondary damage. Attached Figure Description

[0039] Figure 1 This is a functional schematic diagram of an online condition monitoring system for wind turbine generator sets.

[0040] Figure 2 This is a flowchart of data processing and feature extraction from vibration and speed sensor source data sources. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0042] The wind turbine generator online condition monitoring system disclosed in this application is an online condition monitoring system with condition-based maintenance as its core. By sensing the operating status of the wind turbine generator in real time, it uses the deviation between the detected status and the normal operating status to provide early warning monitoring of abnormal operating status, so as to arrange maintenance plans and realize the transformation from passive operation and maintenance to proactive early warning and intervention.

[0043] like Figure 1 As shown, the wind turbine generator online condition monitoring system of this application is based on the Internet of Things layered architecture design. According to its functions and data levels, it can be divided into the following layers from bottom to top: physical sensing layer, data acquisition and transmission layer, data platform layer, and algorithm model layer.

[0044] The physical sensing layer is responsible for sensing the physical state parameters of the wind turbine through sensors deployed on key structures of the wind turbine. The data acquisition and transmission layer converts the analog signals from the physical sensing layer into digital signals that the system can directly process and use, and establishes a communication network for transmission to local or remote data centers. The data platform layer establishes a preprocessing flow for the digital signal data uploaded by the data acquisition and transmission layer, and stores it in a database with a data structure that matches and adapts to the source data. Based on the preset wind turbine structure, equipment, and measurement point coding rules, it completes the labeling of the digital signal data. The algorithm model layer stores, runs, and maintains the online status prediction model of the wind turbine, establishes an algorithm model workflow for fault early warning and diagnosis, and completes the analysis and early warning of the online status of the wind turbine. Through the collaborative work of the aforementioned functional layers, the realization loop from basic data sensing of the wind turbine to online status analysis and early warning is completed.

[0045] To acquire the necessary physical state parameters of the wind turbine, the physical sensing layer in this application includes at least: an integrated circuit piezoelectric accelerometer for monitoring the vibration state of the transmission system (including the wind turbine main shaft, gearbox, and generator bearings); a precise speed sensor for acquiring the low-speed end of the main shaft and the high-speed end of the generator, which may be an eddy current or Hall effect proximity switch; and a temperature sensor for monitoring the gearbox oil sump, generator windings, and main bearing area (a thermocouple or resistance temperature detector may be selected depending on the data range).

[0046] To achieve synchronous data acquisition from multiple sensors within the physical sensing layer and ensure data integrity and acquisition efficiency, this application requires the establishment of a multi-channel synchronous data acquisition mechanism in the data acquisition and transmission layer. Considering cost and application functional requirements, this application employs a data acquisition and transmission module based on PXI (PCI extension for instrument systems) or CompactDAQ (compact data acquisition). Currently, these modules have standardized product structures for each function, allowing for procurement and parameter settings based on actual needs. During application, to ensure the original data acquisition and transmission module can effectively acquire various monitoring data sources from the wind turbine, it should be tailored according to the characteristics of the monitoring data corresponding to the multiple sensors within the physical sensing layer. The acquisition frequencies of the modules are set separately. Specifically: for vibration and speed signals, since the vibration and speed frequencies during wind turbine operation are relatively large, in order to ensure the smooth acquisition of continuous signals under normal and abnormal conditions, the acquisition frequency range should cover twice the expected maximum frequency of the wind turbine, typically 5kHz to 25kHz. At the same time, a hardware anti-aliasing filter should be configured between the data acquisition and transmission module and the vibration sensor input source to ensure that high-order features can be effectively extracted. For temperature signals, the temperature change process is a relatively low-frequency continuous change signal, which can be acquired using a frequency of 5~80Hz, or the temperature average value of a fixed period can be obtained according to the statistical averaging unit built into the module.

[0047] The data acquisition and transmission layer establishes a Modbus TCP network between sensors and data acquisition and transmission modules at locations such as the wind turbine tower base, as well as an OPC UA fiber optic network between the wind turbine and the data platform layer and algorithm model layer located in the regional monitoring center.

[0048] Modbus TCP networks are simple to implement and have low cost, making it easy to establish a real-time interactive network for wind turbine monitoring data in a small-scale master-slave architecture between the physical sensing layer and the data acquisition and transmission layer within the wind turbine.

[0049] The OPC UA fiber optic network establishes bidirectional handshake, disconnection reconnection, data buffering, and security encryption and authentication mechanisms to ensure and control the effectiveness and security of data transmission between layers while transmitting basic sensor monitoring data values. At the same time, it controls and transmits semantic data such as data source stamps, data quality stamps, timestamps, units, ranges, and engineering units established by the data transmission layer or data platform layer through a data protocol library to meet the model training and analysis needs within the algorithm model layer.

[0050] Because wind turbines involve multiple functional structures, effective monitoring of their operating status requires processing a large amount of real-time monitoring data. The data platform layer needs to handle the centralized storage, cleaning, management, and governance of this data; specifically, this includes:

[0051] a. Data preprocessing: Raw data may have various data quality problems during the collection, transmission and storage process. It is necessary to perform necessary preprocessing before storing it in the data path and before application, including outlier removal, missing value handling, data smoothing and noise reduction, and data normalization and standardization.

[0052] Based on the characteristics of wind turbine operating data, outlier removal in this application is mainly based on the Laida criterion (3σ criterion), with the primary purpose of removing outliers caused by sensor momentary failures, lightning interference, etc. Missing value handling employs different methods for different types of data. For short-term missing data, linear interpolation, spline interpolation, or nearest neighbor imputation are used for repair. For long-term missing data, the data is marked and excluded from model training to avoid introducing false information. Data smoothing and denoising primarily target high-frequency random noise in the signal. Depending on the noise source and characteristics, moving average filters, median filters, or Savitzky-Golay filters are used for smoothing.

[0053] Data normalization and standardization are used to address the differences in the dimensions and numerical ranges of physical quantities caused by inconsistencies in sensors or data protocols. By performing dimensionless processing, including Min-Max normalization or Z-score standardization, the data is scaled to the same scale. The ultimate goal is to eliminate the influence of dimensions on the training of subsequent state prediction and analysis models.

[0054] b. Data storage and indexing: By establishing a hybrid storage architecture that combines time-series databases (such as InfluxDB and TDengine) and relational databases (such as MySQL and PostgreSQL), the time-series database is used to store and maintain high-frequency, time-series raw and characteristic data such as vibration, temperature, and rotational speed; the relational database is used to store static or structured relationships such as wind turbine basic information, measurement point configuration, rule base, and maintenance records, and to establish a unified asset hierarchy and equipment measurement point coding system to ensure efficient data query and traceability.

[0055] In particular, considering that wind turbines typically operate in high-altitude, high-wind-speed environments, the continuous changes in wind force and direction cause wind turbines to operate under variable speed, variable load, and high-noise conditions for extended periods. As a result, their signal output is unstable and contains a large amount of background noise and electromagnetic interference. Therefore, during the preprocessing stage, for the associated vibration and speed sensor data, a multi-level noise reduction process based on Complete Ensemble Empirical Mode Decomposition (CEEMDAN) and Variational Mode Decomposition (VMD) is used. The CEEMDAN adaptively decomposes the signal into a series of intrinsic mode functions (IMFs) to separate mode aliasing, and then the VMD is used to separate signal components with similar frequency components, thus filtering out background noise and electromagnetic interference.

[0056] Specifically, in practical applications, to control costs and utilize existing structures, temperature data primarily originates from the wind turbine's built-in SCADA system. Temperature changes in areas such as gearbox oil temperature and main bearing temperature are directly related not only to operating conditions like ambient temperature, wind speed, and active power, but also to their historical states. Therefore, to ensure the subsequent prediction model obtains temperature data features highly correlated with the target temperature, correlation analysis is necessary to select a set of feature variables strongly correlated with the target temperature, including wind speed, ambient temperature, current power, and previous oil temperature. A model for predicting wind turbine temperature data is then established at the algorithm model layer. Anomalies are identified and their severity determined by the residual change trend between real-time temperature monitoring results and temperature prediction output. Based on the characteristics of wind turbine temperature change data, this embodiment establishes a wind turbine temperature monitoring data prediction model based on the TCN-BiLSTM-Attention model at the algorithm model layer to predict temperature.

[0057] In actual condition analysis, due to the high specificity of the fault types and fault trends of each wind turbine, the distribution of actual fault data in the electric field is extremely uneven. Wind turbines of different types and located in different regions often have completely inconsistent fault characteristics, making it difficult to effectively establish fault labels for wind turbines for classification, identification, and prediction. Even if a scheme for regression prediction of key variables can be implemented, it requires massive amounts of computation and analysis work, which is difficult to meet the real-time and high-efficiency requirements of wind turbine condition analysis and fault early warning.

[0058] Therefore, the algorithm model layer in this application is based on the status monitoring data of the wind turbine in the standard operating mode to establish a description of the status characteristics of the wind turbine during normal operation. By analyzing and warning of the abnormal deviation between the operation process and the normal operation, the operating status and faults of the wind turbine are analyzed and warned. A benchmark model that can accurately describe the "behavior" of the equipment in a healthy state is established. When the equipment condition deteriorates, a warning is triggered by behavioral deviation. The algorithm model layer specifically includes a data extraction unit, a model library, and a dynamic monitoring and warning unit. The following will explain the model data source, model structure, and dynamic warning mechanism in detail.

[0059] (a) Data source

[0060] The data extraction unit acquires historical data from the wind turbine generator's healthy and fault-free operating cycle based on the wind turbine generator condition monitoring hardware to establish a normal state database, and allocates the normal state database to the training set, validation set, and test set; the historical data in this application includes vibration, speed, and temperature data; the purpose of the model is to construct a parametric model of the normal operating state, therefore, the model training data should be selected from the preprocessed historical data of the generator within a sufficiently long and confirmed healthy and fault-free operating cycle as the training set.

[0061] (II) Model Structure:

[0062] In this application, the model library establishes a normal behavior model based on historical data and trains the model using a normal state database; the normal behavior model in this application includes:

[0063] ① A multidimensional data distribution feature extraction model for vibration and rotational speed of wind turbines under normal operating conditions based on a support vector data description model;

[0064] The support vector data description model in this application is a feature extraction model based on the architecture of "multi-source data input - multi-layer feature extraction - high-dimensional domain description - three-dimensional boundary output". It can adapt to the multi-source, heterogeneous, nonlinear, and high-dimensional correlation feature attributes of wind turbine operation data. It is used to extract the complex coupling relationship between multi-source sensing data such as vibration, speed, and power, establish the multi-dimensional data distribution characteristics of wind turbine during normal operation, and finally output a three-dimensional bounding body surface feature set that can be directly used for real-time anomaly judgment.

[0065] In this application, the support vector data description model consists of a feature extraction and mapping layer, a high-dimensional decision boundary construction layer, and a three-dimensional bounding volume generation and output layer.

[0066] The feature extraction and mapping layer extracts features by splicing and fusing multidimensional inputs through a one-dimensional convolutional neural network or a lightweight autoencoder. It uses multi-layer feature extraction and nonlinear mapping operations to transform the linearly inseparable unit operation data in the original low-dimensional space into approximately linearly separable features in the high-dimensional feature space.

[0067] One-dimensional convolutional neural networks include multiple convolutional layers, which slide along the time axis of the data through convolutional kernels to extract adjacent time-dimensional related features in parallel; while autoencoders retain key information and learn the optimal low-dimensional features through a compression-reconstruction unsupervised learning process.

[0068] The core function of the feature extraction and mapping layer is to synchronously collect multi-source heterogeneous sensor data that is strongly correlated with the operating status of wind turbine units. The source data of these sensors is collected through the physical sensing layer and then processed and input through the data acquisition and transmission layer.

[0069] The high-dimensional decision boundary construction layer establishes a distribution description of unit operation data under normal conditions in a high-dimensional feature space. By solving a constrained convex optimization problem, it finds a three-dimensional bounding volume that minimizes the volume, and tries to wrap all normal sample points inside the sphere. A high-dimensional decision boundary that distinguishes between normal and abnormal samples is established through the surface of the three-dimensional bounding volume.

[0070] In practice, the high-dimensional decision boundary construction layer maps feature samples through radial basis functions, mapping the original features obtained by the feature extraction mapping layer from the low-dimensional feature space to the high-dimensional reproducing kernel Hilbert space (RKHS). The optimization goal of the model is to find a hypersphere with a minimized volume to achieve the tightest wrapping of normal samples. Its optimization logic is to find an optimal balance between "containing as many normal samples as possible" and "minimizing the hypersphere volume as much as possible". The mathematical essence of this process is to solve a quadratic programming constraint optimization problem with soft boundaries. The objective function of this optimization problem is to minimize the square of the hypersphere radius. The constraint condition is to ensure that most normal sample points are contained inside or on the surface of the hypersphere.

[0071] Normal sample points located on the surface of the minimum hypersphere constitute a high-dimensional feature vector. Then, a linear dimensionality reduction algorithm is used to reduce the high-dimensional feature vector to three dimensions. Based on the coordinate distribution of all normal samples in three-dimensional space after dimensionality reduction, a minimum volume bounding body containing all or most normal samples is calculated, and the complete surface feature set of the minimum volume bounding body is output. The newly input real-time data points are then mapped to three-dimensional space after undergoing the same feature extraction, dimensionality reduction and spatial transformation processes to obtain the distance between the input data and the center of the bounding body.

[0072] ② A data distribution feature extraction model for temperature and its related variables under normal operating conditions of wind turbine units based on the TCN-BiLSTM-Attention model;

[0073] The core architecture of the TCN-BiLSTM-Attention model in this application is as follows: First, the front-end Local Feature Extractor (TCN) extracts multi-scale local temporal features; then, the BiLSTM module captures global temporal dependencies bidirectionally; and finally, the multi-head attention module dynamically strengthens the contribution weights of key features. In the model design, the front-end multi-scale local spatial feature extractor, the BiLSTM module, and the multi-head attention module are connected in a hierarchical manner to form a logically coherent temporal feature extraction pipeline, extracting the key temporal features that have the greatest impact on the temperature prediction results, which are then passed to the subsequent fully connected layers to output the prediction results.

[0074] The TCN-BiLSTM-Attention model, as a wind turbine temperature change prediction model, takes temperature-related operating variables as input and outputs the predicted value of the target temperature.

[0075] The front-end local feature extractor consists of three parts: causal convolution, dilated convolution, and residual connections. Causal convolution performs weighted calculations on the historical input data up to the current time step. Dilated convolution introduces a dilation rate to allow the convolution kernel to perform skip sampling of the input features. Residual connections directly pass the input features of the TCN module across layers to the output of the module, and add them element-wise with the features extracted by the convolutional layer. The front-end multi-scale local spatial feature extractor transforms the high-dimensional raw monitoring data into a set of high-dimensional feature sequences containing sufficient local detail information, thereby obtaining multi-scale features of the monitoring data spatial dimension.

[0076] The BiLSTM module consists of two unidirectional LSTM networks with completely independent structures but opposite processing directions. At each time step, the unidirectional LSTM network outputs its own hidden state. Finally, the hidden state vectors of the two directions are concatenated according to the feature dimension to output bidirectional temporal features. The BiLSTM module performs bidirectional temporal feature extraction on the multi-scale local features extracted by the front-end multi-scale local spatial feature extractor to obtain the feature hidden relationship at mid-to-long-distance time steps.

[0077] The multi-head attention module learns a feature weight allocation strategy through training. It strengthens or weakens the weight of each time step feature or each feature dimension in the bidirectional time series features output by the BiLSTM layer, and assigns higher weights to key time step features that are strongly correlated with the temperature data changes of the wind turbine, thereby obtaining a subset of core features.

[0078] (III) Dynamic Early Warning Mechanism

[0079] The dynamic monitoring and early warning unit in this application establishes a model output early warning threshold setting and dynamic adjustment mechanism for normal and abnormal states, including:

[0080] a. Initial threshold setting based on offline data: Input historical health data into the trained model, calculate the reconstruction error between vibration and rotation speed samples and the surface distance of the 3D bounding body, as well as temperature change data samples and the predictions of the TCN-BiLSTM-Attention model; use statistical methods to set the initial threshold as follows: ,in and These are the mean and standard deviation of the difference sequence of samples under normal conditions. This is the sensitivity coefficient;

[0081] b. Adaptive dynamic threshold adjustment based on online data: Construct a dual sliding window mechanism, defining a fast observation window and a slow baseline window; when the system is running online, the data stream enters the observation window, and the dynamic threshold is adjusted accordingly. Determined by the statistical characteristics of the data within the baseline window. ;in, and These are the mean and standard deviation of the difference series within the current reference window. It is an adjustment factor for controlling sensitivity; when the average level of the difference in the observation window continues to be higher than the baseline window, the baseline window is gradually slid forward, and the statistics are updated with new normal state data, thereby realizing the self-learning and self-adaptation of the threshold to slow normal drift, avoiding false alarms. When the difference suddenly and drastically exceeds the dynamic threshold, the wind turbine state abnormality alarm is triggered.

[0082] As a preferred approach, in order to further configure a feature extraction unit for feature extraction of multi-sensor data in the algorithm model layer based on the core parameter distribution of wind turbine operation characteristics and operation status monitoring, this application uses the feature extraction unit to realize feature extraction of vibration, speed and temperature signals, and complete feature selection and dimensionality reduction processing.

[0083] The vibration and rotation speed signal analysis and feature extraction process is divided into time domain feature extraction, frequency domain feature extraction, and time-frequency domain feature extraction. The features of vibration and rotation speed are formed by time domain features, frequency domain features, and time-frequency domain features.

[0084] The time-domain characteristics of vibration and rotation speed signals are calculated directly from the preprocessed time-domain waveforms, including: ① Dimensional indicators: peak value, peak-to-peak value, mean, effective value (RMS), variance, standard deviation, root mean square value; ② Dimensionless indicators: peak factor, kurtosis, skewness, waveform factor, impulse factor, margin factor.

[0085] Frequency domain features are derived by converting the time-domain signal to the frequency domain using Fast Fourier Transform (FFT), and feature data is extracted through the following data analysis methods: ① Spectrum analysis: visually identifying characteristic frequencies and their harmonics, for example, identifying gear meshing frequencies (and bearing inner ring, outer ring, rolling element, and cage fault characteristic frequencies); ② Envelope spectrum analysis: performing Hilbert transform on the vibration signal, followed by FFT on the envelope signal to obtain the envelope spectrum. The envelope spectrum can effectively extract low-frequency fault characteristic frequencies in the modulated signal; ③ Power spectral density: reflecting the distribution of signal power in the frequency domain, which can be used to analyze the energy concentration band of the signal; ④ Cepstral analysis: performing FFT again on the logarithmic value of the power spectrum can effectively detect periodic components in the spectrum, facilitating the separation and identification of sideband families.

[0086] Time-frequency domain features are obtained through wavelet packet decomposition energy entropy and parameter-adaptive eigenmode decomposition: ① Wavelet packet decomposition energy entropy: Wavelet packet transform can simultaneously perform fine decomposition of the high-frequency and low-frequency components of a signal. The energy of each frequency band is calculated, and the wavelet packet energy entropy is calculated based on information entropy theory to quantify the complexity and uncertainty of the signal's energy distribution in the time-frequency domain; ② Feature extraction based on parameter-adaptive eigenmode decomposition: Addressing the challenge of extracting weak multi-frequency features under strong background noise, parameter-adaptive eigenmode decomposition technology is employed. FMD can adaptively decompose a signal into several physically meaningful modal components. By designing pattern discrimination criteria (such as correlation kurtosis), specific modes strictly corresponding to fault characteristic frequencies can be accurately extracted.

[0087] The feature selection and dimensionality reduction process includes: calculating the correlation between each feature and the device status label, and removing features with low correlation; using the feature importance assessment capability of the machine learning model itself, and combining the average impact value algorithm to set an importance threshold, only retaining features with scores exceeding the threshold to form the final feature set.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. An online condition monitoring system for wind turbine generator sets, characterized in that, It includes an algorithm model layer; the algorithm model layer establishes a state early warning model to describe the wind turbine during normal operation, and analyzes and warns of the wind turbine's operating status and faults by the degree of abnormal deviation between the operating process and the normal operation; the algorithm model layer specifically includes a data extraction unit, a model library, and a dynamic monitoring and early warning unit; The data extraction unit establishes a normal state database based on historical data from the wind turbine's healthy and fault-free operating cycle obtained from the wind turbine's condition monitoring hardware, and allocates the normal state database to the training set, validation set, and test set; the historical data includes vibration, speed, and temperature data; The model library establishes normal behavior models based on historical data and trains the models using a normal state database. The normal behavior model includes: a. A support vector data description model for vibration and rotational speed. The goal of the support vector data description model is to find the smallest volume three-dimensional bounding volume in the high-dimensional space composed of the feature values ​​of vibration and rotational speed, so that it contains as many vibration and rotational speed feature samples as possible under normal conditions, and to determine the optimal state feature set through the surface of the three-dimensional bounding volume. b. For temperature change data, the TCN-BiLSTM-Attention model performs correlation analysis on temperature data and temperature-related operating condition data, selects a set of feature variables that are strongly correlated with temperature data, and identifies anomalies and judges the degree of temperature anomalies by the residual change trend of real-time temperature monitoring results and temperature prediction output. The dynamic monitoring and early warning unit establishes a model for normal and abnormal states, and outputs an early warning threshold setting and dynamic adjustment process, including: a. Initial threshold setting based on offline data: Input historical health data into the trained model, calculate the reconstruction error between vibration and rotation speed samples and the surface distance of the 3D bounding body, as well as temperature change data samples and the predictions of the TCN-BiLSTM-Attention model; use statistical methods to set the initial threshold as follows: ,in and These are the mean and standard deviation of the difference sequence of samples under normal conditions. This is the sensitivity coefficient; b. Adaptive dynamic threshold adjustment based on online data: Construct a dual sliding window mechanism, defining a fast observation window and a slow baseline window; when the system is running online, the data stream enters the observation window, and the dynamic threshold is adjusted accordingly. Determined by the statistical characteristics of the data within the baseline window. ;in, and These are the mean and standard deviation of the difference series within the current reference window. It is an adjustment factor for controlling sensitivity; when the average level of the difference in the observation window continues to be higher than the baseline window, the baseline window is gradually slid forward, and the statistics are updated with new normal state data, thereby realizing the self-learning and self-adaptation of the threshold to slow normal drift, avoiding false alarms. When the difference suddenly and drastically exceeds the dynamic threshold, the wind turbine state abnormality alarm is triggered.

2. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, Based on their functions and data hierarchy, they are divided from bottom to top into: physical sensing layer, data acquisition and transmission layer, and data platform layer; The physical sensing layer is responsible for sensing the physical state quantity analog signal data of the wind turbine's operating status through sensors deployed on various key structures of the wind turbine that are to be monitored and analyzed. The data acquisition and transmission layer uses data acquisition hardware and network communication protocols to convert the analog signals of the physical sensing layer into digital signal data that the system can directly process and use, and establishes a communication network for transmission to local or remote data centers. The data platform layer establishes a preprocessing flow for digital signal data uploaded from the data acquisition and transmission layer, and stores it in a database with a data structure that matches and adapts to the source data. Based on the preset wind turbine structure and equipment and the measurement point coding rules, the labeling of the digital signal data is completed.

3. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, The TCN-BiLSTM-Attention model includes a front-end multi-scale local spatial feature extractor, a BiLSTM module, and a multi-head attention module; The front-end local feature extractor consists of three parts: causal convolution, dilated convolution, and residual connections. Causal convolution performs weighted calculations on the historical input data up to the current time step. Dilated convolution introduces a dilation rate to allow the convolution kernel to perform skip sampling of the input features. Residual connections directly pass the input features of the TCN module across layers to the output of the module, and add them element-wise with the features extracted by the convolutional layer. The front-end multi-scale local spatial feature extractor transforms the high-dimensional raw monitoring data into a set of high-dimensional feature sequences containing sufficient local detail information, thereby obtaining multi-scale features of the monitoring data spatial dimension. The BiLSTM module consists of two unidirectional LSTM networks with completely independent structures but opposite processing directions. At each time step, the unidirectional LSTM network outputs its own hidden state. Finally, the hidden state vectors of the two directions are concatenated according to the feature dimension to output bidirectional temporal features. The BiLSTM module performs bidirectional temporal feature extraction on the multi-scale local features extracted by the front-end multi-scale local spatial feature extractor to obtain the feature hidden relationship at mid-to-long-distance time steps. The multi-head attention module learns a feature weight allocation strategy through training. It strengthens or weakens the weight of each time step feature or each feature dimension in the bidirectional time series features output by the BiLSTM layer, and assigns higher weights to key time step features that are strongly correlated with the temperature data changes of the wind turbine, thereby obtaining a subset of core features.

4. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, The support vector data description model consists of a feature extraction and mapping layer, a high-dimensional decision boundary construction layer, and a three-dimensional bounding volume generation and output layer. The feature extraction and mapping layer extracts features by splicing and fusing multidimensional inputs through a one-dimensional convolutional neural network or a lightweight autoencoder. It uses multi-layer feature extraction and nonlinear mapping operations to transform the linearly inseparable unit operation data in the original low-dimensional space into approximately linearly separable features in the high-dimensional feature space. The high-dimensional decision boundary construction layer maps feature samples through radial basis functions, mapping the original features obtained by the feature extraction mapping layer from the low-dimensional feature space to the high-dimensional regenerating kernel Hilbert space. It constructs the objective function that minimizes the hypersphere and the constraint condition that most normal sample points are contained inside or on the surface of the hypersphere. Through quadratic programming constraint optimization, it solves for the high-dimensional feature vector composed of normal sample points located on the surface of the minimum hypersphere. Then, it reduces the high-dimensional feature vector to three dimensions through a linear dimensionality reduction algorithm. Based on the coordinate distribution of all normal samples in the three-dimensional space after dimensionality reduction, it calculates a minimum volume bounding body that contains all or most of the normal samples and outputs the complete surface feature set of the minimum volume bounding body. The newly input real-time data points are then mapped to three-dimensional space after the same feature extraction, dimensionality reduction and spatial transformation processing to obtain the distance between the input data and the center of the bounding body.

5. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, The data acquisition and transmission layer establishes a multi-channel synchronous data acquisition mechanism. It employs a PXI or CompactDAQ-based data acquisition and transmission module to acquire data at different frequencies based on the characteristics of the monitoring data from multiple sensors within the physical sensing layer. Specifically: For vibration and speed signals, the set acquisition frequency range covers twice the expected maximum frequency of the wind turbine; for temperature signals, a frequency of 5~80Hz is used or the average temperature value of a fixed period is obtained according to the statistical averaging unit built into the module. The data acquisition and transmission layer establishes a Modbus TCP network between the sensors and the data acquisition and transmission module, as well as an OPC UA fiber optic network between the wind turbine and the data platform layer and algorithm model layer located in the regional monitoring center. The OPC UA fiber optic network ensures and controls the effectiveness and security of data transmission between layers while transmitting basic sensor monitoring data values ​​through the establishment of bidirectional handshake, disconnection reconnection, data buffering, and security encryption and authentication mechanisms. At the same time, it controls and transmits semantic data such as data source stamp, data quality stamp, timestamp, unit, range, and engineering unit established by the data transmission layer or data platform layer through the data protocol library.

6. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, The data platform layer needs to handle the centralized storage, cleaning, management, and governance of this data; specifically, this includes: a. Data preprocessing, including outlier removal, missing value handling, data smoothing and denoising, and data normalization and standardization; b. Data storage and indexing: Establish a hybrid storage architecture that combines time-series databases and relational databases. Use time-series databases to store and maintain high-frequency, time-series vibration, speed, and temperature raw and characteristic data. Use relational databases to store wind turbine basic information, measurement point configuration, rule base, and static or structured relationships of maintenance records, and establish a unified asset hierarchy and equipment measurement point coding system.

7. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, The model library also includes a feature extraction unit, which is used to extract features from vibration, rotation speed and temperature signals, and to complete feature selection and dimensionality reduction. The vibration and rotation speed signal analysis and feature extraction process is divided into time domain feature extraction, frequency domain feature extraction and time-frequency domain feature extraction. The vibration and rotation speed features are formed by time domain features, frequency domain features and time-frequency domain features. The time-domain features are directly calculated and statistically obtained from the preprocessed time-domain waveform, including dimensional and dimensionless indicators of vibration and rotational speed. The dimensional indicators of vibration and rotational speed include peak value, peak-to-peak value, mean, RMS value, variance, standard deviation, and root mean square value. The dimensionless indicators of vibration and rotational speed include peak factor, kurtosis, skewness, waveform factor, impulse factor, and margin factor. The frequency-domain features are obtained by converting the time-domain signal to the frequency domain using Fast Fourier Transform, and then performing spectrogram analysis, envelope spectrum analysis, power spectral density analysis, and cepstral analysis. Time-frequency domain features are obtained through wavelet packet decomposition of energy entropy and parameter-adaptive eigenmode decomposition. The feature selection and dimensionality reduction process includes: calculating the correlation between each feature and the device status label, and removing features with low correlation. By leveraging the feature importance assessment capabilities of the machine learning model itself and combining it with the average impact value algorithm to set an importance threshold, only features with scores exceeding the threshold are retained to form the final feature set.

8. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, The preprocessing also includes: for the associated vibration and speed sensor data, adaptively decomposing the signal into a series of intrinsic mode functions through complete set empirical mode decomposition, separating mode aliasing, and then using variational mode decomposition to separate signal components with similar frequency components, jointly filtering out background noise and electromagnetic interference.

9. The online condition monitoring system for wind turbine generator sets according to claim 1, characterized in that, The The physical sensing layer includes at least: an integrated circuit piezoelectric accelerometer for monitoring the vibration state of the transmission system; a precise speed sensor for acquiring the low-speed end of the main shaft and the high-speed end of the generator; a temperature sensor for monitoring the gearbox oil sump, generator windings, and main bearing area; and an oil sensor for monitoring the health status parameters of the wind turbine gearbox. The operating data includes wind speed, ambient temperature, current power, and oil temperature at the previous moment.

Images