Intelligent fault diagnosis method and system for wind power equipment

Abstract

The invention relates to the technical field of intelligent fault diagnosis of wind power equipment, and discloses an intelligent fault diagnosis method and system for wind power equipment. According to the method, an operation data stream and a power grid dispatching instruction stream are processed through time synchronization to generate a graph structure feature tensor, a power grid frequency regulation event is dynamically segmented by using a reinforcement learning agent, and a virtual sample is generated; observation data are decoupled into a forced dynamic response component dominated by a power grid instruction and an inherent mechanical dynamic component determined by a component health state through a physical information neural network, a double-branch diagnosis model is constructed to process causal inference and fault classification respectively, and false correlation is eliminated through a gradient inversion layer. And finally, deploying a digital twin model to simulate fault-free expected data, and fusing and outputting a diagnosis report with a confidence interval through a differential mahalanobis distance and ensemble learning. Normal dynamic response and mechanical faults caused by power grid dispatching are effectively distinguished, the false alarm probability is reduced, and the diagnosis reliability is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for wind power equipment, and in particular to an intelligent fault diagnosis method and system for wind power equipment. Background Technology

[0002] In wind power operation and maintenance scenarios, wind turbines are exposed to variable climate and mechanical loads for extended periods, increasing the risk of component degradation. Therefore, efficient fault diagnosis is crucial for maintaining power generation stability. Intelligent fault diagnosis technology integrates vibration, temperature, and acoustic sensors to collect operational data in real time and uses machine learning algorithms to perform pattern recognition on historical and real-time data streams, thereby inferring abnormal characteristics and fault evolution trends. Based on the analysis results, the system can proactively issue early warnings to guide operation and maintenance personnel in carrying out targeted inspections and repairs, ultimately reducing the probability of unplanned downtime and optimizing full lifecycle management.

[0003] Existing intelligent fault diagnosis technologies for wind power equipment suffer from the following technical challenges: As participants in grid frequency regulation, wind turbines must rapidly adjust their active power output to maintain stability when grid frequency deviations occur. This adjustment leads to drastic fluctuations in turbine power output within a short period. These power fluctuations cause transient changes in operating parameters such as drivetrain torque and speed, resulting in high-frequency oscillations or step-like characteristics in monitoring data. Data-driven fault diagnosis models are typically trained based on historical steady-state operating data, and their decision boundaries are ill-suited to adapting to normal dynamic responses triggered by grid dispatch. The model may misinterpret data patterns corresponding to rapid power fluctuations as mechanical faults. For example, gearbox bearing failure is often accompanied by abnormal torque fluctuations. When the wind turbine suddenly increases power output in response to frequency regulation commands, the model may misidentify normal torque transients as early bearing failures, leading to frequent false alarms and interfering with maintenance decisions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent fault diagnosis method and system for wind power equipment. This invention solves the technical problem that rapid power fluctuations caused by wind turbines participating in grid frequency regulation lead to data-driven fault diagnosis models misjudging normal dynamic responses as mechanical faults.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: In a first aspect, the present invention provides an intelligent fault diagnosis method for wind power equipment, comprising: Step 1: Obtain the wind turbine's operating data stream and the grid dispatch command stream, perform time synchronization processing on the operating data stream and the command stream, and generate a graph structure feature tensor including power commands, wind turbine state variables and grid frequency measurements; Step 2: Using a reinforcement learning agent, the command features in the graph structure feature tensor and the real-time power grid frequency deviation are used as state inputs to dynamically determine the segmentation boundary of the power grid frequency regulation event, obtain the segmented data segments, and adaptively adjust the data window length. At the same time, virtual samples covering different command intensities and wind turbine operating points are generated. Step 3: Input the segmented data fragments into the physical information neural network. Under the constraints of the wind turbine drive chain mechanical equation and the generator electromagnetic transient model, the observed data is decoupled into the forced dynamic response component dominated by the grid command and the inherent mechanical dynamic component determined by the component health status. The component health index, which is independent of grid disturbance, is extracted by the variational autoencoder. Step 4: Construct a dual-branch diagnostic model. The causal inference branch identifies the causal path from the power grid command to the torque response from the forced dynamic response components and calculates the contribution. The fault classification branch processes the component health index. Train the dual-branch diagnostic model and use a gradient inversion layer to make the fault classification features uncorrelated with the power grid contribution. Step 5: Deploy the wind turbine digital twin model, receive real-time grid commands and operating setpoints, simulate expected fault-free operating data, process the actual operating data using a trained dual-branch diagnostic model, obtain preliminary diagnostic results, compare the preliminary diagnostic results with the expected data simulated by the digital twin, calculate the differential Mahalanobis distance, fuse the output probability and differential results of the diagnostic model, and generate a diagnostic report with confidence intervals. A fault alarm is triggered only when the abnormality of the inherent mechanical dynamic component exceeds the expected state and the overall confidence level is higher than the threshold.

[0006] Furthermore, in the intelligent fault diagnosis method for wind power equipment described in this invention, step 1 includes: The high-frequency data bus of the wind turbine main controller acquires millisecond-level active power, transmission chain torque and generator speed sequences to form an operating data stream; Receive timestamped grid dispatch command streams from the wind farm energy management system; The running data stream and the power grid dispatch command stream are aligned using a dynamic time warping algorithm to eliminate transmission delay and sampling rate differences, resulting in aligned data. A graph attention network is used to jointly encode the power command, wind turbine state variables, and grid frequency measurements in the aligned data, generating a spatiotemporally correlated graph structure feature tensor and storing it in a circular buffer.

[0007] Furthermore, in the intelligent fault diagnosis method for wind power equipment described in this invention, step 2 includes: The instruction features and real-time grid frequency deviation in the graph structure feature tensor are used as the state input of the reinforcement learning agent, wherein the state space includes the current instruction value, historical instruction sequence, real-time grid frequency deviation and wind turbine power change rate; Define the action space including extending the window, shortening the window, and the start and end of the confirmation event; Set a reward function to give a positive reward when the data in the segmented window can be explained by the wind turbine response physical model, and give a negative reward when segmenting a continuous dynamic process; The optimal segmentation policy is learned through the policy gradient algorithm to obtain the event segmentation boundary and the corresponding event label. The generative adversarial network uses the event label and random noise as input to synthesize virtual samples.

[0008] Furthermore, in the intelligent fault diagnosis method for wind power equipment described in this invention, step 3 includes: The segmented data fragments are input into a physical information neural network, the loss function of which includes data fitting loss and physical law residual loss. The physical information neural network is trained by backpropagation, and the network output is constrained to conform to the wind turbine motion equation and electromagnetic transient model. The hidden layer of the physical information neural network learns the subspace corresponding to the power grid command and the subspace of the inherent attributes of the mechanical components, which correspond to the forced dynamic response component and the inherent mechanical dynamic component, respectively. The inherent mechanical dynamic components are input into a variational autoencoder for compression and dimensionality reduction, and a low-dimensional component health index manifold is output.

[0009] Furthermore, in the intelligent fault diagnosis method for wind power equipment described in this invention, step 4 includes: The causal inference branch infers the causal direction between variables from the forced dynamic response components using a directional acyclic graph causal discovery algorithm; Based on the stated causal direction, a scalar is used to quantify the contribution of power grid commands to the observed data; The fault classification branch obtains fault classification features from the component health index through forward network processing; During training, the fault classification features are adversarially regularized through a gradient inversion layer, making the fault classification features uncorrelated with the contribution scalar.

[0010] Furthermore, in the intelligent fault diagnosis method for wind power equipment described in this invention, step 5 includes: Deploy digital twin models of wind turbines to receive real-time grid commands and operating setpoints, and simulate expected fault-free operation data; The Mahalanobis distance is obtained by performing a difference calculation between the actual operating data and the expected operating data from the digital twin simulation. The preliminary diagnostic results and the Mahalanobis distance are input into an integrated learning system, which includes multiple base models. The meta-learner is used to weight and fuse the outputs of multiple base models, where the meta-learner learns the optimal weight allocation of each base model output through training. A diagnostic report with confidence intervals is generated based on the fusion results. The fault probability is calculated by a weighted combination of the model output probability and the Mahalanobis distance.

[0011] Furthermore, the intelligent fault diagnosis method for wind power equipment described in this invention also includes: The diagnostic results and confidence scores obtained in step 5 are fed back to the reinforcement learning agent in step 2 to update the segmentation strategy parameters of the reinforcement learning agent. The causal discovery results obtained in step 4 are fed back to the physical information neural network in step 3 to adjust the weights of the physical equations in the physical information neural network.

[0012] Furthermore, the intelligent fault diagnosis method for wind power equipment described in this invention also includes: The diagnostic results and confidence scores from step 5 are fed back to the reinforcement learning agent from step 2 to update the weight parameters in the reward function. The causal discovery results in step 4 are fed back to the physical information neural network in step 3 to adjust the coefficient weights in the physical equations. The feedback mechanism enables the diagnostic system to adjust its diagnostic strategies based on changes in the power grid's operating mode.

[0013] Furthermore, the intelligent fault diagnosis method for wind power equipment described in this invention also includes: The virtual samples include data patterns with different command intensities and wind turbine operating points; The feature decoupling process separates the power grid command response from the component health status; Step 4 involves processing the power grid command response component in the causal inference branch and the component health status component in the fault classification branch.

[0014] Secondly, the present invention provides a wind power equipment intelligent fault diagnosis system, applied to the wind power equipment intelligent fault diagnosis method as described above, including a physical device and a control device, wherein the physical device and the control device establish a communication connection. The physical device includes: The wind turbine main controller is configured to acquire millisecond-level active power, drive train torque, and generator speed sequences via a high-frequency data bus; The wind farm energy management system is configured to receive and store timestamped grid dispatch command streams; A sensor array configured to collect wind turbine operating status data; The controlled device interface is configured to receive control commands and perform fault handling operations. The control device includes: The data synchronization module is configured to perform time synchronization processing on the running data stream and the power grid dispatch command stream, and generate a graph structure feature tensor including power commands, wind turbine state variables and power grid frequency measurements. The event segmentation module is configured to use a reinforcement learning agent, taking the instruction features in the graph structure feature tensor and the real-time power grid frequency deviation as state inputs, to dynamically determine the segmentation boundary of the power grid frequency regulation event, obtain the segmented data fragments, and generate virtual samples. The feature decoupling module is configured to input the segmented data fragments into the physical information neural network. Under the constraints of the mechanical equations of the wind turbine drive chain and the electromagnetic transient model of the generator, the observed data is decoupled into forced dynamic response components and inherent mechanical dynamic components, and the component health index is extracted. The model training module is configured to build a dual-branch diagnostic model, and the training of the dual-branch diagnostic model uses a gradient inversion layer to make the fault classification features uncorrelated with the power grid contribution. The diagnostic verification module is configured to deploy a digital twin model of the wind turbine, use a trained dual-branch diagnostic model to process the actual operating data, compare the preliminary diagnostic results with the expected data of the digital twin simulation, and generate a diagnostic report with confidence intervals. The control device is configured with its modules connected in sequence to form a data processing pipeline. The output of the data synchronization module is connected to the input of the event segmentation module, the output of the event segmentation module is connected to the input of the feature decoupling module, the output of the feature decoupling module is connected to the input of the model training module, and the output of the model training module is connected to the input of the diagnostic verification module.

[0015] Beneficial effects of this invention; The intelligent fault diagnosis method and system for wind power equipment provided by this invention effectively solves the problem of misjudgment caused by rapid power fluctuations due to wind turbines participating in grid frequency regulation through multi-step collaborative processing and closed-loop feedback mechanism. The method first generates graph structure feature tensors through time synchronization processing, providing a unified data foundation for subsequent analysis. It then uses reinforcement learning agents to dynamically segment events and generate virtual samples to enhance the model's generalization ability. Next, it uses a physical information neural network to decouple the grid command response from the component health status, isolating the impact of grid disturbances. The constructed dual-branch diagnostic model combines causal inference and fault classification branches, using a gradient inversion layer to decouple fault features from grid contribution, eliminating false associations. Finally, it uses a digital twin model to simulate fault-free expected data, and outputs a diagnostic report with confidence intervals through differential Mahalanobis distance and ensemble learning. An alarm is triggered only when the inherent mechanical dynamic components are abnormal and the confidence level is high, significantly improving diagnostic accuracy and reducing false alarms. Simultaneously, the feedback mechanism dynamically adapts to changes in grid operation modes, enhancing the system's robustness and practicality. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a flowchart of an intelligent fault diagnosis method for wind power equipment. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Firstly, please refer to Figure 1 The present invention provides an intelligent fault diagnosis method for wind power equipment, comprising: Step 1: Obtain the wind turbine's operating data stream and the grid dispatch command stream, perform time synchronization processing on the operating data stream and the command stream, and generate a graph structure feature tensor including power commands, wind turbine state variables and grid frequency measurements; Step 2: Using a reinforcement learning agent, the command features in the graph structure feature tensor and the real-time power grid frequency deviation are used as state inputs to dynamically determine the segmentation boundary of the power grid frequency regulation event, obtain the segmented data segments, and adaptively adjust the data window length. At the same time, virtual samples covering different command intensities and wind turbine operating points are generated. Step 3: Input the segmented data fragments into the physical information neural network. Under the constraints of the wind turbine drive chain mechanical equation and the generator electromagnetic transient model, the observed data is decoupled into the forced dynamic response component dominated by the grid command and the inherent mechanical dynamic component determined by the component health status. The component health index, which is independent of grid disturbance, is extracted by the variational autoencoder. Step 4: Construct a dual-branch diagnostic model. The causal inference branch identifies the causal path from the power grid command to the torque response from the forced dynamic response components and calculates the contribution. The fault classification branch processes the component health index. Train the dual-branch diagnostic model and use a gradient inversion layer to make the fault classification features uncorrelated with the power grid contribution. Step 5: Deploy the wind turbine digital twin model, receive real-time grid commands and operating setpoints, simulate expected fault-free operating data, process the actual operating data using a trained dual-branch diagnostic model, obtain preliminary diagnostic results, compare the preliminary diagnostic results with the expected data simulated by the digital twin, calculate the differential Mahalanobis distance, fuse the output probability and differential results of the diagnostic model, and generate a diagnostic report with confidence intervals. A fault alarm is triggered only when the abnormality of the inherent mechanical dynamic component exceeds the expected state and the overall confidence level is higher than the threshold.

[0020] In the intelligent fault diagnosis method for wind power equipment, millisecond-level active power, drivetrain torque, and generator speed sequences are first acquired through the high-frequency data bus of the wind turbine main controller, forming an operational data stream. Simultaneously, a timestamped grid dispatch command stream is received from the wind farm energy management system. A dynamic time warping algorithm is used to align the operational data stream and the grid dispatch command stream in time series, eliminating transmission delays and sampling rate differences, resulting in aligned data. A graph attention network is then used to jointly encode the power commands, wind turbine state variables, and grid frequency measurements in the aligned data, generating a spatiotemporally correlated graph structure feature tensor and storing it in a circular buffer. This step provides a unified and synchronized data foundation for subsequent analysis.

[0021] A reinforcement learning agent is used, with command features in a graph structure feature tensor and real-time grid frequency deviation as state inputs. The state space includes the current command value, historical command sequence, real-time grid frequency deviation, and wind turbine power change rate. The action space is defined to include extending the window, shortening the window, and confirming the start and end of events. A reward function is set: positive rewards are given when data within the segmented window can be explained by the wind turbine response physical model, and negative rewards are given when segmenting continuous dynamic processes. The optimal segmentation strategy is learned through a policy gradient algorithm to obtain event segmentation boundaries and corresponding event labels. Simultaneously, a generative adversarial network (GAN) uses event labels and random noise as input to synthesize virtual samples covering different command intensities and wind turbine operating points, thereby enhancing the model's adaptability to boundary scenarios.

[0022] The segmented data fragments are input into a physical information neural network (PIN), whose loss function includes data fitting loss and physical law residual loss. The PSN is trained via backpropagation, constraining the network output to conform to the wind turbine's motion equations and electromagnetic transient model. The hidden layers learn the subspace corresponding to grid commands and the subspace of inherent mechanical component properties, corresponding to the forced dynamic response component and the inherent mechanical dynamic component, respectively. The inherent mechanical dynamic component is input into a variational autoencoder for compression and dimensionality reduction, outputting a low-dimensional component health index manifold, achieving feature decoupling to isolate the impact of grid disturbances.

[0023] A dual-branch diagnostic model is constructed. The causal inference branch infers the causal direction between variables from the forced dynamic response components using a directional acyclic graph causal discovery algorithm, and quantifies the contribution scalar of power grid commands to the observed data based on the causal direction. The fault classification branch obtains fault classification features from the component health index through a feedforward network. During training, an adversarial regularization is applied to the fault classification features using a gradient inversion layer, making the fault classification features uncorrelated with the contribution scalar, thereby eliminating spurious correlations.

[0024] A digital twin model of the wind turbine is deployed to receive real-time grid commands and operating setpoints, simulating expected fault-free operation data. A trained bi-branch diagnostic model is used to process the actual operating data to obtain preliminary diagnostic results. The preliminary diagnostic results are then compared with the expected data simulated by the digital twin to calculate the Mahalanobis distance. The preliminary diagnostic results and the Mahalanobis distance are input into an integrated learning system, which includes multiple base models. A meta-learner is used to weight and fuse the outputs of multiple base models, where the meta-learner learns the optimal weight allocation for each base model output through training. A diagnostic report with confidence intervals is generated based on the fusion result. The fault probability is calculated by a weighted combination of the model output probability and the Mahalanobis distance. A fault alarm is triggered only when the anomaly of the inherent mechanical dynamic component exceeds the expected state and the overall confidence level is higher than a threshold, thus achieving diagnostic reliability.

[0025] Specifically, in the intelligent fault diagnosis method for wind power equipment described in this invention, step 1 includes: The high-frequency data bus of the wind turbine main controller acquires millisecond-level active power, transmission chain torque and generator speed sequences to form an operating data stream; Receive timestamped grid dispatch command streams from the wind farm energy management system; The running data stream and the power grid dispatch command stream are aligned using a dynamic time warping algorithm to eliminate transmission delay and sampling rate differences, resulting in aligned data. A graph attention network is used to jointly encode the power command, wind turbine state variables, and grid frequency measurements in the aligned data, generating a spatiotemporally correlated graph structure feature tensor and storing it in a circular buffer.

[0026] In the intelligent fault diagnosis method for wind power equipment, step 1 involves the acquisition and preprocessing of multimodal data. First, the wind turbine main controller acquires millisecond-level active power, drivetrain torque, and generator speed sequences in real time via a high-frequency data bus. These sequences constitute the operational data stream, with the high-frequency data bus ensuring real-time and accurate data acquisition. Simultaneously, the wind farm energy management system receives command streams from the power grid dispatch center. These command streams carry precise timestamps to record the occurrence time of power grid dispatch events. Due to their different sources, the operational data stream and the power grid dispatch command stream experience transmission delays and sampling rate differences, resulting in inconsistencies in their timelines.

[0027] To eliminate time inconsistencies, a dynamic time warping algorithm is employed to align the time series of operational data streams and power grid dispatch command streams. This algorithm dynamically stretches or compresses time series to find the optimal alignment path, thereby matching data points with different sampling rates and eliminating inherent delays and jitter. The aligned data preserves temporal causality, providing a consistent time benchmark for subsequent analysis.

[0028] The aligned data includes multivariate variables such as power commands, wind turbine state variables, and grid frequency measurements. These variables are jointly encoded using a graph attention network, where the graph structure uses variables as nodes and physical connections or statistical correlations as edges. The graph attention network dynamically captures the spatiotemporal dependencies between variables by calculating attention weights between nodes, generating a spatiotemporally correlated graph structure feature tensor. This graph structure feature tensor is ultimately stored in a circular buffer for rapid access and processing in subsequent steps, forming a complete data preprocessing pipeline.

[0029] Specifically, in the intelligent fault diagnosis method for wind power equipment of the present invention, step 2 includes: The instruction features and real-time grid frequency deviation in the graph structure feature tensor are used as the state input of the reinforcement learning agent, wherein the state space includes the current instruction value, historical instruction sequence, real-time grid frequency deviation and wind turbine power change rate; Define the action space including extending the window, shortening the window, and the start and end of the confirmation event; Set a reward function to give a positive reward when the data in the segmented window can be explained by the wind turbine response physical model, and give a negative reward when segmenting a continuous dynamic process; The optimal segmentation policy is learned through the policy gradient algorithm to obtain the event segmentation boundary and the corresponding event label. The generative adversarial network uses the event label and random noise as input to synthesize virtual samples.

[0030] In the intelligent fault diagnosis method for wind power equipment, step 2 involves event segmentation and virtual sample generation based on reinforcement learning. The reinforcement learning agent uses the command features in the graph structure feature tensor generated in step 1 and the real-time grid frequency deviation as state inputs. The state space specifically includes the current command value, historical command sequence, real-time grid frequency deviation, and wind turbine power change rate. These state variables collectively reflect the dynamic correlation between grid dispatch commands and wind turbine operating status, providing the agent with comprehensive environmental perception.

[0031] The action space is defined to include operations such as extending and shortening the window, and confirming the start and end of events, enabling the agent to flexibly adjust the data segmentation boundaries. The reward function is designed based on the degree of matching between the data within the segmentation window and the physical model of the wind turbine response. A positive reward is given when the data within the window can be effectively interpreted by the physical model, and a negative reward is given when the segmentation action interrupts the continuous dynamic process. This reward mechanism guides the agent to learn a segmentation strategy that adapts to the characteristics of grid frequency regulation events.

[0032] A reinforcement learning agent is trained using a policy gradient algorithm to optimize the segmentation policy with the goal of minimizing reward loss. After training, the agent outputs event segmentation boundaries and corresponding event labels, accurately identifying the start and end points of power grid frequency regulation events. These event labels serve as input to a generative adversarial network (GAN), which, combined with random noise, generates virtual samples covering different command intensities and wind turbine operating points. These virtual samples expand the training dataset, improving the model's generalization ability to rare operating conditions and forming a complete event adaptive segmentation and data augmentation process.

[0033] Specifically, in the intelligent fault diagnosis method for wind power equipment of the present invention, step 3 includes: The segmented data fragments are input into a physical information neural network, the loss function of which includes data fitting loss and physical law residual loss. The physical information neural network is trained by backpropagation, and the network output is constrained to conform to the wind turbine motion equation and electromagnetic transient model. The hidden layer of the physical information neural network learns the subspace corresponding to the power grid command and the subspace of the inherent attributes of the mechanical components, which correspond to the forced dynamic response component and the inherent mechanical dynamic component, respectively. The inherent mechanical dynamic components are input into a variational autoencoder for compression and dimensionality reduction, and a low-dimensional component health index manifold is output.

[0034] In the intelligent fault diagnosis method for wind power equipment, step 3 involves feature decoupling and state reconstruction based on a physical information neural network. The segmented data fragments obtained in step 2 are input into the physical information neural network. The network's loss function consists of data fitting loss and physical law residual loss. The data fitting loss measures the difference between the network output and the actual observed data, while the physical law residual loss assesses whether the network output conforms to physical constraints such as the wind turbine drivetrain mechanical equations and the generator electromagnetic transient model. The physical information neural network is trained using a backpropagation algorithm to minimize the total loss function, ensuring that the network output simultaneously satisfies data accuracy and physical consistency.

[0035] During training, the hidden layers of the physical information neural network learn two feature subspaces: the grid command correspondence subspace and the mechanical component inherent attribute subspace. The activity in the grid command correspondence subspace mainly reflects the dynamic response caused by grid dispatch commands, corresponding to the forced dynamic response component; the activity in the mechanical component inherent attribute subspace characterizes the health status of the wind turbine components, corresponding to the inherent mechanical dynamic component. This feature decoupling separates external grid disturbances from internal mechanical states, reducing interference.

[0036] The decoupled intrinsic mechanical dynamic components are input into a variational autoencoder for compression and dimensionality reduction. The encoder part of the variational autoencoder maps the high-dimensional intrinsic mechanical dynamic components to a low-dimensional latent space and learns their probability distribution; the decoder part reconstructs the data from the latent space, preserving key information. The output is a low-dimensional component health index manifold, which is continuous and smooth, facilitating subsequent processing by the fault classification module, thus forming a complete feature decoupling and state reconstruction process.

[0037] Specifically, in the intelligent fault diagnosis method for wind power equipment of the present invention, step 4 includes: The causal inference branch infers the causal direction between variables from the forced dynamic response components using a directional acyclic graph causal discovery algorithm; Based on the stated causal direction, a scalar is used to quantify the contribution of power grid commands to the observed data; The fault classification branch obtains fault classification features from the component health index through forward network processing; During training, the fault classification features are adversarially regularized through a gradient inversion layer, making the fault classification features uncorrelated with the contribution scalar.

[0038] In the intelligent fault diagnosis method for wind power equipment, step 4 involves the construction and training of a dual-branch diagnostic model. The causal inference branch, derived from the forced dynamic response components decoupled in step 3, uses a directed acyclic graph causal discovery algorithm to infer the causal direction between variables. This algorithm analyzes the conditional independence relationships between variables, identifies the causal path from grid commands to torque response, and clarifies the directionality of the causal relationship.

[0039] Based on the inferred causal direction, a scalar measure of the contribution of power grid commands to the observed data is quantified. The contribution scalar is characterized by calculating the strength or degree of influence of the causal path, for example, by using regression coefficients or causal effect estimation methods to quantify the causal impact of power grid commands into specific numerical values.

[0040] The fault classification branch obtains fault classification features from the component health index extracted in step 3 through a feedforward network. The feedforward network consists of multiple fully connected layers, which extract nonlinear features from the component health index layer by layer, and output a classification feature vector for fault identification.

[0041] During training, an adversarial regularization is applied to the fault classification features using a gradient inversion layer. This layer directly transmits the fault classification features during forward propagation and reverses the gradient direction during backward propagation, making the contribution scalar of the fault classification features unrelated to the causal inference branch calculation. This adversarial regularization forces the fault classification branch to ignore variations caused by power grid commands and focus on the degradation characteristics of the mechanical components themselves, thus improving the robustness of the diagnostic model.

[0042] Specifically, in the intelligent fault diagnosis method for wind power equipment of the present invention, step 5 includes: Deploy digital twin models of wind turbines to receive real-time grid commands and operating setpoints, and simulate expected fault-free operation data; The Mahalanobis distance is obtained by performing a difference calculation between the actual operating data and the expected operating data from the digital twin simulation. The preliminary diagnostic results and the Mahalanobis distance are input into an integrated learning system, which includes multiple base models. The meta-learner is used to weight and fuse the outputs of multiple base models, where the meta-learner learns the optimal weight allocation of each base model output through training. A diagnostic report with confidence intervals is generated based on the fusion results. The fault probability is calculated by a weighted combination of the model output probability and the Mahalanobis distance.

[0043] In the intelligent fault diagnosis method for wind power equipment, step 5 involves online diagnosis and decision confidence assessment assisted by digital twins. A digital twin model of the wind turbine is deployed. This model is built based on the turbine's design parameters and historical health data, receiving real-time grid commands and operating setpoints as inputs to simulate the expected operating data of the turbine under fault-free conditions. The expected operating data includes key parameters such as torque and speed, providing a benchmark reference for comparison with actual data.

[0044] Mahalanobis distance is obtained by calculating the difference between actual operating data and expected operating data from digital twin simulation. Mahalanobis distance is a distance metric that considers the covariance structure of data, quantifies the degree of deviation between actual and expected data, and reflects the correlation between variables, making it suitable for multidimensional data comparison.

[0045] The preliminary diagnostic results obtained in step 4 and the calculated Mahalanobis distance are input into the integrated learning system. The integrated learning system includes multiple base models, such as models based on vibration characteristics and models based on temperature characteristics. Each base model makes a preliminary judgment on the fault from a different perspective, providing diverse diagnostic outputs.

[0046] A meta-learner is used to weight and fuse the outputs of multiple base models. The meta-learner learns the optimal weight allocation for each base model's output during the training process, using historical diagnostic results and corresponding real fault labels as training data. The meta-learner employs a stacked generalization technique, using the base model outputs as meta-features, and learns how to combine these features to optimize prediction accuracy.

[0047] A diagnostic report with confidence intervals is generated based on the fusion results. The fault probability is calculated by a weighted combination of the model output probability and the Mahalanobis distance, with the weights determined during the training process. The confidence interval reflects the uncertainty of the diagnostic results; a fault alarm is triggered only when the abnormality of the inherent mechanical dynamic component exceeds the expected state and the overall confidence level is higher than a threshold, thus improving the reliability of diagnostic decisions.

[0048] Specifically, the intelligent fault diagnosis method for wind power equipment described in this invention further includes: The diagnostic results and confidence scores obtained in step 5 are fed back to the reinforcement learning agent in step 2 to update the segmentation strategy parameters of the reinforcement learning agent. The causal discovery results obtained in step 4 are fed back to the physical information neural network in step 3 to adjust the weights of the physical equations in the physical information neural network.

[0049] The intelligent fault diagnosis method for wind power equipment also includes a closed-loop feedback mechanism to enhance the system's adaptive capability. The diagnostic results and confidence scores obtained in step 5 are fed back to the reinforcement learning agent in step 2. The diagnostic results include fault classification output and confidence scores, which serve as environmental feedback signals for the reinforcement learning agent. The reinforcement learning agent updates the segmentation strategy parameters based on the diagnostic results and confidence scores, adjusts the action selection strategy through an online reinforcement learning algorithm, optimizes the decision accuracy of event segmentation boundaries, and makes event segmentation more consistent with actual operating conditions.

[0050] The causal discovery results obtained in step 4 are fed back to the physical information neural network in step 3. The causal discovery results include causal direction graphs between variables and contribution scalars. The physical information neural network uses these results to adjust the weights of the physical equations, specifically by modifying the weight coefficients in the physical law residual loss function, so that the network training pays more attention to physical constraints with strong causal relationships and improves the accuracy of feature decoupling.

[0051] The aforementioned feedback mechanism forms a two-way optimization loop. The reinforcement learning agent improves the event segmentation strategy through diagnostic feedback, while the physical information neural network optimizes the accuracy of physical modeling through causal feedback. The collaboration between the two enables the diagnostic system to dynamically adapt to changes in power grid dispatch commands and the evolution of wind turbine operating states, thereby improving the stability and adaptability of fault diagnosis.

[0052] Specifically, the intelligent fault diagnosis method for wind power equipment described in this invention further includes: The diagnostic results and confidence scores from step 5 are fed back to the reinforcement learning agent from step 2 to update the weight parameters in the reward function. The causal discovery results in step 4 are fed back to the physical information neural network in step 3 to adjust the coefficient weights in the physical equations. The feedback mechanism enables the diagnostic system to adjust its diagnostic strategies based on changes in the power grid's operating mode.

[0053] The intelligent fault diagnosis method for wind power equipment also includes a closed-loop feedback mechanism to enhance the system's adaptability. The diagnostic results and confidence scores obtained in step 5 are fed back to the reinforcement learning agent in step 2. The diagnostic results include fault classification output and confidence scores, which serve as environmental feedback signals. The reinforcement learning agent uses the diagnostic results and confidence scores to update the weight parameters in the reward function, adjusts the action reward values ​​through an online reinforcement learning algorithm, and optimizes the event segmentation strategy to make the segmentation boundaries more accurately adapt to the actual operating conditions.

[0054] The causal discovery results obtained in step 4 are fed back to the physical information neural network in step 3. The causal discovery results include causal direction graphs between variables and contribution scalars. The physical information neural network uses these results to adjust the coefficient weights in the physical equations, specifically by modifying the coefficients in the physical law residual loss function, so that the network training pays more attention to physical constraints with strong causal relationships and improves the feature decoupling accuracy.

[0055] The aforementioned feedback mechanism forms a two-way optimization loop. The reinforcement learning agent improves the event segmentation strategy through diagnostic feedback, while the physical information neural network optimizes the accuracy of physical modeling through causal feedback. The collaboration between the two enables the diagnostic system to dynamically adapt to changes in power grid dispatch commands and the evolution of wind turbine operating states, thereby improving the stability and adaptability of fault diagnosis.

[0056] Specifically, the intelligent fault diagnosis method for wind power equipment described in this invention further includes: The virtual samples include data patterns with different command intensities and wind turbine operating points; The feature decoupling process separates the power grid command response from the component health status; Step 4 involves processing the power grid command response component in the causal inference branch and the component health status component in the fault classification branch.

[0057] The virtual samples are synthesized by a generative adversarial network, covering data patterns of different grid dispatch command intensities and wind turbine operating points, including rapid power fluctuation scenarios under high command intensities and smooth adjustment scenarios under low command intensities, thereby enhancing the model's generalization ability to boundary conditions. The feature decoupling process is implemented through a physical information neural network. During the training process, the physical information neural network learns the subspace corresponding to the power grid command and the subspace of the inherent attributes of the mechanical components. The activity of the subspace corresponding to the power grid command corresponds to the power grid command response component, and the activity of the subspace of the inherent attributes of the mechanical components corresponds to the component health status component, thereby realizing the separation of the power grid command response and the component health status. In step 4, the causal inference branch uses a directional acyclic graph causal discovery algorithm to infer the causal direction between variables from the power grid command response component, while the fault classification branch obtains fault classification features from the component health status component through forward network processing, thus forming a collaborative diagnostic mechanism.

[0058] Secondly, the present invention provides a wind power equipment intelligent fault diagnosis system, applied to the wind power equipment intelligent fault diagnosis method as described above, including a physical device and a control device, wherein the physical device and the control device establish a communication connection. The physical device includes: The wind turbine main controller is configured to acquire millisecond-level active power, drive train torque, and generator speed sequences via a high-frequency data bus; The wind farm energy management system is configured to receive and store timestamped grid dispatch command streams; A sensor array configured to collect wind turbine operating status data; The controlled device interface is configured to receive control commands and perform fault handling operations. The control device includes: The data synchronization module is configured to perform time synchronization processing on the running data stream and the power grid dispatch command stream, and generate a graph structure feature tensor including power commands, wind turbine state variables and power grid frequency measurements. The event segmentation module is configured to use a reinforcement learning agent, taking the instruction features in the graph structure feature tensor and the real-time power grid frequency deviation as state inputs, to dynamically determine the segmentation boundary of the power grid frequency regulation event, obtain the segmented data fragments, and generate virtual samples. The feature decoupling module is configured to input the segmented data fragments into the physical information neural network. Under the constraints of the mechanical equations of the wind turbine drive chain and the electromagnetic transient model of the generator, the observed data is decoupled into forced dynamic response components and inherent mechanical dynamic components, and the component health index is extracted. The model training module is configured to build a dual-branch diagnostic model, and the training of the dual-branch diagnostic model uses a gradient inversion layer to make the fault classification features uncorrelated with the power grid contribution. The diagnostic verification module is configured to deploy a digital twin model of the wind turbine, use a trained dual-branch diagnostic model to process the actual operating data, compare the preliminary diagnostic results with the expected data of the digital twin simulation, and generate a diagnostic report with confidence intervals. The control device is configured with its modules connected in sequence to form a data processing pipeline. The output of the data synchronization module is connected to the input of the event segmentation module, the output of the event segmentation module is connected to the input of the feature decoupling module, the output of the feature decoupling module is connected to the input of the model training module, and the output of the model training module is connected to the input of the diagnostic verification module.

[0059] The intelligent fault diagnosis system for wind power equipment includes physical equipment and control devices, which establish a communication connection. The physical equipment is responsible for data acquisition and execution of operations, while the control device is responsible for data processing and fault diagnosis.

[0060] The wind turbine main controller in the physical equipment unit acquires millisecond-level active power, drivetrain torque, and generator speed sequences in real time via a high-frequency data bus; these sequences constitute the operational data stream. The wind farm energy management system receives command streams from the power grid dispatch center; these command streams are precisely timestamped to record power grid dispatch events. Sensor arrays collect wind turbine operating status data, such as vibration and temperature parameters. The controlled equipment interface receives control commands and executes fault handling operations, such as shutdown or alarm activation.

[0061] The data synchronization module in the control unit performs time synchronization processing on the operating data stream and the power grid dispatch command stream. The data synchronization module uses a dynamic time warping algorithm to align the time series, eliminate transmission delays and sampling rate differences, and generate a graph structure feature tensor that includes power commands, wind turbine state variables, and power grid frequency measurements. The graph structure feature tensor is generated through a graph attention network to capture the spatiotemporal dependencies between variables.

[0062] The event segmentation module utilizes a reinforcement learning agent, taking command features in a graph structure feature tensor and real-time grid frequency deviation as state inputs, to dynamically determine the segmentation boundaries for grid frequency regulation events. The state space of the reinforcement learning agent includes the current command value, historical command sequences, real-time grid frequency deviation, and wind turbine power change rate. The action space includes extending the window, shortening the window, and confirming the start and end of the event. The reward function is designed based on the degree of matching between the data within the segmentation window and the wind turbine response physical model. The event segmentation module outputs segmented data fragments and virtual samples.

[0063] The feature decoupling module inputs the segmented data fragments into the physical information neural network (PIN), performing feature decoupling under the constraints of the wind turbine drivetrain mechanical equations and the generator electromagnetic transient model. The loss function of the PSN includes data fitting loss and physical law residual loss, and backpropagation training ensures the network output conforms to physical constraints. The hidden layers learn the subspace corresponding to grid commands and the subspace of inherent mechanical component attributes, corresponding to the forced dynamic response component and the inherent mechanical dynamic component, respectively. The feature decoupling module outputs a component health index.

[0064] The model training module constructs a dual-branch diagnostic model, including a causal inference branch and a fault classification branch. The causal inference branch uses a directed acyclic graph causal discovery algorithm to infer the causal direction between variables from the forced dynamic response components and calculates the contribution scalar of the power grid command. The fault classification branch obtains fault classification features from the component health index through a feedforward network. During training, a gradient inversion layer is used to decouple the fault classification features from the contribution scalar.

[0065] The diagnostic verification module deploys a digital twin model of the wind turbine, receives real-time grid commands and operating setpoints, and simulates expected fault-free operation data. A trained bi-branch diagnostic model processes the actual operating data to obtain preliminary diagnostic results. These preliminary diagnostic results are compared with the expected data simulated by the digital twin, and the Mahalanobis distance is calculated. The preliminary diagnostic results and the Mahalanobis distance are input into an ensemble learning system, which includes multiple base models. A meta-learner performs weighted fusion of the base model outputs to generate a diagnostic report with confidence intervals.

[0066] The control unit's modules are sequentially connected to form a data processing pipeline. The output of the data synchronization module is connected to the input of the event segmentation module, the output of the event segmentation module is connected to the input of the feature decoupling module, the output of the feature decoupling module is connected to the input of the model training module, and the output of the model training module is connected to the input of the diagnostic verification module. This pipeline structure enables coherent data processing from acquisition to diagnosis, improving system efficiency.

[0067] This invention addresses the misjudgment problem caused by rapid power fluctuations resulting from wind turbine participation in grid frequency regulation through a multi-step collaborative processing approach. First, time synchronization aligns the wind turbine's operational data stream with the grid dispatch command stream, generating a graph structure feature tensor to eliminate transmission delays and sampling rate differences, providing a consistent data foundation for subsequent analysis. Next, a reinforcement learning agent dynamically determines the segmentation boundaries of grid frequency regulation events, identifying normal dynamic response events caused by grid dispatch and generating virtual samples to enhance the model's generalization ability. Then, a physical information neural network decouples the observed data into a forced dynamic response component dominated by grid commands and an inherent mechanical dynamic component determined by component health status, separating the impact of grid disturbances and extracting component health indices unrelated to the grid. A dual-branch diagnostic model is constructed: the causal inference branch quantifies the contribution of grid commands to the observed data, while the fault classification branch processes the component health index. A gradient inversion layer decouples fault classification features from contribution, eliminating spurious correlations. Finally, a digital twin model of the wind turbine is deployed to simulate expected fault-free operation data. The Mahalanobis distance is calculated by comparing the data with the actual data. The diagnostic results are fused to generate a report with confidence intervals. An alarm is triggered only when the inherent mechanical dynamic components are abnormal and the confidence level is high, thus avoiding misjudging normal dynamic responses as mechanical faults.

Claims

1. A method for intelligent fault diagnosis of wind power equipment, characterized in that, include: Step 1: Obtain the wind turbine's operating data stream and the grid dispatch command stream, perform time synchronization processing on the operating data stream and the command stream, and generate a graph structure feature tensor including power commands, wind turbine state variables and grid frequency measurements; Step 2: Using a reinforcement learning agent, the command features in the graph structure feature tensor and the real-time power grid frequency deviation are used as state inputs to dynamically determine the segmentation boundary of the power grid frequency regulation event, obtain the segmented data segments, and adaptively adjust the data window length. At the same time, virtual samples covering different command intensities and wind turbine operating points are generated. Step 3: Input the segmented data fragments into the physical information neural network. Under the constraints of the wind turbine drive chain mechanical equation and the generator electromagnetic transient model, the observed data is decoupled into the forced dynamic response component dominated by the grid command and the inherent mechanical dynamic component determined by the component health status. The component health index, which is independent of grid disturbance, is extracted by the variational autoencoder. Step 4: Construct a dual-branch diagnostic model. The causal inference branch identifies the causal path from the power grid command to the torque response from the forced dynamic response components and calculates the contribution. The fault classification branch processes the component health index. Train the dual-branch diagnostic model and use a gradient inversion layer to make the fault classification features uncorrelated with the power grid contribution. Step 5: Deploy the wind turbine digital twin model, receive real-time grid commands and operating setpoints, simulate expected fault-free operating data, process the actual operating data using a trained dual-branch diagnostic model, obtain preliminary diagnostic results, compare the preliminary diagnostic results with the expected data simulated by the digital twin, calculate the differential Mahalanobis distance, fuse the output probability and differential results of the diagnostic model, and generate a diagnostic report with confidence intervals. A fault alarm is triggered only when the abnormality of the inherent mechanical dynamic component exceeds the expected state and the overall confidence level is higher than the threshold.

2. The intelligent fault diagnosis method for wind power equipment according to claim 1, characterized in that, Step 1 includes: The high-frequency data bus of the wind turbine main controller acquires millisecond-level active power, transmission chain torque and generator speed sequences to form an operating data stream; Receive timestamped grid dispatch command streams from the wind farm energy management system; The running data stream and the power grid dispatch command stream are aligned using a dynamic time warping algorithm to eliminate transmission delay and sampling rate differences, resulting in aligned data. A graph attention network is used to jointly encode the power command, wind turbine state variables, and grid frequency measurements in the aligned data, generating a spatiotemporally correlated graph structure feature tensor and storing it in a circular buffer.

3. The intelligent fault diagnosis method for wind power equipment according to claim 2, characterized in that, Step 2 includes: The instruction features and real-time grid frequency deviation in the graph structure feature tensor are used as the state input of the reinforcement learning agent, wherein the state space includes the current instruction value, historical instruction sequence, real-time grid frequency deviation and wind turbine power change rate; Define the action space including extending the window, shortening the window, and the start and end of the confirmation event; Set a reward function that gives a positive reward when the data within the segmented window can be explained by the wind turbine response physical model, and gives a negative reward when segmenting a continuous dynamic process; The optimal segmentation policy is learned through the policy gradient algorithm to obtain the event segmentation boundary and the corresponding event label. The generative adversarial network uses the event label and random noise as input to synthesize virtual samples.

4. The intelligent fault diagnosis method for wind power equipment according to claim 3, characterized in that, Step 3 includes: The segmented data fragments are input into a physical information neural network, the loss function of which includes data fitting loss and physical law residual loss. The physical information neural network is trained by backpropagation, and the network output is constrained to conform to the wind turbine motion equation and electromagnetic transient model. The hidden layer of the physical information neural network learns the subspace corresponding to the power grid command and the subspace of the inherent attributes of the mechanical components, which correspond to the forced dynamic response component and the inherent mechanical dynamic component, respectively. The inherent mechanical dynamic components are input into a variational autoencoder for compression and dimensionality reduction, and a low-dimensional component health index manifold is output.

5. The intelligent fault diagnosis method for wind power equipment according to claim 4, characterized in that, Step 4 includes: The causal inference branch infers the causal direction between variables from the forced dynamic response components using a directional acyclic graph causal discovery algorithm; Based on the stated causal direction, a scalar is used to quantify the contribution of power grid commands to the observed data; The fault classification branch obtains fault classification features from the component health index through forward network processing; During training, the fault classification features are adversarially regularized through a gradient inversion layer, making the fault classification features uncorrelated with the contribution scalar.

6. The intelligent fault diagnosis method for wind power equipment according to claim 5, characterized in that, Step 5 includes: Deploy digital twin models of wind turbines to receive real-time grid commands and operating setpoints, and simulate expected fault-free operation data; The Mahalanobis distance is obtained by performing a difference calculation between the actual operating data and the expected operating data from the digital twin simulation. The preliminary diagnostic results and the Mahalanobis distance are input into an integrated learning system, which includes multiple base models. The meta-learner is used to weight and fuse the outputs of multiple base models, where the meta-learner learns the optimal weight allocation of each base model output through training. A diagnostic report with confidence intervals is generated based on the fusion results. The fault probability is calculated by a weighted combination of the model output probability and the Mahalanobis distance.

7. The intelligent fault diagnosis method for wind power equipment according to claim 6, characterized in that, Also includes: The diagnostic results and confidence scores obtained in step 5 are fed back to the reinforcement learning agent in step 2 to update the segmentation strategy parameters of the reinforcement learning agent. The causal discovery results obtained in step 4 are fed back to the physical information neural network in step 3 to adjust the weights of the physical equations in the physical information neural network.

8. The intelligent fault diagnosis method for wind power equipment according to claim 7, characterized in that, Also includes: The diagnostic results and confidence scores from step 5 are fed back to the reinforcement learning agent from step 2 to update the weight parameters in the reward function. The causal discovery results in step 4 are fed back to the physical information neural network in step 3 to adjust the coefficient weights in the physical equations. The feedback mechanism enables the diagnostic system to adjust its diagnostic strategies based on changes in the power grid's operating mode.

9. The intelligent fault diagnosis method for wind power equipment according to claim 8, characterized in that, Also includes: The virtual samples include data patterns with different command intensities and wind turbine operating points; The feature decoupling process separates the power grid command response from the component health status; Step 4 involves processing the power grid command response component in the causal inference branch and the component health status component in the fault classification branch.

10. A wind power equipment intelligent fault diagnosis system, applied to the wind power equipment intelligent fault diagnosis method as described in any one of claims 1 to 9, characterized in that, It includes physical equipment and control devices, wherein the physical equipment and control devices establish a communication connection. The physical device includes: The wind turbine main controller is configured to acquire millisecond-level active power, drive train torque, and generator speed sequences via a high-frequency data bus; The wind farm energy management system is configured to receive and store timestamped grid dispatch command streams; A sensor array configured to collect wind turbine operating status data; The controlled device interface is configured to receive control commands and perform fault handling operations. The control device includes: The data synchronization module is configured to perform time synchronization processing on the running data stream and the power grid dispatch command stream, and generate a graph structure feature tensor including power commands, wind turbine state variables and power grid frequency measurements. The event segmentation module is configured to use a reinforcement learning agent, taking the instruction features in the graph structure feature tensor and the real-time power grid frequency deviation as state inputs, to dynamically determine the segmentation boundary of the power grid frequency regulation event, obtain the segmented data fragments, and generate virtual samples. The feature decoupling module is configured to input the segmented data fragments into the physical information neural network. Under the constraints of the mechanical equations of the wind turbine drive chain and the electromagnetic transient model of the generator, the observed data is decoupled into forced dynamic response components and inherent mechanical dynamic components, and the component health index is extracted. The model training module is configured to build a dual-branch diagnostic model, and the training of the dual-branch diagnostic model uses a gradient inversion layer to make the fault classification features uncorrelated with the power grid contribution. The diagnostic verification module is configured to deploy a digital twin model of the wind turbine, use a trained dual-branch diagnostic model to process the actual operating data, compare the preliminary diagnostic results with the expected data of the digital twin simulation, and generate a diagnostic report with confidence intervals. The control device is configured with its modules connected in sequence to form a data processing pipeline. The output of the data synchronization module is connected to the input of the event segmentation module, the output of the event segmentation module is connected to the input of the feature decoupling module, the output of the feature decoupling module is connected to the input of the model training module, and the output of the model training module is connected to the input of the diagnostic verification module.

Images