Fault early warning method and system for double-device integrated wind turbine generator

Abstract

The invention discloses a fault early warning method and system for a double-unit integrated wind turbine generator, and relates to the technical field of fault early warning of wind turbine generators, and the method comprises the steps: obtaining multi-mode historical operation data of a wind turbine generator, and carrying out the preprocessing of the multi-mode historical operation data to construct a sample data set; constructing a dual-device sensing network; and training the double-device sensing network by adopting the sample data set, embedding the double-device sensing network which is trained to be convergent into a control center of the wind turbine generator, and carrying out fault monitoring and early warning. The technical problems that in the prior art, wind turbine generator fault early warning lags behind, diagnosis precision is insufficient, and early faults are difficult to accurately recognize are solved, and the technical effects that accurate early monitoring and diagnosis of the wind turbine generator faults are achieved, and fault early warning timeliness and diagnosis accuracy are improved are achieved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine fault early warning technology, specifically to a fault early warning method and system for a wind turbine with dual integrated units. Background Technology

[0002] In wind turbine operation and maintenance management, the timeliness of fault early warning and the accuracy of diagnosis directly affect the unit's operational safety, power generation efficiency, and operation and maintenance cost control. Currently, traditional wind turbine fault early warning technologies have significant limitations: on the one hand, most methods rely on multimodal operational data from the main control SCADA system, but the data preprocessing stage often fails to effectively remove invalid data such as shutdowns and power curtailments, and lacks precise screening of key features, resulting in inconsistent data quality and affecting the reliability of model input; on the other hand, traditional early warning models mostly adopt a single deep learning architecture, making it difficult to simultaneously and efficiently mine the spatiotemporal correlation features of the data, and lacking an adaptive optimization mechanism designed for non-stationary operating environments, leading to insufficient sensitivity to early faults and frequent problems such as delayed warnings, false alarms, or missed alarms.

[0003] Existing technologies for wind turbines suffer from delayed fault warnings, insufficient diagnostic accuracy, and difficulty in accurately identifying early-stage faults. Summary of the Invention

[0004] This application provides a fault early warning method and system for wind turbine generators with dual-unit integration, which is used to address the technical problems of delayed fault early warning, insufficient diagnostic accuracy, and difficulty in accurately identifying early faults in existing wind turbine generators.

[0005] In view of the above problems, this application provides a fault early warning method and system for a dual-unit integrated wind turbine.

[0006] The first aspect of this application provides a fault early warning method for a dual-unit integrated wind turbine generator, the method comprising: Multimodal historical operating data of the wind turbine are acquired from the main control SCADA system and preprocessed to construct a sample dataset. A dual-sensor network is constructed based on CNN, LSTM, an autoencoder based on an attention mechanism, and an online adaptive Kalman filter. The dual-sensor network is trained using the sample dataset and then embedded into the control center of the wind turbine for fault monitoring and early warning.

[0007] A second aspect of this application provides a fault early warning system for a dual-unit integrated wind turbine generator, the system comprising: The sample dataset construction module is used to acquire multimodal historical operating data of the wind turbine from the main control SCADA system and preprocess it to construct a sample dataset; the dual-sensor network construction module is used to construct a dual-sensor network based on CNN, LSTM, an autoencoder based on an attention mechanism, and an online adaptive Kalman filter; the fault monitoring and early warning module is used to train the dual-sensor network using the sample dataset, and embed the converged dual-sensor network into the control center of the wind turbine for fault monitoring and early warning.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: Multimodal historical operating data of the wind turbine is acquired from the main control SCADA system and preprocessed to construct a sample dataset. A dual-generator sensing network is constructed based on CNN, LSTM, an attention-based autoencoder, and an online adaptive Kalman filter. The dual-generator sensing network is trained using the sample dataset, and the converged network is embedded into the control center of the wind turbine for fault monitoring and early warning. This achieves accurate early monitoring and diagnosis of wind turbine faults, improving the timeliness of fault warnings and the accuracy of diagnosis. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic flowchart of a fault early warning method for a dual-unit integrated wind turbine provided in this application embodiment; Figure 2 This is a schematic diagram of a fault early warning system for a dual-unit integrated wind turbine provided in an embodiment of this application.

[0011] Figure labeling: Sample dataset construction module 10, dual-device sensing network construction module 20, fault monitoring and early warning module 30. Detailed Implementation

[0012] This application provides a fault early warning method and system for wind turbines with dual integrated units, which addresses the technical problems of delayed fault early warning, insufficient diagnostic accuracy, and difficulty in accurately identifying early faults in existing wind turbine technologies.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] Example 1, as Figure 1 As shown, this application provides a fault early warning method for a dual-unit integrated wind turbine, the method comprising: Step S100: Obtain multimodal historical operating data of the wind turbine from the main control SCADA system and preprocess it to construct a sample dataset.

[0015] Specifically, the system first acquires multimodal historical operating data of the wind turbine from the main control SCADA system. The data covers 46 relevant variables, including active power, wind speed, gearbox temperature, and generator winding temperature, with a sampling frequency of 5 minutes and a collection period of no less than one year to ensure data integrity. Then, preprocessing is performed to construct a sample dataset: invalid data records during shutdown, start-up / shutdown processes, and power curtailment are deleted to generate the first processed data. Next, the importance score of each SCADA feature variable and the core state parameters of the unit is calculated using the XGBoost algorithm. The top N key feature variables, such as 14, are selected to form the input feature set. Finally, this set is normalized to eliminate the dimensional differences between the feature variables, resulting in a representative and consistent sample dataset, laying the foundation for subsequent model training.

[0016] Step S200: Construct a dual-sensor network based on CNN, LSTM, an attention-based autoencoder, and an online adaptive Kalman filter.

[0017] Specifically, the focus is on constructing a dual-device perceptron network, using CNN, LSTM, attention-based autoencoders, and online adaptive Kalman filters as core technical components. A dual-device architecture consisting of a global perceptron and a specialized diagnostic module is built modularly: First, a global perceptron is constructed by building an encoder consisting of a one-dimensional convolutional layer, a max-pooling layer, a bidirectional LSTM layer, and an attention mechanism layer connected sequentially to extract key spatiotemporal features of the input sequence. Then, a decoder consisting of a one-dimensional convolutional layer and an upsampling layer is built to reconstruct the feature representation output by the encoder back to the original input dimension. Simultaneously, an online adaptive Kalman filter is embedded between the encoder and decoder to refine the feature sequence. The system performs state estimation and noise filtering, and the global sensor assesses the overall operating status of the unit through the reconstruction error output. On the other hand, a specialized diagnostic tool is constructed to obtain the high-frequency vibration raw signals of the preset subsystems from the state monitoring device independent of the main control SCADA system. These signals are aligned with the SCADA multimodal data by timestamp and fused at the feature level. A subsystem topology graph is constructed with sensor measurement points as nodes and sensor associations as edges. The spatiotemporal patterns of fault evolution are learned by combining graph convolutional networks and gated recurrent units to achieve deep fault diagnosis and location. Finally, the global sensor and the specialized diagnostic tool are connected to establish a collaborative triggering mechanism to form a complete dual-sensor network.

[0018] Step S300: Train the dual-sensor network using the sample dataset, and embed the converged dual-sensor network into the control center of the wind turbine for fault monitoring and early warning.

[0019] Specifically, to complete model training and fault early warning deployment, the dual-sensor network is trained in stages using a constructed sample dataset: First, normal operating condition samples are extracted to perform unsupervised pre-training on the autoencoder of the global sensor, minimizing the reconstruction error between input and output; then, the weights of the pre-trained autoencoder are initialized, and end-to-end joint supervised training is performed with an online adaptive Kalman filter and a specialized diagnostic tool that incorporates fault label data, simultaneously optimizing the parameters of the entire network until the model converges; the converged dual-sensor network is embedded in the wind turbine control center, accessing the unit's operating data in real time. The global sensor monitors the unit's status based on a health baseline and dynamic alarm thresholds. When an anomaly is detected, the specialized diagnostic tool is automatically triggered to accurately locate the fault and output early warning information, realizing real-time monitoring and early warning of wind turbine faults.

[0020] In one possible implementation, step S200 further includes: Step S210: CNN is used to extract local spatial features of the data, LSTM is used to capture the dependencies between time series, and a global perceptron is constructed with a lightweight CNN-BiLSTM-attention autoencoder as the core. The global perceptron is used to evaluate the overall operating status of the wind turbine and the output reconstruction error is used as a health indicator.

[0021] Step S220: A specialized diagnostic tool is constructed using a model that integrates high-frequency vibration signals and a time-series neural network for in-depth fault diagnosis and localization.

[0022] Step S230: Connect the global sensor to the specialized diagnostic device and establish a collaborative triggering mechanism to obtain the dual-sensor network.

[0023] Specifically, an encoder structure is first constructed, sequentially connecting a one-dimensional convolutional layer, a max-pooling layer, a bidirectional LSTM layer, and an attention mechanism layer. The one-dimensional convolutional CNN layer extracts local spatial features from the multimodal historical operating data of the wind turbine, such as active power, wind speed, and equipment temperature. The max-pooling layer further reduces data dimensionality while preserving key patterns. The bidirectional LSTM layer captures the forward and backward dependencies of data across time steps, fully exploring the dynamic patterns in the time-series data. The attention mechanism layer assigns differentiated weights to different parts of the sequence, prioritizing key features strongly correlated with the turbine's operating state, thus improving the model's sensitivity to core information. A decoder structure is then constructed, restoring the temporal resolution of the original input sequence through an upsampling layer, and refining the features and reconstructing the output data using a one-dimensional convolutional layer. Simultaneously, an online adaptive Kalman filter is embedded between the encoder and decoder to accurately estimate the state and filter noise in the encoder's output feature sequence, ultimately forming a global perceptron. This perceptron can comprehensively assess the overall operating state of the wind turbine, using the reconstruction error between the model input and output as a health indicator to intuitively reflect the degree to which the turbine deviates from normal operating conditions.

[0024] A model fusing high-frequency vibration signals and a time-series graph neural network is adopted: First, high-frequency vibration raw signals of preset subsystems, such as gearbox and generator subsystems, are obtained from a condition monitoring device independent of the main control SCADA system. These signals are then aligned with multimodal historical operating data obtained from the main control SCADA system by timestamp and fused at the feature level to generate a fused feature vector that combines temporal operating characteristics and vibration characteristics. Next, using each sensor measurement point within the preset subsystem as a node, and based on the fused feature vector, a topology graph accurately reflecting the subsystem structure is constructed with the physical connections or functional associations between sensors as edges. Finally, based on this subsystem topology graph, a graph convolutional network is used to capture the spatial dependencies between nodes, and a gated recurrent unit is combined to jointly learn the dynamic temporal patterns of the graph structure during fault evolution, ultimately establishing a specialized diagnostic tool capable of deep fault diagnosis and precise location.

[0025] By closely linking the two through a data interaction interface, the activation logic of the specialized diagnostic tool is clearly defined, and it only starts running when the global sensor detects an anomaly, avoiding unnecessary resource consumption. Subsequently, a collaborative triggering mechanism is established: during the fault-free operation phase of the wind turbine, the reconstruction error distribution output by the global sensor is continuously statistically analyzed, and a health status baseline is established based on this distribution; using the health status baseline as a reference, the alarm threshold for the current time window is dynamically calculated using an exponentially weighted moving average method to ensure that the threshold can adapt to changes in the unit's operating environment; when the real-time reconstruction error exceeds the alarm threshold for a preset number of consecutive times, the unit is judged to be in an abnormal state, the specific subsystem to which the abnormal variable belongs is automatically analyzed, and the corresponding specialized diagnostic tool is triggered to start running, accurately carrying out in-depth fault diagnosis and location, and finally forming a dual-sensor network of global monitoring and specialized diagnosis working in synergy.

[0026] In one possible implementation, step S210 further includes: Step S211: Construct an encoder consisting of a one-dimensional convolutional layer, a max pooling layer, a bidirectional LSTM layer, and an attention mechanism layer connected in sequence, which is used to extract key spatiotemporal features from the input sequence.

[0027] Step S212: Construct a decoder consisting of a one-dimensional convolutional layer and an upsampling layer to reconstruct the feature representation output by the encoder back to the dimension of the original input sequence.

[0028] Step S213: Embed an online adaptive Kalman filter between the encoder and the decoder to perform state estimation and noise filtering on the feature sequence output by the encoder, thereby obtaining the global perceptron.

[0029] Specifically, the functional layers are connected sequentially in the order of one-dimensional convolutional layer, max pooling layer, bidirectional LSTM layer, and attention mechanism layer to form an architecture for collaborative extraction of key spatiotemporal features: First, a one-dimensional convolutional layer is connected to capture the correlation features of local spatial dimensions of the data, such as the local coupling relationship between different operating parameters, for the multimodal historical operating data of wind turbines, such as time-series data of active power, wind speed, and gearbox temperature. Then, a max pooling layer is connected to downsample the feature map output by the one-dimensional convolutional layer, which reduces the data dimensionality and computational complexity while retaining key feature patterns. To avoid redundant information interference, a bidirectional LSTM layer is then connected to learn data dependencies from both the forward and reverse directions of the time series. This captures the impact of historical moments on the current state and also considers the correlation between future trends and current features, fully exploring the dynamic patterns of the time series. Finally, an attention mechanism layer is connected to assign differentiated weights to the features at different time steps output by the bidirectional LSTM layer. This allows the model to prioritize key features strongly correlated with the unit's health status, such as the time series changes of core parameters like gearbox oil temperature and generator winding temperature. Ultimately, this achieves the goal of accurately extracting key spatiotemporal features from the input sequence.

[0030] The above-mentioned sampling layer and one-dimensional convolutional layer structure realizes feature reconstruction: First, the low-dimensional feature representation output by the encoder is upsampled by the upsampling layer, and the temporal resolution of the data is restored by interpolation sampling, so that it gradually approaches the temporal length of the original input sequence; then, the one-dimensional convolutional layer is connected to refine the features after the dimensionality is increased. The convolution operation corrects the feature details and makes up for the information loss in the upsampling process. Finally, the abstract feature representation output by the encoder is reconstructed back to the same dimension as the original input sequence, so as to ensure that the unit's operating status can be measured by comparing the input and output errors.

[0031] The integration of the filtering module with the global perceptron is completed: an online adaptive Kalman filter is used as the core filtering unit and embedded in the feature transfer path between the encoder and decoder. After receiving the key spatiotemporal feature sequence output by the encoder, the filter first dynamically tracks and predicts the effective information in the feature sequence through a state estimation algorithm. Then, it filters out redundant information such as sensor noise and environmental interference that may exist in the wind turbine operation data. By adjusting the filtering parameters in real time, such as process noise covariance and measurement noise covariance, it adapts to the dynamic changes of the data and improves the purity and stability of the feature sequence. The filtered and optimized feature sequence is then fed into the decoder for reconstruction, ultimately forming a complete link of encoder feature extraction, filter feature optimization, and decoder sequence reconstruction, resulting in a global perceptron with accurate feature processing capabilities.

[0032] In one possible implementation, step S220 further includes: Step S221: Obtain the original high-frequency vibration signal of the preset subsystem from the state monitoring device independent of the main control SCADA system, align it with the multimodal historical operation data obtained from the main control SCADA system on the timestamp, and perform splicing and fusion at the feature level to obtain the fused feature vector.

[0033] Step S222: Based on the sensor measurement points within the preset subsystem as nodes, and based on the fused feature vector, construct a subsystem topology graph with the physical connections or functional associations between sensors as edges.

[0034] Step S223: Based on the subsystem topology graph, use graph convolutional networks to capture the spatial dependencies between nodes, and combine them with gated recurrent units to jointly learn the dynamic temporal patterns of the graph structure in the fault evolution process, and establish the specialized diagnostic tool.

[0035] Specifically, high-frequency vibration raw signals from pre-defined subsystems, such as gearboxes and generator subsystems, are collected from dedicated condition monitoring devices independent of the main control SCADA system, such as vibration sensors. These signals accurately reflect the mechanical operating status of the equipment, such as minor faults like bearing wear and abnormal gear meshing. Simultaneously, multi-modal historical operating data of the wind turbine, such as active power, wind speed, equipment temperature, and pressure, are obtained from the main control SCADA system. Then, using timestamps as a reference, the high-frequency vibration signals and multi-modal historical operating data are strictly aligned to ensure consistency in the time dimension of data from different sources. Finally, the two types of data are spliced ​​and fused at the feature level, integrating vibration features, such as vibration amplitude and frequency spectrum features, with operating features, such as temperature change rate and power fluctuation, into a unified fused feature vector. This retains detailed information about the mechanical status of the equipment while covering the overall operating background of the unit, providing comprehensive data support for subsequent fault diagnosis.

[0036] Construction of the subsystem topology: Using pre-defined subsystems, such as various sensor measurement points within the gearbox system, including vibration sensors at the gearbox drive end, vibration sensors at the non-drive end, and oil temperature sensors, as nodes in the topology, the attributes of each node are assigned by the feature data of the corresponding measurement point in the obtained fused feature vector. Then, based on the actual physical connection relationship between sensors, such as vibration sensors at different positions on the same shaft system or functional associations, such as oil temperature sensors and vibration sensors jointly reflecting the gearbox lubrication and mechanical operation status, edges are established between corresponding nodes. This constructs a topology that accurately reflects the physical structure and functional associations of the subsystem. This topology can intuitively present the spatial correlation of the data from each measurement point, laying the foundation for capturing the propagation path of faults within the subsystem.

[0037] Using the constructed subsystem topology as the core input, the graph convolutional network (GCN) is first used to learn features from the topology. The GCN effectively captures the spatial dependencies between nodes by weightedly fusing the features of each sensor measurement point node with the features of its neighboring nodes. For example, in a gearbox system, the feature transmission patterns formed by the physical association between the drive-end vibration sensor and the oil temperature sensor can accurately uncover the spatial distribution characteristics and propagation paths of faults within the subsystem. Subsequently, the feature sequence containing spatial association information output by the GCN is fed into a gated recurrent unit (GRU). The GRU, with its gating mechanism, can adaptively retain or forget key information in the time-series data, efficiently learning the graph structure in fault evolution. The model captures dynamic temporal patterns during the fault transformation process, such as the trend and correlation of the characteristics of each measurement point changing over time from the initial budding stage to the obvious deterioration stage. By extracting spatial features through graph convolutional networks and modeling temporal features through gated recurrent units, a deep fusion of spatial and temporal dual-dimensional features of the fault is achieved. Finally, the Softmax function is used in the final output layer of the model to map the fused features to the probability distribution of the subsystem under different fault modes, such as gearbox bearing wear, abnormal gear meshing, and generator winding overheating. The fault mode with the highest probability value is the potential fault type of the current subsystem determined by the model. In the end, a specialized diagnostic tool with both spatial correlation capture capability and temporal dynamic learning capability is established.

[0038] In one possible implementation, step S300 further includes: Step S310: Extract sample data under normal operating conditions from the sample dataset, perform unsupervised pre-training on the autoencoder in the global perceptron, and minimize the reconstruction error between the input and output.

[0039] Step S320: Initialize the pre-trained autoencoder weights and perform end-to-end joint supervised training with the online adaptive Kalman filter and the specialized diagnostic tool that incorporates labeled data.

[0040] Specifically, sample data under normal operating conditions of wind turbine units are selected from the constructed sample dataset, such as SCADA multimodal data when the unit is running without faults and sensor data without abnormal vibrations. This type of data can reflect the characteristic distribution pattern of the unit under healthy conditions. The normal operating condition sample data is input into the autoencoder, and unsupervised training is carried out with the goal of minimizing the reconstruction error between the input data and the decoder reconstruction output data. During the training process, the encoder continuously optimizes its ability to extract spatiotemporal features of normal operating condition data, and the decoder continuously improves the feature reconstruction accuracy. The trainable parameters of the encoder's one-dimensional convolutional layer, bidirectional LSTM layer, attention mechanism layer and the decoder's one-dimensional convolutional layer and upsampling layer are adjusted iteratively through backpropagation until the reconstruction error converges to the preset threshold, so that the autoencoder can stably learn the data feature patterns under normal operating conditions, laying the foundation for subsequent joint training.

[0041] End-to-end joint supervised training of the entire network was conducted: First, the weights of the pre-trained autoencoder were fixed and initialized to ensure that its learned normal operating condition feature extraction capabilities were retained. Then, the autoencoder was integrated with the embedded online adaptive Kalman filter and the constructed specialized diagnostic tool into a complete network architecture. At the same time, sample data with fault labels were introduced, such as operating data labeled with fault types such as gearbox bearing wear and generator winding overheating. Supervised training was carried out with the goal of accurately outputting fault types. During the training process, the online adaptive Kalman filter dynamically optimized the noise filtering and state estimation effects on the encoder output features. The specialized diagnostic tool combined the spatial dependence captured by the graph convolutional network and the temporal pattern learned by the gated recurrent unit to continuously adjust the parameters to improve the fault identification accuracy. Through end-to-end backpropagation, the noise covariance parameters of the Kalman filter and the parameters of the graph convolutional layer and GRU layer of the specialized diagnostic tool were updated synchronously. At the same time, the key parameters of the autoencoder were fine-tuned to adapt to the fault data features until the fault diagnosis error of the entire network converged, and the training of the dual-sensor network was finally completed.

[0042] In one possible implementation, step S300 further includes: Step S330: During the fault-free operation phase of the wind turbine, statistically analyze the distribution of reconstruction error output by the global sensor and establish a health status baseline.

[0043] Step S340: Based on the health status baseline, dynamically calculate the alarm threshold for the current time window using the exponentially weighted moving average method.

[0044] Step S350: When the real-time reconstruction error exceeds the alarm threshold for a preset number of consecutive times, the unit is determined to be in an abnormal state, and the corresponding special diagnostic device is automatically triggered to start operation according to the subsystem to which the abnormal variable belongs.

[0045] Specifically, during the stable phase of wind turbine operation without faults, reconstruction error data output by the global sensor is continuously collected. This error is the input of the autoencoder in the global sensor, i.e., the deviation between the normal operating data and the reconstruction output of the decoder, which can intuitively reflect the degree of feature fitting under the healthy state of the unit. By statistically analyzing the distribution characteristics of the reconstruction error in this phase, such as the mean, variance, and probability density distribution, the normal fluctuation range of the reconstruction error during the healthy operation of the unit is determined, thereby establishing a baseline of the healthy state, which serves as the core reference standard for subsequent judgment of whether the unit is abnormal.

[0046] Based on the established health status baseline, the exponentially weighted moving average (EWMA) method is used to process the reconstruction error collected in real time. This method dynamically tracks the trend of error changes by assigning higher weights to recent error data and lower weights to long-term data, effectively avoiding the interference of random fluctuations on threshold determination. For different operating scenarios, such as wind speed changes and load fluctuations, corresponding time windows are set. The EWMA value is calculated in real time within each time window, and combined with the normal fluctuation range of the health baseline, the alarm threshold of the current window is determined. This allows the threshold to adapt to the dynamic changes in the unit's operating environment, avoiding false alarms caused by fixed thresholds and preventing missed alarms of potential anomalies.

[0047] The system monitors the reconstruction error output by the global sensor in real time and compares it with the calculated current alarm threshold. When the real-time reconstruction error exceeds the alarm threshold for a preset number of times (e.g., 3 times), and after ensuring that accidental interference is eliminated, the unit is determined to be in an abnormal state. At the same time, by analyzing the variables contributing to the abnormal error in the global sensor, such as gearbox oil temperature and generator vibration, the specific subsystem to which the abnormality belongs is determined, such as the gearbox system or the generator subsystem. Subsequently, the system automatically triggers the dedicated diagnostic tool corresponding to the subsystem to be put into operation. The dedicated diagnostic tool quickly performs in-depth fault diagnosis and location by fusing high-frequency vibration signals with time-series neural networks, accurately identifying the cause and type of abnormality, and providing clear guidance for subsequent operation and maintenance.

[0048] In one possible implementation, step S100 further includes: Step S110: Delete invalid data records in the multimodal historical operation data where the unit is in the process of shutdown, startup, shutdown, or power restriction operation, and generate the first processing data.

[0049] Step S120: Using the XGBoost algorithm, calculate the importance score of each SCADA feature variable and the core status parameter of the unit on the first processed data, and select the top N feature variables in terms of importance to form the input feature set.

[0050] Step S130: Normalize the input feature set.

[0051] Specifically, the focus is on eliminating invalid data: The multimodal historical operating data obtained from the main control SCADA system includes records of the unit during shutdown, startup, shutdown processes, and power curtailment. This type of data is invalid because the unit is not in a normal and stable operating condition, such as large speed fluctuations during startup or abnormal power output during power curtailment. It cannot truly reflect the characteristic patterns of the unit's healthy operation or fault state. Through data labeling and filtering algorithms, the above invalid data records are deleted one by one, and only the data of the unit in a normal power generation and stable operating state is retained to generate the first processing data, ensuring the accuracy of subsequent data processing and model training.

[0052] For the first processed data, the XGBoost extreme gradient boosting algorithm is used to evaluate feature importance. This algorithm constructs multiple decision trees and calculates the contribution of each SCADA feature variable, such as active power, wind speed, gearbox temperature, and generator winding temperature, to the core state parameters of the unit during iterative training, generating corresponding importance scores. Based on the model training requirements and feature discriminability, the top N feature variables in terms of importance are selected, such as the top 14 key parameters. The specific value of N can be adjusted according to the actual subsystem complexity to form the input feature set. This reduces the impact of redundant features on model training efficiency while retaining the key information that plays a core role in judging the unit's state.

[0053] Since the dimensions and numerical ranges of different feature variables in the selected input feature set differ, directly inputting them into the model could lead to features with large numerical differences having an excessive impact on model training, interfering with the model's learning of key features. Therefore, standardization or normalization algorithms, such as Min-Max normalization and Z-Score standardization, are used to map the values ​​of each variable in the input feature set to a uniform range, such as [0, 1], to eliminate the influence of differences in dimensions and numerical levels. This allows each feature to participate in model training with equal weight, improving the stability and accuracy of the model's learning of data features.

[0054] Example 2, based on the same inventive concept as the fault early warning method for a dual-unit integrated wind turbine in the foregoing examples, such as... Figure 2 As shown, this application provides a fault early warning system for a dual-unit integrated wind turbine. The system and method embodiments in this application are based on the same inventive concept. The system includes: The sample dataset construction module 10 is used to obtain multimodal historical operating data of wind turbines from the main control SCADA system and perform preprocessing to construct the sample dataset.

[0055] The dual-sensor network building module 20 is used to build dual-sensor networks based on CNN, LSTM, attention-based autoencoders, and online adaptive Kalman filters.

[0056] The fault monitoring and early warning module 30 is used to train the dual-device sensing network using the sample dataset, and embed the converged dual-device sensing network into the control center of the wind turbine for fault monitoring and early warning.

[0057] Furthermore, the system is also used to implement the following functions: A global perceptron is constructed using CNN to extract local spatial features of the data and LSTM to capture the dependencies between time series. A lightweight CNN-BiLSTM-attention autoencoder is used as the core to evaluate the overall operating status of the wind turbine and the output reconstruction error is used as a health indicator. A specialized diagnostic tool is constructed using a model that integrates high-frequency vibration signals and time-series graph neural networks for deep fault diagnosis and localization. The global perceptron and the specialized diagnostic tool are connected and a collaborative triggering mechanism is established to obtain the dual-device perception network.

[0058] Furthermore, the system is also used to implement the following functions: An encoder consisting of a one-dimensional convolutional layer, a max-pooling layer, a bidirectional LSTM layer, and an attention mechanism layer is constructed to extract key spatiotemporal features from the input sequence. A decoder consisting of a one-dimensional convolutional layer and an upsampling layer is constructed to reconstruct the feature representation output by the encoder back to the dimension of the original input sequence. An online adaptive Kalman filter is embedded between the encoder and the decoder to perform state estimation and noise filtering on the feature sequence output by the encoder, thus obtaining the global perceptron.

[0059] Furthermore, the system is also used to implement the following functions: High-frequency vibration raw signals from a pre-defined subsystem are acquired from a state monitoring device independent of the main SCADA system and aligned with multimodal historical operating data acquired from the main SCADA system on the timestamp. These signals are then spliced ​​and fused at the feature level to obtain a fused feature vector. Based on the sensor measurement points within the pre-defined subsystem as nodes, and using the fused feature vector as edges along with the physical connections or functional associations between sensors, a subsystem topology graph is constructed. Based on the subsystem topology graph, a graph convolutional network is used to capture the spatial dependencies between nodes and combined with a gated recurrent unit to jointly learn the dynamic temporal patterns of the graph structure during fault evolution, thereby establishing the specialized diagnostic tool.

[0060] Furthermore, the system is also used to implement the following functions: Sample data under normal operating conditions is extracted from the sample dataset, and the autoencoder in the global perceptron is pre-trained unsupervised to minimize the reconstruction error between the input and output. The weights of the pre-trained autoencoder are initialized and then jointly supervised end-to-end with the online adaptive Kalman filter and the specialized diagnostic device that incorporates labeled data.

[0061] Furthermore, the system is also used to implement the following functions: During the fault-free operation phase of the wind turbine, the distribution of reconstruction errors output by the global sensor is statistically analyzed to establish a health status baseline. Based on the health status baseline, the alarm threshold for the current time window is dynamically calculated using an exponentially weighted moving average method. When the real-time reconstruction error exceeds the alarm threshold for a preset number of consecutive times, the turbine is determined to be in an abnormal state, and the corresponding specialized diagnostic device is automatically triggered to operate according to the subsystem to which the abnormal variable belongs.

[0062] Furthermore, the system is also used to implement the following functions: Invalid data records in the multimodal historical operation data that indicate the unit is in a shutdown, startup, shutdown process, or power rationing state are deleted to generate the first processed data; using the XGBoost algorithm, the importance scores of each SCADA feature variable and the unit's core status parameters are calculated on the first processed data, and the top N feature variables in terms of importance are selected to form the input feature set; the input feature set is then normalized.

[0063] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0064] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0065] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A fault early warning method for a wind turbine generator with dual integrated units, characterized in that, include: The multimodal historical operation data of the wind turbine were obtained from the main control SCADA system and preprocessed to construct a sample dataset. A dual-sensor network is constructed based on CNN, LSTM, an attention-based autoencoder, and an online adaptive Kalman filter. The dual-sensor network is trained using the sample dataset, and the converged dual-sensor network is embedded into the control center of the wind turbine for fault monitoring and early warning.

2. The fault early warning method for a dual-unit integrated wind turbine as described in claim 1, characterized in that, A dual-sensor network is constructed based on CNN, LSTM, an attention-based autoencoder, and an online adaptive Kalman filter, including: CNN is used to extract local spatial features of data, LSTM captures the dependencies between time series, and a global perceptron is constructed with a lightweight CNN-BiLSTM-attention autoencoder as the core. The global perceptron is used to evaluate the overall operating status of wind turbine units and the output reconstruction error is used as a health indicator. A specialized diagnostic tool was constructed using a model that integrates high-frequency vibration signals and time-series neural networks for in-depth fault diagnosis and localization. The global sensor is connected to the specialized diagnostic device, and a collaborative triggering mechanism is established to obtain the dual-sensor network.

3. The fault early warning method for a dual-unit integrated wind turbine as described in claim 2, characterized in that, We utilize CNNs to extract local spatial features from data, and LSTMs to capture dependencies between time series. A global perceptron is constructed around a lightweight CNN-BiLSTM-attention autoencoder, including: An encoder consisting of a one-dimensional convolutional layer, a max pooling layer, a bidirectional LSTM layer, and an attention mechanism layer is constructed to extract key spatiotemporal features from the input sequence. A decoder consisting of one-dimensional convolutional layers and upsampling layers is constructed to reconstruct the feature representation output by the encoder back to the dimension of the original input sequence; An online adaptive Kalman filter is embedded between the encoder and the decoder to perform state estimation and noise filtering on the feature sequence output by the encoder, thereby obtaining the global perceptron.

4. The fault early warning method for a dual-unit integrated wind turbine as described in claim 1, characterized in that, A specialized diagnostic tool is constructed using a model that integrates high-frequency vibration signals and a time-series neural network for in-depth fault diagnosis and localization, including: The high-frequency vibration raw signal of the preset subsystem is obtained from the state monitoring device independent of the main control SCADA system, and is aligned with the multimodal historical operation data obtained from the main control SCADA system on the timestamp, and spliced ​​and fused at the feature level to obtain the fused feature vector. Based on the sensor measurement points within the preset subsystem as nodes, and based on the fused feature vector, a subsystem topology graph is constructed with the physical connections or functional associations between sensors as edges. Based on the subsystem topology graph, a graph convolutional network is used to capture the spatial dependencies between nodes, and combined with a gated recurrent unit to jointly learn the dynamic temporal pattern of the graph structure in the fault evolution process, thereby establishing the specialized diagnostic tool.

5. A fault early warning method for a dual-unit integrated wind turbine as described in claim 2, characterized in that, Training the dual-sensor network using the sample dataset includes: Extract sample data under normal working conditions from the sample dataset, and perform unsupervised pre-training on the autoencoder in the global perceptron to minimize the reconstruction error between the input and output. The pre-trained autoencoder weights are initialized and then subjected to end-to-end joint supervised training with the online adaptive Kalman filter and the specialized diagnostic tool that incorporates labeled data.

6. The fault early warning method for a dual-unit integrated wind turbine as described in claim 2, characterized in that, Establish a collaborative triggering mechanism, including: During the fault-free operation phase of the wind turbine, the distribution of reconstruction error output by the global sensor is statistically analyzed to establish a health status baseline. Based on the aforementioned health status baseline, the alarm threshold for the current time window is dynamically calculated using an exponentially weighted moving average method. When the real-time reconstruction error exceeds the alarm threshold for a preset number of consecutive times, the unit is determined to be in an abnormal state, and the corresponding special diagnostic tool is automatically triggered to start operation according to the subsystem to which the abnormal variable belongs.

7. The fault early warning method for a dual-unit integrated wind turbine as described in claim 1, characterized in that, The system acquires multimodal historical operating data of the wind turbine from the main SCADA system and performs preprocessing, including: Delete invalid data records from the multimodal historical operation data where the unit is in the process of shutdown, startup, shutdown, or power restriction operation, and generate the first processed data; Using the XGBoost algorithm, the importance scores of each SCADA feature variable and the core status parameters of the unit are calculated for the first processed data, and the top N feature variables in terms of importance are selected to form the input feature set. The input feature set is normalized.

8. A fault early warning system for a dual-unit integrated wind turbine generator, characterized in that, The system is used to implement the fault early warning method for a dual-unit integrated wind turbine as described in any one of claims 1-7, the system comprising: The sample dataset construction module is used to obtain multimodal historical operating data of wind turbines from the main control SCADA system and preprocess it to construct the sample dataset. A dual-sensor network building module is used to construct dual-sensor networks based on CNN, LSTM, attention-based autoencoders, and online adaptive Kalman filters. The fault monitoring and early warning module is used to train the dual-sensor network using the sample dataset, and embed the converged dual-sensor network into the control center of the wind turbine for fault monitoring and early warning.

Images