Wind power plant unit state monitoring and fault early warning system and method based on deep learning

Abstract

The invention provides a wind power plant unit state monitoring and fault early warning system and method based on deep learning, and belongs to the field of wind power generation and artificial intelligence. According to the system, a cloud edge collaborative architecture is adopted, an edge computing terminal operates a data-driven space-time prediction model and a physical digital twinborn model in parallel, and abnormity is preliminarily screened by calculating a double-track residual error and comparing the double-track residual error with a dynamic early warning threshold value. And when an exception occurs, the cloud platform receives multi-modal data including a sensor, a model state and an operation and maintenance text, performs deep root cause analysis by using a diagnosis model fused with a wind power fault knowledge graph, and generates an interpretable diagnosis report. According to the method, deep fusion of data and a physical model is realized, and the accuracy of fault monitoring, the interpretability of diagnosis and the intelligent level of operation and maintenance decision are remarkably improved through a data-physical double-track driving mode.

Description

Technical Field

[0001] This invention relates to the intersection of wind power generation technology and artificial intelligence technology, and in particular to a system and method for condition monitoring, fault diagnosis and predictive maintenance of wind turbine generator sets using deep learning models and cloud-edge collaborative architecture. Background Technology

[0002] Wind turbines are typically deployed in harsh environments, where critical components such as gearboxes, generators, and blades are subjected to immense mechanical stress and environmental corrosion over long periods, leading to frequent failures. These failures not only incur high maintenance and replacement costs but also cause significant power generation losses, impacting the overall efficiency of wind farms. Therefore, developing a system capable of precise condition monitoring and intelligent fault early warning for wind turbines, enabling predictive maintenance, has become a pressing technical challenge in this field.

[0003] To address the aforementioned issues, various wind farm fault early warning schemes have been disclosed in the existing technology.

[0004] For example, Chinese patent CN115809866A discloses a method for early warning by obtaining the "state deviation" of a wind turbine unit through machine learning. While this method considers both the unit's own time-dimensional data and the spatial-dimensional data of adjacent units, its machine learning model is broadly defined, and the simple deviation calculation fails to capture the complex and nonlinear dynamic relationships within the wind field, resulting in limited fault diagnosis capabilities. Furthermore, its fully centralized data processing architecture faces severe challenges in terms of communication bandwidth and computational latency when dealing with massive amounts of high-frequency data across the entire wind field.

[0005] For example, Chinese patent CN114154567A discloses a method for anomaly identification based on historical data. Before training the model, this method requires pre-labeling the data as "normal" or "abnormal" using physical mechanisms, statistical methods, and clustering algorithms. The bottleneck of this approach lies in the complexity of the pre-labeling process, its high dependence on prior knowledge, its poor generalization ability for unknown or novel fault modes, and its primary identification of statistical anomalies at the data level rather than the root causes of physical faults in the equipment.

[0006] For example, Chinese patent CN116090626A creatively proposed a "cloud-edge collaborative" computing architecture, deploying edge computing terminals on the turbine side to execute real-time tasks, while the cloud platform is responsible for model training and global optimization. This architecture effectively alleviates data transmission latency and cloud computing load. However, the algorithm models used in this solution (such as ARIMA and KPCA) are relatively traditional, and for highly complex and nonlinear spatiotemporally coupled systems like wind turbines, there is still considerable room for improvement in the model's expressive power and prediction accuracy.

[0007] In summary, existing technologies for wind farm condition monitoring suffer from drawbacks such as static spatiotemporal correlation characterization, poor interpretability of fault diagnosis, lagging digital twin calibration, rigid dynamic (RUL) early warning thresholds, and shallow multimodal data fusion. These shortcomings make it difficult to adapt to the complex operating conditions and intelligent operation and maintenance needs of wind farms. Therefore, a new technology is urgently needed to solve the problems of existing technologies. Summary of the Invention

[0008] This invention provides a method for wind farm unit condition monitoring and fault early warning based on deep learning and its application, which addresses the problems of current technology such as single model association characterization, poor interpretability of fault root cause diagnosis, lagging physical model adaptation, rigid operation and maintenance decision-making, and shallow multimodal data fusion.

[0009] The core technology of this invention mainly adopts a two-level computing architecture of "cloud-edge collaboration". It realizes real-time status monitoring and anomaly screening through a multi-factor dynamic graph spatiotemporal neural network (ST-GNN). It combines a Transformer model that integrates wind power fault knowledge graph to complete fault root cause localization. It is equipped with digital twin real-time calibration, dual-track residual deviation correction, dynamic RUL threshold calculation and multimodal cross-attention fusion mechanism to build a closed-loop wind turbine status monitoring and fault early warning system of "monitoring-early warning-diagnosis-operation and maintenance".

[0010] In a first aspect, the present invention provides a method for wind farm turbine condition monitoring and fault early warning based on deep learning, the method comprising the following steps: Real-time monitoring steps at the edge: S1: Obtain real-time operating data of the wind turbine at the edge computing terminal deployed on the wind turbine; S2: Based on the edge computing terminal, the data-driven spatiotemporal prediction model and the physical mechanism-based digital twin model are run in parallel to obtain the data prediction state and the physical prediction state, respectively. S3: Calculate the first residual between the real-time running data and the data prediction state, and the second residual between the real-time running data and the physical prediction state, and combine the first residual and the second residual into a dual-track residual; S4: Compare the dual-track residual with the dynamic early warning threshold to determine abnormal events; Cloud-based in-depth diagnostic steps: S5: When an abnormal event is detected, the edge computing terminal uploads multimodal abnormal data, which includes at least two of the sensor data, model status, and operation and maintenance text, to the cloud platform. S6: On the cloud platform, a diagnostic model that integrates a fault knowledge graph in the wind power field is used to analyze multimodal anomaly data in order to locate the root cause of the fault. S7: Generate a fault diagnosis report based on the root cause of the fault.

[0011] Furthermore, the data-driven spatiotemporal prediction model is a spatiotemporal graph neural network model, in which the wind farm is constructed as a dynamic topology graph, and the weights of the connection edges between units in the dynamic topology graph are dynamically adjusted according to at least one of the wake effect, the degree of unit aging, and the maintenance history.

[0012] Furthermore, the dynamic early warning threshold is calculated based on a preset maintenance decision objective function, which comprehensively considers the estimated power generation loss and maintenance costs.

[0013] Furthermore, the diagnostic model is an improved Transformer model that includes a cross-modal attention layer for calculating the association weights between different modalities in multimodal anomaly data.

[0014] Furthermore, it also includes a real-time calibration step for the digital twin model: When the deviation between the physical prediction state and the real-time running data exceeds the preset calibration threshold, the edge computing terminal is triggered to upload high-resolution data to the cloud platform. The cloud platform optimizes the key parameters of the digital twin model and sends the optimized parameters back to the edge computing terminal for updating.

[0015] Furthermore, it also includes a closed-loop optimization step: feeding the actual maintenance results based on the fault diagnosis report as new cases back to the cloud platform to iteratively update the diagnostic model and knowledge graph.

[0016] Secondly, the present invention provides a wind farm turbine condition monitoring and fault early warning system based on deep learning, comprising: An edge computing terminal, deployed on a wind turbine, is configured to perform the edge-end real-time monitoring steps of claim 1; A cloud platform that communicates with an edge computing terminal is configured to perform the cloud-based deep diagnostic steps of claim 1.

[0017] Furthermore, the data-driven spatiotemporal prediction model is a spatiotemporal graph neural network model, in which the wind farm is constructed as a dynamic topology graph, and the weights of the connection edges between the units in the dynamic topology graph are dynamically adjusted according to at least one of the wake effect, the degree of unit aging, and the maintenance history.

[0018] Furthermore, the system is configured to calculate dynamic early warning thresholds based on a preset maintenance decision objective function, which comprehensively considers the estimated power generation loss and maintenance costs.

[0019] Furthermore, the cloud platform and edge computing terminals are also configured to perform a real-time calibration step for the digital twin model, which includes: When the deviation between the physical prediction state and the real-time running data exceeds a preset calibration threshold, the edge computing terminal is triggered to upload high-resolution data to the cloud platform. The cloud platform optimizes the key parameters of the digital twin model and sends the optimized parameters back to the edge computing terminal for updating.

[0020] The main contributions and innovations of this invention are as follows: 1. Higher precision spatiotemporal correlation characterization and anomaly detection: The multi-factor dynamic edge weight spatiotemporal graph neural network (ST-GNN) used in this invention can dynamically adjust the correlation weight between units according to actual operating conditions such as wind speed and unit aging. Compared with the simple deviation calculation or traditional model in the prior art, it can more accurately capture the complex spatiotemporal dynamic pattern of the wind field, improve the early weak fault detection rate by 25%, and reduce the false alarm rate to below 4%.

[0021] 2. Enhanced Interpretability of Root Cause Diagnosis: This invention deeply integrates the Transformer diagnostic model with a wind power fault knowledge graph and a cross-modal attention mechanism. It not only outputs the fault type but also pinpoints the key sensor combinations and critical time points that triggered the fault, and correlates them with relevant maintenance documentation. Compared to the "black box" diagnostic methods of existing technologies, its root cause localization misjudgment rate is reduced from 8% to below 3%, significantly improving the interpretability and decision support value of the diagnostic report.

[0022] 3. More timely physical model adaptation: The invention's unique edge-end digital twin real-time calibration mechanism can quickly coordinate with the cloud to complete parameter optimization when the unit encounters sudden changes in operating conditions such as gusts of wind. Compared with existing technologies that use physical models with fixed parameters or delayed cloud updates, its physical residual calculation error is reduced from 15% to below 8%, and its adaptability to sudden operating conditions is improved by 60%.

[0023] 4. Smarter O&M Closed-Loop Optimization: This invention dynamically calculates early warning thresholds based on a maintenance decision objective function and prioritizes maintenance based on fault severity. Compared to fixed early warning thresholds in existing technologies, it effectively avoids over-maintenance or delayed maintenance, achieving optimal allocation of O&M resources, reducing wind farm O&M costs by 25% and power generation losses by 15%.

[0024] 5. Deeper Multimodal Data Fusion: This invention, through a cross-modal attention layer, can uncover the deep intrinsic connections between sensor data, operational text, and digital twin model status. Compared to the simple feature stitching of existing technologies, its diagnostic accuracy for "text-sensor" correlated faults is improved from 85% to 96%.

[0025] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a deep learning-based wind farm turbine condition monitoring and fault early warning method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a web-based display according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a heat map according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a snapshot of associated data according to an embodiment of the present invention. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0028] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0029] Example 1 This invention aims to propose a deep learning-based method for wind farm turbine condition monitoring and fault early warning. Specifically, it refers to... Figure 1 The method employs a two-stage monitoring and diagnostic strategy.

[0030] Phase 1: Real-time state monitoring and anomaly detection based on ST-GNN (edge ​​execution) Step 1: Constructing a dynamic wind farm map (cloud-based): In the cloud, the entire wind farm is abstracted as a dynamic graph G=(V,E,X). Where: Each node v∈V represents a wind turbine unit; Node feature X is a multidimensional time series of all sensor measurements for the unit; Edge e∈E represents the spatial association between units, and its weight is calculated using a multi-factor dynamic function:

[0031] in: The original wake function (Jensen / Larsen model) is incorporating the wind speed v(t) gradient (the wake effect is enhanced by 20% when the wind speed is >12 m / s); specifically: : The straight-line distance between unit i and unit j (unit: meters). The closer the distance, the stronger the wake effect of i on j. The larger the value. : Wind direction at time t (unit: degrees). The wake will only affect j when the wind direction is from i to j (j is downwind of i); if the wind direction is opposite (j is upwind of i). The value is close to 0 (no correlation). v(t): Average wind speed at time t (unit: m / s).

[0032] It is evident that the greater the wind speed, the greater the intensity and range of the wake. For example, "when the wind speed > 12 m / s, the wake's influence increases by 20%." The value will be multiplied by a coefficient of 1.2. It is based on the classic Jensen / Larsen wake model in the wind farm field, incorporating dynamic adjustments for real-time wind speed and direction. For example: when =500m (much larger than the wake's influence range), when the angle between the wind direction and the i→j direction is >90° =0.1 (weak association); when =100m (in the wake core region), wind direction is the same as i→j, wind speed =15m / s =0.8 (strong correlation).

[0033] The unit aging factor is adjusted based on the unit's operating time (the weighting coefficient for units >5 years increases by 0.3) and the historical failure count of key components (gearbox, main bearing); specifically: : The cumulative operating time of unit i (in years), and the number of historical failures of key components (gearbox, main bearing); : The cumulative operating time of unit j and the number of historical failures of key components.

[0034] : To maintain the correlation factor, if both unit i and j complete maintenance within 3 months, the weighting coefficient is reduced by 0.1 (to reduce short-term correlation misjudgments after maintenance); specifically: The most recent maintenance time (e.g., "gearbox lubricant changed on 2024-05-10") and maintenance type of unit i; : The time and type of the most recent maintenance for unit j.

[0035] : Dynamic coefficients (summing up to 1), optimized in real time through cloud-based self-supervised training (e.g., α increases to 0.6 during windy weather, β increases to 0.4 when the proportion of aging units is >30%, and during peak maintenance periods (e.g., within one month after quarterly maintenance). Upgraded to 0.2 (significant maintenance impact)).

[0036] The edge weights of the dynamic graph are updated every minute.

[0037] Step 2: Supervised Model Training (Cloud-based): The ST-GNN model is trained based on massive amounts of historical data under normal operating conditions. This model integrates: Graph Attention Networks (GAT) capture multi-factor dynamic spatial dependencies among units; The gated recurrent unit (GRU) captures the temporal evolution of each sensor sequence; Training objective: Given the state of the entire field graph at time T in the past, accurately predict the state of the entire field graph at time T+k in the future, without the need for manual annotation of fault data.

[0038] Through this self-supervised approach, the model can autonomously learn the complex spatiotemporal dynamic patterns of wind farms under normal operation without the need for manual annotation of fault data.

[0039] Step 3: Edge Real-Time Inference and Digital Twin Calibration: ① The ST-GNN model trained in the cloud is quantized and pruned and then deployed to the edge terminal. At the same time, a lightweight digital twin simplified model is distributed (the core physical equations are retained, and the parameter dimension is reduced to 15+, such as friction coefficient, thermal resistance, and fatigue coefficient). ② The edge terminal collects real-time data from the local unit and the status of adjacent units, and inputs it into the ST-GNN model to obtain data-driven theoretical normal prediction values; ③ Synchronously run the lightweight digital twin model, input real-time operating parameters (speed, torque, etc.) to obtain the theoretical normal prediction value of physical drive; ④ Digital twin real-time calibration trigger: When the physical residual (X) is triggered... actual - X predicted_physics If the dynamic threshold is exceeded by more than 1.2 times for 5 consecutive seconds, or if the sensor data changes abruptly (such as a sudden increase in wind speed of 5 m / s), the edge terminal will upload high-resolution data (1 time / time) of the abrupt change period to the cloud. The cloud will then output the optimized key parameters through Bayesian optimization (iterations ≤ 20 times, time ≤ 30 seconds) and send them to the edge terminal to update the simplified model.

[0040] Step 4: Dual-track residual analysis and initial screening of dynamic threshold anomalies: ① Calculate the dual-track residual: Data residuals = |X actual - X predicted_data | (ST-GNN output); Physical residual = |X actual - X predicted_physics (Digital twin output); ② Calculation of Dynamic (RUL) Early Warning Threshold: Calculate the optimal threshold based on the maintenance decision objective function: Cost(t)=λ1·Loss(t)+λ2·Maint(t)−λ3·Power(t) In the formula: Loss(t) is the power generation loss due to fault when RUL=t, Main(t) is the maintenance cost at time t, Power(t) is the predicted power generation of the wind farm at time t, and λ1 / λ2 / λ3 are dynamic weights; Take the RUL value corresponding to the minimum value of Cost(t) as the warning threshold (e.g., the threshold for high power generation period is increased from 100h to 150h). ③ Anomaly detection: When the fusion norm (such as weighted Euclidean distance) of the dual-track residual exceeds the dynamic threshold for 3 consecutive seconds, it is determined to be an abnormal event and triggers the second stage.

[0041] Phase Two: Intelligent Fault Diagnosis and Root Cause Analysis Based on Transformer (Executed in the Cloud) Step 5: Upload multimodal anomaly data: The edge terminal packages and uploads high-resolution data (1200 time points × 100+ dimensional sensor data) from 10 minutes before and after the anomaly, synchronized digital twin status, and recent maintenance text (such as "2024-05-10 Replaced gearbox lubricant") to the cloud platform, with all data aligned by timestamp.

[0042] Step Six: Cross-modal attention fusion and deep diagnostics: ① The cloud platform has launched an improved Transformer diagnostic model, which adds a cross-modal attention layer and a knowledge graph embedding sub-network: Cross-modal attention layer: Calculates weights between different modalities (e.g., the weight of the oil temperature sensor corresponding to the text "lubricating oil replacement" is increased to 0.6, and the weight of the vibration Z-axis corresponding to the digital twin "transmission efficiency decrease" is increased to 0.5), replacing traditional feature stitching; Knowledge Graph Embedded Sub-Network: Construct a "Wind Power Fault Knowledge Graph" (entities: components / sensors / fault types; relations: "fault caused" "sensor association"), and use the TransE model to map entities / relationships into low-dimensional vectors, which are then concatenated with cross-modal features; ②The Transformer encoder introduces a knowledge constraint term: If the attention weight is directed to a sensor combination with no fault propagation relationship (such as "wind speed sensor → gearbox oil temperature"), the weight is reduced by a penalty term, and the focus is on the association that conforms to the domain rules.

[0043] Step 7: Root cause identification, maintenance priority ranking, and explanation report generation: ①Model output: Fault category probability (e.g., "insufficient gearbox lubricating oil pressure": 0.89); Heatmap of key sensor contributions (e.g., "pitch motor 1 current" and "hydraulic station pressure" have the highest weights); Critical time points (e.g., coordination anomalies at XX:XX). ② Maintenance Priority Ranking: The cloud generates a maintenance queue based on the Cost(t) of each unit and the severity of the fault (gearbox fault priority > sensor fault), and outputs the optimal maintenance window period (e.g., if the wind speed is predicted to be low the next day, maintenance will be scheduled for that period first). ③ Diagnostic report: such as "Warning: Hydraulic fault in the pitch system of Unit 23 (confidence level 96%). Root cause: The current of pitch motor 1 and the pressure of the hydraulic station were abnormally synchronized at XX:XX, and the associated maintenance record '2024-05-05 hydraulic oil not replaced as scheduled'. Recommendation: Check the hydraulic system and replace the hydraulic oil and pressure sensor between 9:00-12:00 the next day (during low wind speed period)".

[0044] It is worth mentioning that the core idea of ​​this invention lies in the "cloud-edge collaborative dual-track monitoring and multimodal diagnostic framework", and ST-GNN and Transformer are the preferred methods to implement this framework.

[0045] Example 2 This embodiment is based on Embodiment 1. The present invention will be described in detail below with reference to a specific embodiment.

[0046] Scenario: A wind farm containing 50 3MW wind turbine units.

[0047] Step 1: System Deployment and Initialization An embedded edge computing terminal is installed in the control cabinet at the base of each unit tower. Terminal deployment: ① Lightweight ST-GNN model (INT8 quantization); ② Lightweight digital twin simplified model (core physical equations of gearbox / main bearing, parameters: friction coefficient, thermal resistance, fatigue coefficient and 15 other items); ③ Dynamic threshold calculation module and maintenance priority judgment module; The terminal connects to the PLC via industrial Ethernet (for second-level / sub-second-level data reading) and connects to the cloud via 5G / 4G, where a wind power fault knowledge graph (containing 50+ entities and 80+ relationships) is deployed.

[0048] Step 2: Cloud-based model training Graph construction: Collect data from the entire dataset for one year and calculate the dynamic edge weights w. ij (t) (α=0.5, β=0.3, γ=0.2; 12 aging units (operating for >5 years) (The coefficient increases by 0.3). ST-GNN training: Input the state of the entire field over the past 60 minutes, predict the state over the next 5 minutes, train 2 layers of GAT + 3 layers of GRU, use mean squared error as the loss function, and deploy quantized after convergence. Transformer training: Integrating cross-modal attention layer and knowledge graph embedding sub-network, training data includes historical fault cases + simulated fault data + operation and maintenance text (1000+ entries), training objectives: fault classification + sensor contribution regression; Digital twin full model training: Ansys is used to build a gearbox multiphysics model, Bayesian optimization of calibration parameters (such as thermal resistance h=12W / (m·K)), and simplified before being sent to the edge.

[0049] Step 3: Model Deployment and Edge Monitoring Edge terminals enter real-time monitoring loop: Collects data from 100+ measuring points per second to obtain the status of adjacent units; Run the ST-GNN and digital twin model to output dual-track predictions; Calculate the dual-track residual and calculate the dynamic RUL threshold based on Cost(t) (if the predicted power generation is high on the day, the threshold is increased from 100h to 150h). If the residual exceeds the threshold for 3 consecutive seconds, an anomaly report will be triggered (e.g., an anomaly is detected by the edge terminal of Unit 23).

[0050] Step 4: Anomaly Reporting and Cloud-based Diagnosis Uploaded by terminal 23: Data for 10 minutes before and after the anomaly (1200 time points × 100+ dimensions), digital twin status (transmission efficiency decreased by 5%), and maintenance text ("Hydraulic oil not changed on 2024-05-05"). Cloud-based Transformer model: Cross-modal attention layer: The text "Hydraulic oil not changed" corresponds to a pressure sensor weight of 0.6 for the hydraulic station, and the digital twin "transmission efficiency decreased" corresponds to a vibration Z-axis weight of 0.5. Knowledge constraint: Eliminate invalid associations such as "wind speed sensor → hydraulic failure"; Model output: Fault category: "Pitch system hydraulic failure" (0.96); Key sensors: Pitch motor 1 current, hydraulic station pressure; Key time point: XX:XX; Maintenance priority ranking: Cost(t) calculation shows that the maintenance cost is lowest from 9:00 to 12:00 the next day (wind speed <5m / s), so it is ranked first. Diagnostic report: "Hydraulic fault in the pitch system of Unit 23 (confidence level 96%). Root cause: The current of pitch motor 1 and the pressure of the hydraulic station were abnormally synchronized at XX:XX, and the associated maintenance record '2024-05-05 hydraulic oil not replaced as scheduled' was found. Recommendation: Check the hydraulic system from 9:00 to 12:00 the next day, and replace the hydraulic oil and pressure sensor." Step 5: Closed-loop optimization Maintenance personnel perform repairs according to the report (replacing hydraulic oil and pressure sensors) and report the repair results to the cloud. The cloud-based system adds this case study to the Transformer training set and fine-tunes the model. Every two weeks: Train the ST-GNN with the incremental new data and update the dynamic edge weights α / β / γ; The full digital twin model is calibrated using all the data, and a simplified model is then distributed to the edge. Optimize and maintain the weights λ1 / λ2 / λ3 of the decision objective function.

[0051] Example 3 To further improve the system's prediction accuracy, physical interpretability, and robustness to unknown operating conditions, a physical mechanism-driven digital twin model is introduced based on Example 1 and deeply integrated with a data-driven AI model to construct a "twin-data" hybrid-driven monitoring and remaining useful life prediction system. The specific steps are as follows: Step 1: Building and calibrating component-level digital twins (cloud + edge): ① Model building: For key components such as gearbox and main bearing, multiphysics models (such as gearbox thermal balance equations) are built using Simulink / Ansys. ② Personalized calibration: Bayesian optimization and calibration of all model parameters are performed in the cloud, and a simplified model is deployed at the edge and real-time calibration is performed (see step three of the first stage of Example 1). ③ Dynamic spatiotemporal graph construction and ST-GNN training: Same as steps one and two of the first stage above (multi-factor dynamic edge weights).

[0052] Preferably, since component-level digital twins are the core foundation of "physical residual calculation," a balance needs to be struck between "model accuracy" and "engineering practicality." To achieve this, the following operations are required: (a) Screening criteria for core physical parameters (taking gearbox as an example) (1) Screening principles Based on the "priority of physical mechanisms affecting the core operating state of components" and the "lightweight requirements of edge devices", a three-dimensional screening model of "mechanism necessity-data sensitivity-computational complexity" is adopted to eliminate secondary parameters that have less than 5% impact on output accuracy and are computationally time-consuming, while retaining key parameters.

[0053] (2) Details of 15 core physical parameters of the gearbox and selection criteria (see Tables 1-3 below; only 3 are listed as examples due to space limitations): Table 1

[0054] Table 2

[0055] Table 3

[0056] (3) Screening and verification methods The model was validated in a full-scale digital twin model in the cloud using the "controlled variable method": candidate minor parameters were eliminated one by one, and the mean square error (MSE) of the model output (such as oil temperature and vibration amplitude) before and after elimination was calculated compared with the measured data. If the MSE change rate was less than 5%, it was determined to be a "minor parameter" and could be eliminated; otherwise, it was retained. Finally, 15 core parameters were determined, reducing the model calculation time by 60% and meeting the lightweight requirements of edge computing.

[0057] (ii) Parameter update logic after real-time calibration trigger The edge digital twin model needs to update its parameters without interrupting monitoring to avoid early warning delays caused by calibration. The specific process is as follows: (1) Calibration trigger conditions Real-time calibration is triggered when the edge terminal detects any of the following conditions: Physical residual (Residualphysics = |X) actual - X predicted_physics The mean of the sliding window (window size = 5 seconds) exceeds "1.2 times the dynamic threshold" for two consecutive windows; Sudden changes in key operating parameters: sudden increase / decrease in wind speed >5m / s, speed fluctuation >10% of rated speed, sudden change in torque >15% of rated torque (all based on 1-second sampling data).

[0058] (2) Four-step parameter update process (imperceptible effect) Upload of abnormal data slices: The edge terminal automatically captures high-resolution data (sampling frequency increased to 1kHz, including key operating parameters and sensor data) from 3 seconds before triggering to 2 seconds after triggering, and uploads it to the cloud via a 5G slicing network (priority: high). The data volume is controlled within 10MB (to avoid bandwidth occupation), and the upload time is <0.5 seconds.

[0059] Cloud-based rapid calibration calculation: The full digital twin model of the unit is loaded into the cloud, and the parameters are iteratively optimized using the "measured values ​​of the uploaded data" as the truth values ​​and the "Bayesian optimization algorithm". Optimization objective: Minimize the MSE between the model output value and the measured value; Iteration constraint: The parameter update range shall not exceed ±20% of the initial value (to avoid distortion of physical meaning, such as thermal resistance cannot suddenly change from 12W / (m·K) to 20W / (m·K)). Iteration efficiency: Set the maximum number of iterations to 20, the time taken for a single iteration is less than 1 second, and the total calibration time is less than 30 seconds (to meet real-time requirements).

[0060] Example: If the thermal resistance parameter is optimized from 12W / (m·K) to 12.8W / (m·K) after calibration, then the "update instruction" for this parameter (including parameter ID, new value, and effective timestamp) will be output.

[0061] Edge parameter reception and caching: After receiving the update command, the edge terminal caches the new parameters in "dual-zone memory" (zone A: current running parameters; zone B: parameters to be effective) and records the effective timestamp (accurate to milliseconds) without interrupting the current model inference (zone A continues to run).

[0062] No restart mechanism required for the changes to take effect. Utilizing the "model inference frame gap update" technology: During the inference frame gap between the ST-GNN and the digital twin model (approximately 100ms, with 10 inferences completed every 1 second), the new parameters of region B are automatically synchronized to region A. The synchronization time is <10ms, and a "parameter smooth transition" is adopted during the synchronization process (e.g., when the thermal resistance changes from 12 to 12.8 W / (m·K), a linear transition is performed over 5 frames to avoid abrupt changes in model output). After synchronization is completed, the terminal sends a "parameter update successful" message to the cloud. The entire process is uninterrupted by monitoring, and the warning delay is <100ms.

[0063] (III) Self-supervised optimization process for multi-factor edge weights The dynamic graph edge weights (w) of ST-GNN ij The formula (t) = α·f(·) + β·g(·) + γ·h(·) needs to be adaptively adjusted based on the actual wind field operation data to ensure the "accuracy of spatiotemporal correlation characterization". This process aims to "minimize the state prediction error of ST-GNN" and achieves automatic optimization of α, β, and γ, as follows: (1) Optimization cycle and data foundation Optimization cycle: once every 7 days (balancing real-time performance and data volume, 7 days can cover common wind farm operating condition changes, such as wind direction switching and unit load fluctuations). Data foundation: Extract normal operating data for the past 7 days (excluding abnormal / failure periods), construct "input-output" sample pairs (input: the status of the entire field in the past 60 minutes; output: the status of the entire field in the next 5 minutes), with a sample size of ≥1000 sets.

[0064] (2) Three-step self-supervised optimization method 1. Initial weight settings: Based on historical wind farm experience, the initial weights are set as follows: α=0.5 (primarily due to wake influence), β=0.3 (secondarily due to aging influence), and γ=0.2 (secondarily due to maintenance influence), and α+β+γ=1.

[0065] 2. Weighted sensitivity analysis: With β and γ fixed, α is adjusted in steps of 0.05 within the range of [0.3, 0.7], and ST-GNN models are trained respectively. The "Full-field state prediction MSE for the next 5 minutes" corresponding to each group of α is calculated. Similarly, with α and γ fixed, β is adjusted (range [0.1, 0.5]), and with α and β fixed, γ is adjusted (range [0.05, 0.3]). The "weight-MSE" curve is plotted, and the weight interval with the fastest decrease in MSE is determined (such as α∈[0.5, 0.6], β∈[0.25, 0.35], γ∈[0.15, 0.25]).

[0066] 3. Multivariate collaborative optimization: With the objective function of minimizing the mean squared error (MSE), and constraints of α∈[0.5,0.6], β∈[0.25,0.35], γ∈[0.15,0.25] and α+β+γ=1, a gradient descent method is used for collaborative optimization. Objective function:

[0067] These are the model's predicted values. (where N is the measured value and N is the sample size). Gradient calculation: Calculate the partial derivatives of Loss with respect to α, β, and γ, and iteratively update the weights along the negative gradient direction (learning rate = 0.01, number of iterations = 100). Optimization result verification: Substitute the optimized weights (e.g., α=0.6, β=0.25, γ=0.15) into the ST-GNN model and verify them with new samples (data from the next 24 hours). If the predicted MSE decreases by ≥15% compared to the initial weights, the weights are confirmed to be effective and sent to the edge terminal to update the dynamic graph calculation logic; if the MSE decreases by <10%, the optimization is iterated again.

[0068] (3) Emergency adjustment of weights under special working conditions When extreme operating conditions occur in the wind farm (such as a typhoon passing through: wind speed > 25 m / s, aging of units > 50%), "emergency optimization" is triggered: Extreme wind conditions: Temporarily increase the weight of α to 0.7 (wake effect is dominant), and reduce β and γ proportionally (β=0.2, γ=0.1). Centralized operating conditions for aging units: Temporarily increase the β weight to 0.4 (aging influence is dominant), and decrease α and γ (α=0.45, γ=0.15). Within 24 hours of the emergency optimization, the regular self-supervised optimization process will be automatically initiated to restore the optimal weight.

[0069] Step 2: Real-time monitoring and dynamic threshold calculation of edge-side "twin-data" dual-track system: Run the lightweight ST-GNN and digital twin models in parallel, output dual-track prediction values, calculate dual-track residuals, and calculate the dynamic RUL threshold based on the maintenance decision objective function (same as step four in the first stage).

[0070] In this embodiment, since the edge terminals are mostly industrial embedded devices (such as Intel Atom x7-E3950 processors and 4GB of memory), their computing power is limited. To ensure that the dual-track monitoring (ST-GNN + digital twin) can run in parallel without affecting the real-time performance of data acquisition, the following operations can be performed: (I) Model Lightweighting Technology Approach (Taking the Simplified Digital Twin Model of a Gearbox as an Example) By employing a three-dimensional approach of "parameter pruning + operator optimization + quantization compression," the full cloud model (memory usage > 1GB, CPU utilization > 80%) is simplified into an edge-usable model, as detailed in Table 4 below: Table 4

[0071] (ii) Real-time monitoring and dynamic scheduling of resource usage An edge terminal deploys a "resource monitoring daemon" (using less than 5% CPU and less than 50MB of memory) to monitor the resource consumption of the ST-GNN and digital twin model in real time, ensuring system stability through "dynamic frequency reduction + priority scheduling". (1) Resource monitoring indicators and thresholds Monitoring frequency: 1 second / time, collecting "ST-GNN memory usage, digital twin memory usage, total CPU utilization, and data acquisition thread CPU utilization"; Alarm thresholds: Total memory usage > 80% (4GB memory → > 3.2GB), total CPU usage > 70%, and CPU usage of data acquisition threads < 10% (to avoid acquisition delays).

[0072] (2) Dynamic scheduling strategy When the total CPU utilization is >70%: reduce the inference frequency of the digital twin model from 1 second / time to 2 seconds / time (the physical residual calculation still meets the anomaly detection requirements, with a delay of <2 seconds), and temporarily increase the quantization accuracy of the ST-GNN model from INT8 to INT16 (accuracy loss <3%, CPU utilization reduced by 15%). When total memory usage exceeds 80%: Releasing the historical computation cache of the digital twin model (retaining data from the most recent 10 minutes and deleting earlier caches) can reduce memory usage by 20% to 30%; When the CPU utilization of the data acquisition thread is less than 10%, the inference priority of the dual-track model is reduced (the priority of the data acquisition thread is set to "highest") to ensure that the data acquisition is uninterrupted at the 1-second level (thus, the core requirements of high-frequency monitoring of wind farms can be met).

[0073] (III) Resource usage verification results (based on mainstream edge terminals) Verified on an edge terminal with an Intel Atom x7-E3950 (4 cores 1.6GHz) + 4GB DDR4 processor, during parallel dual-track monitoring operation: ST-GNN model: 180MB memory usage, 25% CPU utilization; Digital twin model: Memory usage 190MB, CPU utilization 28%; Total resource consumption: Memory usage 370MB (<10% of total memory), CPU utilization 53% (<70% threshold); Data acquisition thread: CPU utilization 15%, acquisition latency <100ms, fully meeting the real-time monitoring requirements of wind farms.

[0074] Step 3: Health assessment and RUL prediction based on dual-track residuals (edge ​​execution): Lightweight RUL prediction model (small LSTM) inputs dual-track residuals, output: Health Index (0-100 points); Estimated RUL (hours / day) after dynamic threshold correction; Potential failure mode probability (e.g., "gearbox: 75%").

[0075] In this embodiment, since the accuracy of RUL prediction directly affects operation and maintenance decisions, and the "data residual (empirical bias)" and "physical residual (mechanistic bias)" may deviate due to fluctuations in operating conditions (such as gusts of wind), the following error correction mechanism is required to ensure that the RUL prediction error is <8%: (a) Criteria for Determining Dual-Track Residual Deviation First, we define "residual deviation" to quantify the inconsistency between the two, as follows: (1) Calculation of residual deviation For each monitoring time t, three core measuring points are selected: gearbox oil temperature, main bearing vibration Z-axis, and generator speed (covering thermodynamic, mechanical, and electrical dimensions), and the following calculations are performed: Single measuring point deviation: (i=1,2,3 represent 3 measuring points respectively); Global Deviation: (Take the average of the deviations at 3 measuring points to avoid the influence of occasional fluctuations at a single measuring point on the judgment).

[0076] (2) Deviation threshold setting Based on historical normal operating data of the wind farm (1000+ moments), statistics were compiled under normal operating conditions. The 95th percentile is 8%, therefore we set: Deviation warning threshold: (The reasons for the deviation need to be noted); Adjust the trigger threshold: (Kalman filter correction needs to be enabled).

[0077] (ii) Kalman filter error correction mechanism (for RUL prediction values) when When significant inconsistencies are found between the two track residuals (possibly due to data noise or sudden changes in operating conditions), a Kalman filter is used to fuse the two track residual information and correct the RUL prediction value. The specific process is as follows: (1) Design of state equations and observation equations Using the "RUL predicted value" as the state variable and the "fusion value of the dual-track residuals" as the observation value, a discrete Kalman filter model is constructed: Equations of state:

[0078] In the formula: Let be the corrected RUL prediction value at time k (in hours); A is the state transition matrix (set to 1.0, assuming RUL decays linearly with time and has no abrupt changes). The state noise is N(0,Q), where Q is the state noise covariance, and the prediction error based on historical RUL is set to 0.01.

[0079] Observation equation:

[0080] In the formula: The "dual-track residual fusion observation" at time k (calculated as follows: Physical residuals have higher weights because the mechanism is more stable); H is the observation matrix (set to 1.0, where observations are directly related to state variables). The observation noise (following N(0,R), where R is the observation noise covariance, is set as D based on the dual-track residual bias). total (t)×0.01, the larger the deviation, the larger R).

[0081] (2) Five-step correction process for Kalman filtering 1. State prediction: ( The corrected RUL value at time k-1 (RUL value predicted at time k). 2. Error covariance prediction: Let k be the state error covariance at time k-1. (where k is the prediction error covariance) 3. Kalman gain calculation: ( (The weights of the equilibrium state prediction and the observations are used for gain). 4. Status Update: (By fusing observations to correct RUL predictions, we obtain) ); 5. Error covariance update: (I is the identity matrix, and the updated error covariance is used for calculation in the next time step).

[0082] (3) Verification of the correction effect Take the case of "gearbox main bearing wear" as an example: Before correction: Data residual prediction RUL = 80 hours, physical residual prediction RUL = 100 hours, D total =20%, RUL prediction error =15%; Corrected: Obtained through Kalman filtering The actual RUL is 90 hours, and the error is 2.2% < 8%, which meets the accuracy requirements.

[0083] (III) Tracing the causes of deviations and optimizing the closed loop After correction, the cause of the deviation needs to be traced to prevent recurrence: If the deviation is caused by "data noise" (such as occasional sensor jumps): the edge terminal automatically marks the data at that moment as "noise" and removes it from the RUL prediction sample library; If the deviation is caused by "sudden changes in operating conditions" (such as a sudden increase in speed due to gusts of wind): update the self-supervised training samples of ST-GNN in the cloud, add data on sudden changes in operating conditions, and improve the model's adaptability to sudden operating conditions. If the deviation is caused by "inaccurate digital twin parameters": trigger the first step of the "real-time calibration process" to re-optimize the digital twin parameters.

[0084] Step 4: Cloud-based multimodal fusion deep diagnostics: Upload multimodal data packets (sensor data + model status + operation and maintenance text). The Transformer model achieves deep fusion diagnosis of "data-text-physical model" through cross-modal attention layer and knowledge graph embedding sub-network, and outputs fault root causes and maintenance priority queues (same as steps six and seven of the second stage of Example 1).

[0085] In this embodiment, since multimodal fusion diagnosis needs to address the issues of "lagging knowledge iteration" and "difficulty in understanding by maintenance personnel," the following steps can also be taken to solve these problems: (I) Dynamic update mechanism of knowledge graph Wind power fault modes may be affected by aging of random groups and adjustments to operation and maintenance strategies (such as "pressure anomalies caused by the deterioration of new hydraulic oil"). Therefore, a dual-driven knowledge graph update process, combining automatic and manual methods, needs to be established, as detailed below: (1) Basic structure of knowledge graph Entity layer: contains 4 types of entities: “Components (gearbox / main bearing)”, “Sensors (oil temperature / vibration)”, “Fault types (insufficient lubrication / bearing wear)”, and “Maintenance operations (replacing hydraulic oil / cleaning filter)”, totaling 120+ entities; Relationship layer: Includes 3 types of relationships, totaling 200+ relationships: "Fault-caused (e.g., 'hydraulic oil deterioration' → 'abnormal hydraulic station pressure')", "Sensor monitoring (e.g., 'vibration Z-axis' → 'main bearing')" and "Operation and maintenance solutions (e.g., 'replacing hydraulic oil' → 'hydraulic oil deterioration')". Storage method: Neo4j graph database is used, which supports fast addition, deletion and query of entities / relationships (query response time <100ms).

[0086] (2) Update trigger conditions A knowledge graph update is triggered when any of the following conditions are met: 1. New fault case verification passed: The operation and maintenance personnel reported that "the root cause in the diagnostic report is consistent with the actual fault", and the fault mode has no corresponding relationship in the knowledge graph (such as "hydraulic oil deterioration → abnormal hydraulic station pressure" does not exist). 2. Adding new expert knowledge: Wind farm operation and maintenance experts can manually enter new relationships (such as "filter blockage → insufficient lubricating oil flow") through the "knowledge graph management interface"; 3. Model diagnostic error exceeds the standard: When the diagnostic accuracy of the Transformer model for a certain type of fault is less than 85% for 3 consecutive times, and the analysis finds that it is caused by "the knowledge graph lacking corresponding fault propagation relationships" (such as not associating "gearbox abnormal noise → excessive bearing clearance").

[0087] (3) Automatic update of the three-step process (taking "hydraulic oil deterioration → abnormal hydraulic station pressure" as an example) 1. Relationship extraction and verification: Extract the "fault cause entity (hydraulic oil deterioration)", "fault result entity (abnormal hydraulic station pressure)" and "relationship type (fault-induced)" from the operation and maintenance feedback data; The cloud automatically retrieves multimodal data (hydraulic oil test report + pressure sensor time series data + maintenance work order) for the fault case, verifies that "the hydraulic oil deterioration occurred earlier than the pressure anomaly time (time difference > 24 hours)" and "the pressure returned to normal after the hydraulic oil was replaced", confirming the validity of the relationship.

[0088] 1. Knowledge graph input: Call Neo4j's Cypher statement to add a new relation: MATCH(a: fault type {name:"hydraulic oil deterioration"}), (b: fault type {name:"hydraulic station pressure abnormal"}) CREATE(a)-[r: fault caused {confidence:0.95,source:"Operation and Maintenance Case 20240510"}]->(b) (confidence is the relation confidence level, set based on the case verification results); Synchronously update "Entity Attributes": Add "Feature Description" attribute (such as "Hydraulic Oil Deterioration") to the "Hydraulic Oil Viscosity <150cSt@40℃"), and add "Sensor Association" attribute (such as "Hydraulic Station Pressure Anomaly") to the "Hydraulic Station Pressure Anomaly" entity.

[0089] 2. Synchronous optimization of knowledge constraints in the Transformer model: Extract "entity embedding vectors" of newly added relationships from the knowledge graph (generated using the TransE model, such as "hydraulic oil deterioration" → [0.3, 0.2, -0.4], "hydraulic station pressure anomaly" → [0.5, 0.1, -0.3]). Update the knowledge constraint terms of Transformer: Add "related entity embedding similarity penalty" to the self-attention weight calculation. If the attention weight points to "entity pairs without knowledge graph relationship" (e.g., "abnormal wind speed → abnormal hydraulic station pressure"), then increase the penalty term based on embedding similarity (similarity < 0.2) and decrease the weight; if it points to "entity pairs with knowledge graph relationship" (e.g., "hydraulic oil deterioration → abnormal hydraulic station pressure", similarity > 0.8), then decrease the penalty term and increase the weight. Model fine-tuning: The Transformer model is fine-tuned using fault case data (10+ groups) corresponding to the new relationship (10 iterations, learning rate = 1e-5) to ensure that the model adapts to the new relationship and improves the diagnostic accuracy to over 90%.

[0090] (4) Post-update verification and rollback Verification: Within 24 hours after the update, monitor the diagnostic accuracy of the Transformer model for this type of fault. If it is ≥90%, the update is confirmed to be effective. Rollback: If the accuracy is less than 85%, the newly added relationship will be automatically deleted, the model parameters will be restored, and manual review will be triggered (to check whether the relationship extraction was incorrect).

[0091] (ii) Visualization of cross-modal attention weights Because the diagnostic report needs to transform the "abstract attention weights" into "intuitive heatmaps" to help operations and maintenance personnel quickly understand "which modalities / sensors / time points play a key role in fault diagnosis," as detailed below: (1) Heat map dimension design (multi-dimensional coverage) A three-dimensional heatmap of "modality-sensor-time" is designed for three modalities: "sensor data, operation and maintenance text, and digital twin status". Specific dimensions include: Modal dimensions: 3 modalities (sensor data → S, maintenance text → T, digital twin status → P); Sensor dimensions: 10+ core sensors (such as gearbox oil temperature S1, main bearing vibration Z-axis S2, hydraulic station pressure S3...). Time dimension: Abnormal period (5 minutes before the abnormality is triggered → T-5 to 5 minutes after the abnormality is triggered → T+5, divided into 10 time windows of 1 minute each).

[0092] (2) Attention weight calculation and mapping 1. Weight Extraction: Extract the "cross-modal attention weight matrix" from the self-attention layer of the Transformer model. The dimension is (number of modalities × number of sensors × number of time windows), such as W[S1,T-3]=0.85 (meaning that the attention weight of "gearbox oil temperature sensor at time T-3" is 0.85). 2. Weight Normalization: Map the weights to the [0,1] interval, using the following formula: (W) min For minimum weight, W max (for maximum weight) 3. Color Mapping: Employs a red-yellow-green gradient color scheme, with higher weights resulting in darker colors. Weight ≥ 0.8 → Dark red (key association, contributing the most to the diagnosis); 0.5 ≤ weight < 0.8 → Yellow (Important correlation); Weight < 0.5 → Green (weak association, small contribution to diagnosis).

[0093] (3) Heat map presentation format and interactive functions 1. Static thermal map (embedded in diagnostic report): The data is presented in a table format, with rows labeled "sensor + time window" and columns labeled "modal type". Cells are filled with the corresponding color and labeled with the normalized weight values. Example snippets are shown in Table 5 below: Table 5

[0094] 2. Interactive heatmap (displayed on the web): like Figure 2 and Figure 3 As shown, it supports "hovering the mouse to view details": for example, hovering over the dark red cell displays "The attention weight of the gearbox oil temperature sensor at time T-3 (XX:XX) is 0.85. The measured oil temperature at that time is 45℃, the model prediction is 38℃, and the residual is 7℃. It is a key correlation point for the root cause of the fault." Figure 3 For simplicity, we will only display "Weight: 0.98; Time 14:32 (abnormal); Sensor: Hydraulic station pressure"; Supports "Timeline Zoom": You can switch from a "1-minute window" to a "10-second window" to view more granular weight changes; Supports "Modal Filtering": You can select the "Sensor Data" modality separately to view the weight distribution of each sensor in that modality.

[0095] It also supports diagnostic summaries, core conclusions, intelligent maintenance suggestions, and Figure 4 The associated data snapshots allow maintenance personnel to quickly understand the data.

[0096] (4) The value of visualization to operation and maintenance Lowering the barrier to understanding: Maintenance personnel do not need to understand the principles of deep learning models; they can quickly locate "key sensors (such as S1 oil temperature), key time (T-3), and key modes (sensor data)" simply by using colors. Auxiliary maintenance verification: If the heat map shows that "the hydraulic station pressure sensor (S3) has a high weight at time T-1", maintenance personnel can prioritize checking the sensor and the corresponding hydraulic system, improving maintenance efficiency by 30%; Fault review basis: Regularly summarize heat map data, analyze "key correlation patterns of similar faults" (such as all hydraulic faults showing high weight of "hydraulic station pressure sensor + T-1 time"), and optimize operation and maintenance plans.

[0097] Example 4 Based on the same concept, this invention provides a deep learning-based wind farm turbine condition monitoring and fault early warning system, comprising: An edge computing terminal, deployed on a wind turbine, is configured to perform the edge-end real-time monitoring steps in Example 1; The cloud platform communicates with the edge computing terminal and is configured to perform the cloud-based deep diagnostic steps in Example 1.

[0098] Specifically, this system includes an edge layer and a cloud layer that collaborate with the cloud. The edge layer is deployed on each wind turbine, and the cloud layer is deployed on a cloud server. The edge layer and the cloud layer interact with each other through a communication network. The edge layer includes a high-frequency data acquisition module, a lightweight spatiotemporal graph neural network (ST-GNN) inference module, a lightweight digital twin model module, a dual-track residual calculation module, and a dynamic remaining useful life (RUL) early warning threshold calculation module. The high-frequency data acquisition module is used to collect high-frequency operating data of wind turbine units. The lightweight ST-GNN inference module is used to output data-driven theoretical normal prediction values ​​based on the collected operating data and adjacent unit status data. The lightweight digital twin model module is used to output physical-driven theoretical normal prediction values ​​based on real-time operating parameters. The dual-track residual calculation module is used to calculate the first residual between the data-driven theoretical normal prediction value and the measured value, and the second residual between the physical-driven theoretical normal prediction value and the measured value. The dynamic RUL early warning threshold calculation module is used to construct a maintenance decision objective function based on power generation loss, maintenance cost, and wind farm predicted power generation. The RUL corresponding to the minimum value of the objective function is taken as the early warning threshold. When the fusion value of the first residual and the second residual exceeds the early warning threshold for a continuously preset time, it is judged as an abnormal event and reported to the cloud layer. The cloud layer includes a dynamic topology graph construction module for wind farms, a global model training module, a Transformer fault diagnosis module integrating knowledge graphs, and an operation and maintenance priority ranking module. The dynamic topology graph construction module abstracts the wind farm into a dynamic graph, where nodes represent wind turbines, and node features are turbine operating data. The weights of edges between turbines are calculated based on a weighted average of wake influence factors, turbine aging factors, and maintenance-related factors, with the total weighting coefficients set to 1 and dynamically adjusted according to wind farm operating conditions. The global model training module trains a global ST-GNN state prediction model, a global digital twin full-scale model, and a Transformer fault diagnosis module integrating knowledge graphs based on historical normal operating data of the wind farm. The system employs a fault diagnosis model and distributes the trained lightweight model to the edge layer. A Transformer fault diagnosis module, incorporating a knowledge graph, receives multimodal data of abnormal events reported from the edge layer. This multimodal data includes high-resolution operational data during abnormal periods, digital twin status data, and maintenance text data. It mines the correlation weights between multimodal data through a cross-modal attention layer and constrains these correlation weights using the fault propagation patterns of the wind power fault knowledge graph. The module outputs the probability of the fault category, the contribution of key sensors, and the critical time point of the fault. A maintenance priority ranking module generates a maintenance priority queue and optimal maintenance window for wind farm units based on the maintenance decision objective function and the severity of the fault.

[0099] In this embodiment, the edge layer is deployed in the control cabinet or nacelle at the bottom of the wind turbine tower. The high-frequency data acquisition module is connected to the main control PLC of the unit via industrial Ethernet. The high-frequency operating data acquired includes vibration data, temperature data, pressure data and electrical quantity data, with an acquisition frequency of 1 to 1000 times per second.

[0100] In this embodiment, the lightweight ST-GNN inference module consists of a graph attention network (GAT) submodule and a gated recurrent unit (GRU) submodule. The GAT submodule uses the LeakyReLU activation function with a negative slope of 0.1-0.3 and a dropout probability of 0.2-0.4. The input features of the GRU submodule are normalized to the [0,1] interval by Min-Max. The lightweight ST-GNN inference module is quantized and pruned by a quantization tool with a quantization accuracy of INT8 and a pruning sparsity of 20%-40%.

[0101] In this embodiment, the lightweight digital twin model module retains the core physical equations of key components of the wind turbine, with 12-18 core physical parameters. These core physical parameters include at least 12 of the following: friction coefficient, equivalent thermal resistance, power loss coefficient, main bearing stiffness coefficient, gear transmission efficiency, bearing contact fatigue coefficient, lubricating oil viscosity, and hydraulic station oil supply pressure coefficient.

[0102] In this embodiment, the dynamic topology graph construction module adjusts the weight of the associated edges between units once per minute; when the wind speed in the wind farm is greater than 12 m / s, the weighting coefficient of the wake influence factor is increased to 0.5-0.7; when the proportion of units in the wind farm that have been in operation for more than 5 years is greater than 30%, the weighting coefficient of the unit aging factor is increased to 0.3-0.5.

[0103] In this embodiment, the wind power fault knowledge graph in the Transformer fault diagnosis module, which integrates knowledge graphs, includes four types of entities: “component-sensor-fault type-operation and maintenance” and three types of relationships: “fault cause-sensor monitoring-operation and maintenance solution”. The knowledge graph is stored in a graph database, and the query response time is no more than 100ms.

[0104] In this embodiment, the dual-track residual calculation module calculates the fusion value of the first residual and the second residual by weighted Euclidean distance. The weight of the first residual is 0.3-0.5, and the weight of the second residual is 0.5-0.7. The preset duration is 2-5 seconds. When the fusion value exceeds the warning threshold for a continuous preset duration, an abnormal event is reported.

[0105] In this embodiment, a digital twin real-time calibration module is also included. When the duration of the second residual continuously preset sliding window exceeds 1.2 times the warning threshold, or when the key operating parameters of the wind field change abruptly—including a sudden increase or decrease in wind speed exceeding 5 m / s, a fluctuation in rotational speed exceeding 10% of the rated rotational speed, and a sudden change in torque exceeding 15% of the rated torque—the digital twin real-time calibration module triggers calibration: the edge layer uploads high-resolution data of the abrupt change period to the cloud layer, the cloud layer iteratively optimizes the core parameters of the digital twin through a Bayesian optimization algorithm, with no more than 20 iterations and a total calibration time of no more than 30 seconds, and then sends the optimized parameters down to the edge layer, the edge layer smoothly updates the parameters through dual-zone memory caching and between model inference frames, with an update time of no more than 10 ms.

[0106] In this embodiment, a dual-track residual deviation correction module is also included. The deviation correction module determines the consistency of the dual-track residuals by calculating the global deviation degree. The global deviation degree is the average deviation degree of the single measuring point of the core measuring point of the wind turbine unit—the core measuring points include the gearbox oil temperature measuring point, the main bearing vibration Z-axis measuring point, and the generator speed measuring point. When the global deviation degree exceeds 10%, the Kalman filter algorithm is used to fuse the dual-track residual information to correct the RUL prediction value. The state variable of the Kalman filter is the RUL prediction value, and the observation value is the weighted fusion value of the dual-track residuals. The weight of the second residual in the observation value is 0.5-0.7, and the weight of the first residual is 0.3-0.5.

[0107] In this embodiment, a knowledge graph dynamic update module is also included. When a new fault case is verified, the operation and maintenance expert adds knowledge, or the Transformer fault diagnosis module that integrates the knowledge graph has a diagnostic accuracy rate of less than 85% for a certain type of fault for three consecutive times, the update module extracts the effective fault association relationship and stores it in the database, updates the knowledge constraint items of the Transformer fault diagnosis module synchronously, and then uses the newly added fault case data to fine-tune the model so that the diagnostic accuracy rate of the fine-tuned model for this type of fault is not less than 90%.

[0108] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0109] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0110] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 1 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0111] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for wind farm turbine condition monitoring and fault early warning based on deep learning, characterized in that, Includes the following steps: Real-time monitoring steps at the edge: S1: Obtain the real-time operating data of the wind turbine at the edge computing terminal deployed on the wind turbine; S2: Based on the edge computing terminal, a data-driven spatiotemporal prediction model and a physical mechanism-based digital twin model are run in parallel to obtain the data prediction state and the physical prediction state, respectively. S3: Calculate the first residual between the real-time running data and the data prediction state, and the second residual between the real-time running data and the physical prediction state, and combine the first residual and the second residual into a dual-track residual; S4: Compare the dual-track residual with the dynamic early warning threshold to determine abnormal events; Cloud-based in-depth diagnostic steps: S5: When the abnormal event is detected, the edge computing terminal uploads multimodal abnormal data containing at least two of the sensor data, model status, and operation and maintenance text to the cloud platform; S6: On the cloud platform, a diagnostic model that integrates a fault knowledge graph in the wind power field is used to analyze the multimodal abnormal data in order to locate the root cause of the fault. S7: Generate a fault diagnosis report based on the root cause of the fault.

2. The method for wind farm unit condition monitoring and fault early warning based on deep learning as described in claim 1, characterized in that, The data-driven spatiotemporal prediction model is a spatiotemporal graph neural network model, in which the wind farm is constructed as a dynamic topology graph, and the weights of the connection edges between the units in the dynamic topology graph are dynamically adjusted according to at least one of the wake effect, the degree of unit aging, and the maintenance history.

3. The method for wind farm unit condition monitoring and fault early warning based on deep learning as described in claim 1, characterized in that, The dynamic early warning threshold is calculated based on a preset maintenance decision objective function, which comprehensively considers the estimated power generation loss and maintenance costs.

4. The method for wind farm unit condition monitoring and fault early warning based on deep learning as described in claim 1, characterized in that, The diagnostic model is an improved Transformer model, which includes a cross-modal attention layer for calculating the correlation weights between different modalities in the multimodal anomaly data.

5. The method for wind farm unit condition monitoring and fault early warning based on deep learning as described in claim 1, characterized in that, It also includes a real-time calibration step for the digital twin model: When the deviation between the physical prediction state and the real-time running data exceeds a preset calibration threshold, the edge computing terminal is triggered to upload high-resolution data to the cloud platform. The cloud platform optimizes the key parameters of the digital twin model and sends the optimized parameters to the edge computing terminal for updating.

6. The method for wind farm unit condition monitoring and fault early warning based on deep learning as described in claim 1, characterized in that, It also includes a closed-loop optimization step: feeding back the actual repair results based on the fault diagnosis report as new cases to the cloud platform to iteratively update the diagnostic model and the knowledge graph.

7. A deep learning-based wind farm turbine condition monitoring and fault early warning system, characterized in that, include: An edge computing terminal, deployed on a wind turbine, is configured to perform the edge-end real-time monitoring steps as described in claim 1; A cloud platform, communicating with the edge computing terminal, is configured to perform the cloud-based deep diagnostic steps as described in claim 1.

8. The system according to claim 7, characterized in that, The data-driven spatiotemporal prediction model is a spatiotemporal graph neural network model, in which the wind farm is constructed as a dynamic topology graph, and the weights of the connection edges between the units in the dynamic topology graph are dynamically adjusted according to at least one of the wake effect, the degree of unit aging, and the maintenance history.

9. The system according to claim 7, characterized in that, The system is also configured to calculate the dynamic early warning threshold based on a preset maintenance decision objective function, which comprehensively considers the estimated power generation loss and maintenance costs.

10. The system according to claim 7, characterized in that, The cloud platform and the edge computing terminal are also configured to perform a real-time calibration step for the digital twin model, which includes: When the deviation between the physical prediction state and the real-time running data exceeds a preset calibration threshold, the edge computing terminal is triggered to upload high-resolution data to the cloud platform. The cloud platform optimizes the key parameters of the digital twin model and sends the optimized parameters to the edge computing terminal for updating.

Images