Multi-source data fusion wind turbine anomaly detection method and development platform

Abstract

This application discloses a method and development platform for anomaly detection in wind turbine generators based on multi-source data fusion. The algorithm includes: collecting multi-source data related to wind turbine generators; preprocessing the collected data to generate a multi-source dataset; using a hybrid algorithm model to detect and analyze the multi-source dataset and output anomaly detection results; specifically, extracting features from the input multi-source dataset and incorporating them into a global feature pool, employing a shared feature encoder, and jointly using a multi-task branch decoder to complete the detection and output of anomaly detection results for each of the multi-task branches; generating early warning prompts for anomaly types and targeted governance suggestions based on the anomaly detection results and a historical case library; acquiring multi-source data during the governance period, setting quantitative evaluation indicators, obtaining evaluation results of the targeted governance suggestions and providing feedback to dynamically update the parameters of the hybrid algorithm model. This application enables effective fusion of multi-source data from wind turbine generators, accurate location of anomaly causes, and completion of closed-loop management intelligent applications.

Description

Technical Field

[0001] This application relates to the field of data processing and equipment monitoring technology, specifically to a method and development platform for detecting anomalies in wind turbine generators using multi-source data fusion. Background Technology

[0002] As the core equipment of wind power generation, the operating status and power generation performance of wind turbines directly affect the economic benefits and energy supply stability of the entire wind farm. The power generation performance of wind turbines is affected by various factors. Increased operating time leads to a gradual decline in power generation performance. Simultaneously, changes in wind resources, aging controllers and software, and incorrect adjustments to control parameters can all adversely affect the power generation performance of wind turbines. Therefore, accurate detection, analysis, and evaluation of the power generation performance of wind turbines are of significant practical importance.

[0003] In traditional wind turbine operation and maintenance, conventional methods are used to detect and evaluate the power generation performance of wind turbines. For anomaly detection, traditional methods often rely on a single data source, such as analyzing only SCADA data. In handling multi-source data, there is a lack of a unified data fusion mechanism, making it difficult to efficiently process heterogeneous and inconsistent data formats, such as SCADA data, laser ranging data, and event logs. Traditional wind turbine detection platforms typically only have anomaly alarm functions, focusing primarily on detecting abnormal situations and issuing alerts. Furthermore, when inspecting wind turbines of different models and operating years, they rarely fully consider the performance differences between them, instead employing more general detection methods.

[0004] This reveals significant shortcomings in traditional operation and maintenance methods. Traditional wind turbine anomaly detection relies on a single data source, resulting in insufficient data dimensions and low accuracy in anomaly identification. It also fails to cover various anomaly scenarios, such as anemometer malfunctions and yaw system failures. Existing detection methods lack a unified data fusion mechanism; multi-source data exhibits heterogeneous formats and inconsistent standards, leading to low data preprocessing efficiency and impacting the real-time performance of anomaly detection. Traditional platforms only provide anomaly alarm functions, lacking a closed-loop management capability encompassing "detection-analysis-diagnosis-remediation," thus failing to provide accurate anomaly root cause localization and targeted remediation suggestions for operation and maintenance. Furthermore, cross-model compatibility is poor; performance differences between different models and years of operation are not fully considered, resulting in insufficient generalization ability of the detection model and difficulty in adapting to the diverse needs of complex wind farms. Summary of the Invention

[0005] To achieve effective fusion of multi-source data from wind turbines, accurate location of anomaly causes, intelligent application of closed-loop management, and improved cross-model adaptability, this application provides a wind turbine anomaly detection method and development platform based on multi-source data fusion.

[0006] Firstly, this application provides a method for detecting anomalies in wind turbine generators based on multi-source data fusion, including: Collect SCADA data, laser ranging data, and event log data from wind turbine generators; The collected data is cleaned, aligned, and standardized to form a multi-source dataset in a unified format. A hybrid algorithm model is used to detect and analyze multi-source datasets, outputting anomaly detection results including anomaly type and location information. The hybrid algorithm model extracts features from the input multi-source dataset and incorporates them into a global feature pool. A shared feature encoder is used to extract global universal feature vectors, and a multi-task branch decoder is used to complete the anomaly detection results for each task branch. Cross-validation rules are used to verify the anomaly detection results of each branch. Furthermore, the decoder for each task branch adapts the detection algorithm according to a pre-defined hybrid strategy for the corresponding anomaly type. The training process of the hybrid algorithm model includes: selecting the main wind turbine model, acquiring historical operating data of the main turbine model under normal and abnormal conditions in different scenarios, and training the shared feature encoder; designing a joint loss function, establishing dedicated training data for different wind turbine models under different scenario conditions, and jointly fine-tuning the multi-task branch decoder. Based on anomaly detection results and historical case databases, combined with fault tree analysis and root cause analysis methods, early warning prompts and targeted governance suggestions are generated for anomaly types. During the governance period, SCADA data, laser ranging data and event log data of wind turbines are obtained, quantitative evaluation indicators are set, evaluation results of targeted governance suggestions are obtained and fed back, so as to dynamically update the parameters of the hybrid algorithm model.

[0007] By adopting the above scheme, standardized integration of multi-source heterogeneous data is achieved, providing a high-quality data foundation for anomaly detection; a hybrid model is designed to collaboratively cover anomaly detection across all scenarios, enabling precise location of anomaly causes; targeted governance suggestions are generated by combining fault tree analysis and root cause analysis methods, data during the governance period is acquired and evaluated to update model parameters; an algorithm system is designed for basic algorithm selection, hybrid strategy implementation, cross-scenario migration, and multi-detection collaboration, realizing full automation of the anomaly detection, governance scheme generation, and effect evaluation process, completing intelligent applications of closed-loop management and improving cross-model adaptability.

[0008] Preferably, the hybrid algorithm model includes the following steps: extracting features from the input multi-source dataset and incorporating them into a global feature pool: A three-level classification feature extraction strategy is set up, including: environmental investigation features, core performance features, and auxiliary features; For each level of classification features, common shared original features are pre-extracted, and then derivative features are extracted for the exclusive derivative features constructed for each anomaly type. The exclusive derivative features constructed for each anomaly type include: a first derivative feature characterizing the association between the corresponding detection target and environmental features of the anomaly type, a second derivative feature characterizing the correlation between core performance and the corresponding anomaly degree of the anomaly type, and a third derivative feature characterizing the association between auxiliary features and core performance. The auxiliary features include pitch features and large component temperature features. Features with variance less than the variance threshold, importance less than the importance threshold, and correlation with core performance less than the correlation threshold are removed. The extracted general shared original features and various anomaly detection-specific derived features are included in the feature pool and divided into three-level classification sub-feature pools. Each of the three-level classification sub-feature pools contains the corresponding general shared original features and various anomaly detection-specific derived features, and is labeled with normalization markers, scene adaptation markers, environmental interference amount, core performance deviation amount, and auxiliary correlation value.

[0009] By adopting the above scheme, a three-level classification feature extraction strategy can be set up to systematically classify features of multi-source datasets, which facilitates subsequent processing. For each level of classification features, general shared original features and specific derived features are extracted to comprehensively and specifically mine feature information. Furthermore, by removing features that do not meet the standards of variance, importance, and relevance, useless or low-value features are removed to improve feature quality.

[0010] Preferably, the step of using a shared feature encoder to extract the global universal feature vector includes: Informer, CNN, and TabNet were selected as the running models for the shared feature encoder. The steps for extracting the fully generalized feature vector using Informer, CNN, and TabNet include: The Informer model is used to process the temporal features in multi-source datasets and extract a general temporal feature vector, including predicted future temporal feature values. A CNN model is used to process static or structured features in multi-source datasets, and spatial general feature vectors at different scales are extracted by setting different convolutional layers. The TabNet model is used to concatenate the extracted temporal and spatial general features, and the feature selection mask of TabNet is enabled. Pre-weighted features are applied to different scenarios and detection tasks. The temporal and spatial general features are then fused through an attention mechanism to achieve adaptive feature selection and fusion, and the final general feature vector is output.

[0011] By adopting the above scheme and setting up multiple models, different types of features in multi-source datasets are processed in a targeted manner, which can better extract temporal and spatial general features, realize adaptive feature selection and fusion, and output the final general feature vector. This provides a more comprehensive and accurate feature foundation for subsequent anomaly detection, thereby improving the accuracy and reliability of anomaly detection.

[0012] Preferably, the joint multi-task branch decoder completes the anomaly detection results for each of the multi-task branches by outputting the following: For the task branch decoder of abnormal power generation performance, the CNN algorithm and the XGBoost algorithm are adapted. The global general feature vector output by the shared encoder, the general shared original features extracted from the global feature pool, and the power generation performance abnormality detection-specific derived features are concatenated and input into the CNN algorithm model to obtain the preliminary estimate. Then, the preliminary estimate is added and input into the XGBoost algorithm model. The XGBoost algorithm is used to perform a multi-classification task and output the various abnormal classes and their confidence scores. For the blade aerodynamic imbalance anomaly type task branch decoder, the random forest and gradient boosting tree algorithms are adapted. The global general feature vector output by the shared encoder, the general shared original features extracted from the global feature pool, and the blade aerodynamic imbalance anomaly type-specific derived features are concatenated and input into the random forest algorithm model to obtain the preliminary estimate; then the preliminary estimate is input into the gradient boosting tree algorithm model to output the post-estimate; and finally, the weighted fusion is used to obtain the final anomaly detection result. For the yaw deviation anomaly type task branch decoder, the random forest algorithm model and the LSTM algorithm model are adapted. The global general feature vector output by the shared encoder and the static feature tensor in the yaw deviation anomaly type-specific derived features extracted from the global feature pool are concatenated and input into the random forest algorithm model to output a fixed initial deviation value. The global general feature vector output by the shared encoder and the time-series feature tensor in the general shared original features extracted from the global feature pool and the yaw deviation anomaly type-specific derived features are concatenated and input into the LSTM algorithm model to output a dynamic deviation distribution. Based on the output fixed initial deviation value and dynamic deviation distribution, fusion verification is performed, and the verified deviation value is output.

[0013] By adopting the above scheme, different algorithm combinations are adapted to the task branch decoder for different anomaly types, so as to adapt to the detection and classification of different anomaly types, and output various anomaly detection results more accurately, thereby improving the accuracy and reliability of anomaly detection.

[0014] Preferably, the verification of anomaly detection results for each branch output based on cross-validation rules includes: An anomaly correlation graph is constructed based on historical anomaly detection results of different anomaly types to obtain the logical basis for cross-validation. Define the core dimensions of cross-validation, including result consistency verification, feature complementarity verification, and logical constraint verification; Based on the obtained cross-validation logic foundation and core dimensions of cross-validation, core cross-validation rules are formulated. Each rule includes validation conditions, judgment logic, quantitative score, and conclusion. The core cross-validation rules include: causal forward validation, reverse validation, and feature complementarity validation between power generation performance pitch anomaly and blade aerodynamic imbalance; causal forward validation, reverse validation, and feature complementarity validation between power generation performance yaw anomaly and yaw wind deviation; conflict-free validation and trend consistency validation between blade aerodynamic imbalance and yaw wind deviation; and global logic constraint validation. The system matches the anomaly detection results output by each branch with the established core cross-validation rules. Cross-validation is then performed between the corresponding branches based on the matched core cross-validation rules. The total cross-validation score between each branch is calculated and compared with the preset cross-validation score threshold to determine whether the cross-validation passes. If the cross-validation fails, a detection review prompt is generated.

[0015] By adopting the above scheme, an anomaly correlation graph is constructed using historical anomaly detection results. Cross-validation rules are formulated in conjunction with core dimensions, and the anomaly detection results output by each branch are cross-validated. This can improve the accuracy and reliability of anomaly detection results and reduce the false positive rate.

[0016] Preferably, the generation of targeted remediation suggestions for different anomaly types based on anomaly detection results and a historical case library, combined with fault tree analysis and root cause analysis methods, includes: Using the anomaly detection results output from each branch as input, the system obtains the top, intermediate, and bottom events through fault tree analysis; then, it determines the root cause of the anomaly through root cause analysis; it matches the historical case library and generates targeted governance suggestions based on the wind turbine model, wind turbine anomaly type, and wind turbine scenario type, with different levels of anomaly, scenario-specific adaptation, and step-by-step governance.

[0017] By adopting the above approach, fault tree analysis and root cause analysis methods are used to accurately identify the root causes of anomalies. Based on the wind turbine model, anomaly type, and location scenario, targeted management suggestions are generated that are adapted to different scenarios, differentiate anomaly levels, and have step-by-step management features. This provides precise and effective solutions for managing wind turbines with different anomaly types, improving the efficiency and effectiveness of anomaly management.

[0018] Preferably, the steps of acquiring SCADA data, laser ranging data, and event log data of wind turbines during the governance period, setting quantitative evaluation indicators, obtaining evaluation results of targeted governance suggestions, and providing feedback to dynamically update the parameters of the hybrid algorithm model include: Set up general quantitative assessment indicators, specific assessment indicators for each anomaly type, and assessment indicators for the governance process; The weighted calculation of the evaluation scores obtained according to various quantitative evaluation indicators yields the final evaluation score for targeted governance recommendations. The anomaly detection results, root causes, governance measures, results, and evaluation results of each branch are archived to generate new cases and fed back to the historical case library. The multi-source data corresponding to the governance results in the new cases with final evaluation scores greater than the preset score threshold are used as incremental data to train and update the parameters of the hybrid algorithm model.

[0019] By adopting the above approach, targeted governance recommendations are quantitatively evaluated to obtain a final evaluation score. The relevant information is then archived to generate new cases and fed back into the historical case database. At the same time, the parameters of the hybrid algorithm model are updated using new case data that meet the conditions, continuously optimizing the anomaly detection model and improving detection accuracy and governance effectiveness.

[0020] Secondly, this application provides a multi-source data fusion wind turbine anomaly detection development platform, including: a data middle platform and an anomaly detection and governance platform; the data middle platform includes a multi-source data acquisition and storage unit, and the anomaly detection and governance platform includes a multi-source data processing unit, an anomaly detection unit, a governance suggestion unit, and a feedback optimization unit; The multi-source data acquisition and storage unit is used to acquire SCADA data, laser ranging data and event recording data of the wind turbine. The multi-source data processing unit is used to clean, align, and standardize the collected data to form a multi-source dataset in a unified format. The anomaly detection unit is used to perform detection and analysis on multi-source datasets using a hybrid algorithm model, and output anomaly detection results including anomaly type and location information. The hybrid algorithm model extracts features from the input multi-source dataset and incorporates them into a global feature pool. It uses a shared feature encoder to extract global general feature vectors, and a multi-task branch decoder to complete the anomaly detection results of each task branch. The anomaly detection results of each branch are verified based on cross-validation rules. The decoder of each task branch completes the hybrid adaptation of the detection algorithm according to the corresponding anomaly type preset hybrid strategy. The training process of the hybrid algorithm model includes: selecting the main wind turbine model, obtaining historical operating data of the main turbine model under normal and abnormal conditions in different scenarios, and training the shared feature encoder; designing a joint loss function, establishing exclusive training data for different wind turbine models under different scenario conditions, and jointly fine-tuning the multi-task branch decoder. The governance suggestion unit is used to generate targeted governance suggestions for different anomaly types based on the anomaly detection results, combined with fault tree analysis and root cause analysis methods. The feedback optimization unit is used to acquire SCADA data, laser ranging data and event record data of wind turbines during the governance period, set quantitative evaluation indicators, obtain evaluation results of targeted governance suggestions and provide feedback, so as to dynamically update the parameters of the hybrid algorithm model.

[0021] By adopting the above solution, we can achieve effective fusion of multi-source data from wind turbine units, accurate location of anomalies, intelligent application of closed-loop management, and improved cross-model adaptability.

[0022] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.

[0023] Fourthly, this application provides a computer device, the computer device including a memory, a processor and a program stored in the memory and executable thereon, the program being executed by the processor to implement the steps of the method described above.

[0024] In summary, this application has the following beneficial effects: 1. By standardizing and integrating multi-source heterogeneous data, a high-quality data foundation is provided for anomaly detection; a hybrid model is used to collaboratively cover multiple types of anomalies such as blade, yaw, and power generation performance, accurately locating the causes of anomalies; and targeted remediation suggestions are generated by combining fault tree analysis and root cause analysis methods. Data during the remediation period is acquired and evaluated to update model parameters, thus achieving full automation of the anomaly detection, remediation plan generation, and effect evaluation process. 2. In the process of anomaly detection using a hybrid model, from the perspective of feature pool construction, a three-level classification feature extraction strategy is set up and feature extraction and derived feature extraction are performed to mine more comprehensive feature information. From the perspective of feature extraction, a shared feature encoder is used to adaptively filter and fuse the extracted features to obtain a general feature vector. The multi-task branch decoder is combined with an appropriate algorithm combination for different anomaly type branches to effectively complete the detection output of multiple anomaly types. Attached Figure Description

[0025] Figure 1 This is a flowchart of the wind turbine anomaly detection method based on multi-source data fusion described in a specific embodiment; Figure 2 This is a schematic diagram illustrating the design principle of the data acquisition system in the wind turbine anomaly detection method based on multi-source data fusion described in a specific embodiment. Figure 3 This is a schematic diagram of the structure of the wind turbine anomaly detection development platform based on multi-source data fusion described in a specific embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] like Figure 1 As shown in the figure, this application discloses a method for detecting anomalies in wind turbines using multi-source data fusion, including steps such as data acquisition and fusion, anomaly detection output, governance suggestion generation, and governance result feedback optimization. By using multi-source data fusion to detect anomalies in wind turbines and implementing closed-loop management, the method achieves the effects of improving the accuracy of anomaly detection, enabling cross-model adaptation, and providing precise governance suggestions. The following is a further detailed description of each step of this application.

[0028] S1 collects multi-source data, including SCADA data, laser ranging data, and event log data from wind turbine generators.

[0029] like Figure 2 As shown, a data acquisition system is constructed. Specifically, various sensors installed on the wind turbine, such as wind speed sensors, power sensors, and temperature sensors, can be used for data acquisition via SCADA, obtaining second / minute-level data such as wind direction, wind speed, power, and rotational speed. It can also interface with anemometer towers to acquire data such as wind speed, wind direction, and turbulence intensity. A portable laser rangefinder (with a measurement range of 50m-300m, angular accuracy better than ±0.1°, distance accuracy better than 10cm, supporting simultaneous measurement of at least two targets, and data transmitted in real-time via gigabit Ethernet) is selected to acquire laser ranging data, including data on blade pitch angle deviation, tower top vibration, and impeller rotational speed. An event logging system is used to record event data, such as fault and maintenance records, as well as basic turbine information (model, rated parameters, commissioning time, etc.). The multi-source data acquired in real-time is integrated and transmitted through a data transmission line. The data transmission line can be a wired network, such as Ethernet, or a wireless network, such as Wi-Fi or Bluetooth.

[0030] like Figure 2 As shown, to better complete multi-source data acquisition, a unified data standard is established. Specifically, the acquisition frequency (e.g., SCADA data ≤ 30 seconds / data, laser ranging data > 100kHz), accuracy requirements (wind speed accuracy ±0.1m / s, power accuracy ±1%), and format specifications (timestamps are unified to UTC format, and numerical units are standardized) for various data types are defined; data semantic mapping rules are established to unify the definition and interpretation of key data items such as "power" and "wind speed". In addition, algorithm selection and fusion can be pre-set to assist in subsequent anomaly detection.

[0031] S2. The collected data is cleaned, aligned, and standardized to form a multi-source dataset with a unified format.

[0032] To better assist in anomaly detection, the collected multi-source data undergoes preprocessing. Specifically, data cleaning involves using a combination of statistical analysis and machine learning algorithms to identify and remove outliers and duplicates, such as negative wind speeds or wind speeds exceeding the cutoff speed, and negative power values. Missing values ​​are handled using mean interpolation, regression interpolation, or LSTM time series prediction imputation. Data alignment ensures the continuity of the time series by interpolating missing values. Data standardization uses Z-Score standardization to eliminate dimensional differences, outputting a high-confidence dataset. Data transformation involves format conversion, unit conversion (e.g., unifying different power units to kW), and encoding conversion to transform unstructured data (event record text) into structured data. Data quality checks establish rules for completeness (missing rate ≤2%), consistency (no logical contradictions), and accuracy (deviation from standard value ≤5%), generating a data quality report.

[0033] S3. Use a hybrid algorithm model to perform anomaly detection analysis on multi-source datasets and output the anomaly detection results.

[0034] Specifically, existing anomaly detection methods for multi-source datasets typically employ relatively simple machine learning algorithms, lacking adaptive detection designs for different wind turbine models, wind turbines operating in different scenarios (different service life, wind farm environmental data), and different anomaly types. This often results in significant errors in the detection results. To achieve more accurate anomaly detection for wind turbines, a four-layer algorithm system is designed, encompassing basic algorithm selection, algorithm hybridization strategies, cross-scenario learning transfer, and multi-detection collaborative fusion. This system adapts to differences in models, service life, and geographical climate, and enables collaborative verification and joint learning of multiple detection branches.

[0035] The specific steps for anomaly detection and analysis using a hybrid algorithm model are as follows: First, features are extracted from the input multi-source dataset and incorporated into a global feature pool. To comprehensively and specifically mine feature information for accurate anomaly analysis, a three-level classification feature extraction strategy is implemented to systematically classify the features of the multi-source dataset. For each level of classification feature extraction, both shared original features and specific derived features are extracted. Specifically, this includes: Firstly, a three-tiered feature extraction strategy is established, including: environmental investigation features (such as ambient temperature and air pressure), core performance features (such as power generation and wind speed), and auxiliary features (pitch characteristics, temperature characteristics of major components, etc.). This three-tiered classification (environmental investigation → core performance → auxiliary features) corresponds to the following steps in comprehensive analysis: eliminating interference (by extracting environmental investigation features, eliminating environmental interference not caused by the unit itself, and distinguishing between environmental fluctuations and unit anomalies) → locating core anomalies (by extracting core performance features, capturing the core operating patterns of the unit, locating anomalies in core components / systems, and directly associating them with the detection target) → improving the logic of anomaly tracing (extracting auxiliary features, improving anomaly tracing, supplementing the deficiencies of core performance features, and locating the root cause of anomalies, such as aging, climate, and multi-component linkage), thus achieving comprehensive feature acquisition.

[0036] Among these, environmental exclusion feature extraction corresponds to wind resource analysis, completing unit wind speed frequency analysis, unit wind rose analysis, and unit average wind speed analysis, and extracting environmental exclusion features such as wind speed frequency, wind rose, and average wind speed. Core performance feature extraction corresponds to power curve analysis, rated performance analysis, and other performance feature analysis, extracting power curve-related features, rated power and wind speed-related features, and other performance features (CP / λ, speed-torque, tip speed ratio, etc.). Auxiliary feature extraction corresponds to strain pitch characteristic analysis, temperature performance analysis, and large component temperature rise analysis, obtaining auxiliary features related to pitch and temperature.

[0037] Secondly, for each level of classification features, common shared original features are pre-extracted, and then exclusive derived features are extracted for each type of anomaly and incorporated into the global feature pool.

[0038] Specifically, for the first-level environmental investigation features, general shared original features are extracted. These features form the basis of environmental investigation and do not require adjustment for individual detections; they are directly included in the global feature pool for use by all detection branches. Examples include: wind speed related features: frequency of each wind speed interval, hourly / daily / monthly average wind speed, wind speed standard deviation, etc.; wind direction related features: interval wind direction frequency, dominant wind direction, wind direction duration, etc.; wind field environment related features: air density, ambient temperature, humidity, air pressure, number of extreme weather days per year, etc. Then, specific derived features are extracted for each anomaly type and included in the feature pool. For example, the first derived feature characterizes the association between the corresponding detection target and environmental features for each anomaly type. The core is to associate environmental features with the detection target and quantify the degree of environmental interference.

[0039] In this embodiment, taking the detection targets corresponding to three types of anomalies—abnormal power generation performance, blade aerodynamic imbalance, and yaw deviation—as examples, the first derived feature corresponding to each anomaly type is extracted. For example, the first derived feature for abnormal power generation performance includes: wind speed difference between adjacent wind turbines, wind speed-power correlation coefficient, and temperature difference between ambient temperature and major components, to exclude power / temperature fluctuations caused by low regional wind speed and sudden changes in ambient temperature; the first derived feature for abnormal blade aerodynamic imbalance includes: correlation coefficient between ambient wind speed and tower vibration, air density and tip speed ratio correction value, and blade torsion angle deviation after extreme weather, to exclude sudden changes in ambient wind speed, differences in air density at high altitudes, and temporary interference from extreme weather; the first derived feature for abnormal yaw deviation includes: wind rose diagram entropy value, difference between the prevailing wind direction and the turbine's windward angle, and difference in windward angle between adjacent wind turbines, to exclude windward deviation caused by chaotic wind direction and errors in the global wind direction sensor.

[0040] Specifically, for the second-level core performance features, general shared original features are extracted. These features directly reflect the core operating status of the unit. First, dimensionless normalization is performed according to the unit's design parameters, and then they are included in the global feature pool. Examples include: power-related features: active power, reactive power, raw power curve data (wind speed-power correspondence), etc.; speed / speed ratio-related features: generator speed, tip speed, tip speed ratio, etc.; angle-related features: pitch angle, windward angle, blade torsion angle, etc.; and operating status-related features: unit operating mode, percentage of effective power generation time, etc. Then, specific derived features are extracted for each anomaly type and included in the feature pool, such as the second derived feature characterizing the correlation between core performance and the corresponding degree of anomaly of the anomaly type.

[0041] In this embodiment, a second derived feature corresponding to each anomaly type is extracted. For example, the (exclusive) second derived feature for the power generation performance anomaly type includes: power curve deviation rate for each wind speed range, blade tip speed ratio-power correlation coefficient, power compliance rate at rated wind speed, and anemometer anomaly prediction value, to quantify the degree of core performance deviation; the second derived feature for the blade aerodynamic imbalance anomaly type includes: three-blade pitch angle difference, average blade torsion angle deviation, aerodynamic imbalance prediction value, and blade tip speed ratio-tower torsional vibration correlation coefficient, to quantify the core aerodynamic performance anomaly of the blade; the second derived feature for the yaw wind deviation anomaly type includes: wind deviation angle, crosswind power attenuation rate, fixed deviation prediction value, and yaw response lag wind deviation, to quantify the core wind performance anomaly of yaw.

[0042] Specifically, for the third-level auxiliary features, general shared original features are extracted and incorporated into the global feature pool. These features supplement the core performance features. The core logic is to extract original features from auxiliary systems (pitch system, temperature control system) / unit aging / climate scenarios, and construct linked and related derived features to support anomaly root cause tracing and collaborative verification. Examples include: temperature-related features: gearbox oil temperature, generator stator temperature, converter temperature, ambient temperature sensor data; pitch-related features: pitch motor current, target pitch angle, actual pitch angle, pitch response time; aging / maintenance-related features: years of operation, cumulative power generation, major component replacement records, maintenance frequency; vibration / other-related features: tower torsion / vertical vibration, yaw motor current, generator torque. Furthermore, specific derived features are extracted for each anomaly type and incorporated into the feature pool, such as the third derived features characterizing the correlation between auxiliary features and core performance.

[0043] In this embodiment, a third derived feature corresponding to each anomaly type is extracted. For example, the third derived feature for the power generation performance anomaly type includes: the temperature rise slope of large components, the correlation coefficient between pitch motor current and pitch angle, the difference between ambient temperature sensor and adjacent wind turbine, and the aging correction coefficient, to support the root cause tracing of the anomaly; the third derived feature for the blade aerodynamic imbalance anomaly type includes: pitch response lag time, the correlation coefficient between tower vibration amplitude and operating years, the torsion angle deviation after blade maintenance, and the estimated value of humidity and blade surface fouling, to support the root cause tracing of the blade imbalance; the third derived feature for the yaw wind deviation anomaly type includes: yaw motor current fluctuation rate, yaw response lag time, yaw lubricating oil viscosity correction value under low temperature environment, and the wind deviation value after yaw maintenance, to support the root cause tracing of the yaw deviation.

[0044] Third, features with variance less than the variance threshold, importance less than the importance threshold, and correlation with core performance less than the correlation threshold are removed. The extracted general shared original features and various anomaly detection-specific derived features are included in the feature pool and divided into three-level classification sub-feature pools. Each of the three-level classification sub-feature pools contains the corresponding general shared original features and various anomaly detection-specific derived features, and is labeled with normalization markers, scene adaptation markers, environmental interference amount, core performance deviation amount, and auxiliary correlation value.

[0045] Specifically, for the screening and inclusion of environmental investigation features, low-fluctuation features with variance < variance threshold (0.01) are removed. The corresponding general shared original features and the first derived features of the three types of detection are included in the environmental investigation sub-pool of the global feature pool, and an environmental interference quantification value (0-1, higher values ​​indicate greater environmental interference) is labeled. For the screening and inclusion of second-level core performance features, redundant features with importance < importance threshold (1%) are removed. The corresponding general shared original features and the second derived features of the three types of detection are included in the core performance sub-pool of the global feature pool, and a core performance deviation value (0-1, higher values ​​indicate more significant anomalies) is labeled. For the screening and inclusion of third-level auxiliary features, weakly correlated features with correlation < correlation threshold (0.3) to core performance features are removed. The corresponding general shared original features and the third derived features of the three types of detection are included in the auxiliary feature sub-pool of the global feature pool, and an auxiliary correlation value (0-1, higher values ​​indicate stronger correlation with core anomalies) is labeled.

[0046] Secondly, a joint learning model of a shared feature encoder and a multi-task branch decoder is set as the basic architecture of the hybrid model. The inputs and outputs of the shared feature encoder and the multi-task branch decoder are clearly defined to obtain anomaly detection results, including anomaly type, anomaly level, and specific anomaly location. The anomaly detection results of each branch are verified based on cross-validation rules. Therefore, the feature extraction and detection analysis output process of the hybrid model is as follows: Input layer → Shared feature encoder (extracts global general features) → General feature vector layer → Multi-task branch decoder (power generation branch + blade branch + yaw branch) → Output layer.

[0047] Firstly, the model input of the input layer comes from the features of the global feature pool after environmental investigation, core performance, and auxiliary feature selection. It is divided into two types of data (temporal features + static / structured features), which are concatenated and then input into the model. Input constraints are set, including: filtering samples with environmental interference quantization values ​​> 0.5, and using sample balancing with oversampling (a few abnormal samples) and undersampling (most normal samples).

[0048] Secondly, a shared feature encoder is used to extract global universal feature vectors.

[0049] Specifically, the shared feature encoder extracts a universal feature vector applicable to all three types of detection from the mixed features of the input. Informer, CNN, and TabNet are selected as the operating models for the shared feature encoder. The steps for extracting the universal feature vector using Informer, CNN, and TabNet include: The Informer model is used to process the temporal features in the multi-source dataset. Through the encoder-decoder structure, the temporal feature values ​​for the next hour are predicted. At the same time, the hidden correlation patterns of the temporal features (such as the synchronous change pattern of blade tip speed ratio and temperature) are extracted, and a general temporal feature vector is extracted. This vector contains the temporal patterns of unit operation, such as the temperature trend of power generation performance, the timing of blade pitch angle, and the timing of yaw wind angle.

[0050] A CNN model is used to process static or structured features in multi-source datasets. Different convolutional layers are used to extract spatially general feature vectors with varying granularities (scales). For example, three convolutional layers with kernel sizes of 3, 5, and 7 are used to extract fine-grained patterns of a single feature (e.g., local fluctuations in pitch angle difference), correlation patterns between two features (e.g., the correlation between wind speed frequency and power deviation), and global correlation patterns among multiple features (e.g., the linkage between temperature, speed, and power). Dimensionality is reduced using pooling layers, outputting a spatially general feature vector that contains the static correlation patterns of wind turbine operation.

[0051] The TabNet model is used to concatenate the extracted temporal and spatial general features, and TabNet feature selection masking is enabled. Weighted preset features are applied to different scenarios and detection tasks. An attention mechanism is then used to fuse the temporal and spatial general features to achieve adaptive feature selection and fusion, outputting the final general feature vector. Specifically, feature selection encoding can be performed by weighting preset high-value features (e.g., power deviation rate for power generation performance detection, and pitch angle difference for blade detection) for different scenarios (e.g., combinations of turbine type / year / wind field climate environment) and different detection tasks (three types of detection), while maintaining shared underlying weights.

[0052] Third, the multi-task branch decoder is combined to complete the anomaly detection results of each of the multi-task branches.

[0053] Specifically, in the multi-task branch decoder, each task branch corresponds to one type of anomaly detection. For different anomaly detection types, a corresponding preset mixing strategy can be set to mix the selected base model, thereby adapting the mixed model to specific anomaly detection types. In this embodiment, the multi-task branch decoder corresponds to three types of detection, with each branch corresponding to one type of detection. The input is the globally universal feature vector output by the encoder and the specific features of the corresponding type of detection, and the output is the anomaly result of each branch. The algorithm adaptation for each type of detection is described below.

[0054] For tasks involving abnormal power generation performance, a branch decoder is adapted to both CNN and XGBoost algorithms to construct a multi-scenario anomaly detection model, covering various anomaly scenarios such as anemometer anomalies, yaw system anomalies, pitch system anomalies, large component temperature rise anomalies, and environmental temperature sensing anomalies. Specifically, the input to this branch model includes: a globally common feature vector output by the shared encoder, universal shared original features extracted from the global feature pool, and power generation performance anomaly detection-specific derived features. The concatenation operation involves concatenating the globally common feature vector output by the shared encoder, the universal shared original features extracted from the global feature pool, and the power generation performance anomaly detection-specific derived features (first, second, and third derived features) and inputting them into the CNN algorithm model to obtain a preliminary estimate, i.e., a preliminary estimated anomaly type. Then, based on the input features, the preliminary estimate is added and input into the XGBoost algorithm model. The XGBoost algorithm performs a multi-class classification task, outputting each anomaly class and its confidence score. The XGBoost algorithm model uses a multi-class cross-entropy loss, trains on data from each class of abnormal and normal states, and outputs the confidence score (0-1) for each anomaly, the final anomaly type (the class with the highest confidence), and a global anomaly score.

[0055] For the task of detecting blade aerodynamic imbalance anomalies, a branch decoder is adapted to random forest and gradient boosting tree algorithms to detect and obtain the degree of blade aerodynamic imbalance and the numerical values ​​of pitch angle deviation. Specifically, the input of this branch model includes: the global general feature vector output by the shared encoder, and the general shared original features extracted from the global feature pool and the blade aerodynamic imbalance anomaly-specific derived features (first, second, and third derived features). The concatenation operation involves concatenating the global general feature vector output by the shared encoder, the general shared original features extracted from the global feature pool, and the blade aerodynamic imbalance anomaly-specific derived features into the random forest algorithm model to obtain preliminary estimates, i.e., the degree of imbalance and the pitch angle deviation. Then, based on the preliminary estimates, valid samples are selected (extreme values ​​are removed) and input into the gradient boosting tree algorithm model. A leaf-by-leaf growth strategy is adopted, with a loss function of MSE, an iteration count of 1000, and a learning rate of 0.01, outputting a refined estimate and a final estimate. A weighted fusion (e.g., initial estimate * 0.3 + final estimate * 0.7) is then performed to obtain the final anomaly detection results, including the degree of imbalance and the pitch angle deviation.

[0056] For tasks involving yaw deviation anomalies, the branch decoder adapts to both random forest and LSTM algorithm models, outputting a fixed wind deviation value and a dynamic distribution of the deviation. Specifically, the inputs to the branch model include: the globally common feature vector output by the shared encoder, the general shared original features extracted from the global feature pool, and the yaw deviation anomaly type-specific derived features (first, second, and third derived features); the concatenation operation concatenates the globally common feature vector output by the shared encoder and the yaw deviation anomaly type-specific derived features extracted from the global feature pool. The following methods are employed: First, the global general feature vector output by the shared encoder and the static feature tensor from the yaw-wind deviation anomaly type-specific derived features extracted from the global feature pool are concatenated and input into a random forest algorithm model to train a fixed deviation value regression model, outputting a fixed initial deviation value and a static anomaly score. Second, the global general feature vector output by the shared encoder and the temporal feature tensor from the yaw-wind deviation anomaly type-specific derived features extracted from the global feature pool are concatenated and input into a (improved) LSTM algorithm model (Bi-LSTM+Attention) to capture forward / backward temporal dependencies, outputting a dynamic deviation distribution (hourly mean / variance / temporal curve). Third, fusion verification is performed based on the output fixed initial deviation value and dynamic deviation distribution. For example, if the difference between the dynamic deviation mean and the fixed deviation value is ≤1°, it is considered valid, and the verified deviation value is output.

[0057] Fourth, the abnormal detection results of each branch output are verified based on the cross-validation rules.

[0058] Specifically, the anomaly detection results from the different branches can be further weighted to obtain the final global anomaly score. To further ensure the accuracy of the output, collaborative verification can be performed based on the outputs of each branch. The collaborative verification steps are explained below: Considering that the results of the three types of detection branches are not independent, but have clear causal chains, accompanying relationships, and complementary relationships, anomaly correlation graphs are constructed based on the historical anomaly detection results of different anomaly types to obtain the logical basis for cross-validation. For example: Power generation performance anomaly: pitch system anomaly, yaw system anomaly, other strongly correlated branch anomalies include: blade aerodynamic imbalance (pitch anomaly), yaw wind deviation (yaw anomaly), the corresponding correlation logic is: blade aerodynamic imbalance → pitch angle deviation → pitch system response anomaly → power generation performance degradation; yaw wind deviation → crosswind power attenuation → power generation performance degradation. Blade aerodynamic imbalance: pitch angle deviation, aerodynamic imbalance, other strongly correlated branch anomalies include: power generation performance anomaly (pitch anomaly). Yaw wind deviation: fixed wind deviation, dynamic deviation, other strongly correlated branch anomalies include: power generation performance anomaly (yaw anomaly).

[0059] The core dimensions of cross-validation are defined, including result consistency verification, feature complementarity verification, and logical constraint verification. Specifically, result consistency verification verifies whether the anomaly types, anomaly scores, and anomaly levels of the three branches conform to the causal relationships in the correlation graph; feature complementarity verification verifies whether the deviation values, trends, and correlations of shared features in the global feature pool shared by the three branches are consistent across different branches, avoiding feature contradictions between branches; logical constraint verification includes verifying whether there are logical contradictions in the branch results based on the operating rules of wind turbines, such as: high blade imbalance score but no pitch anomaly in power generation performance, or large yaw deviation but no crosswind power attenuation in power generation performance.

[0060] Based on the obtained cross-validation logic and core dimensions, core cross-validation rules are formulated. Each rule includes validation conditions, judgment logic, quantitative score, and conclusion. The core cross-validation rules include: causal forward validation, reverse validation, and feature complementarity validation between power generation performance pitch anomalies and blade aerodynamic imbalance; causal forward validation, reverse validation, and feature complementarity validation between power generation performance yaw anomalies and yaw-wind deviation; conflict-free validation (whether there is a contradiction) and trend consistency validation (whether the shared feature trends are consistent) between blade aerodynamic imbalance and yaw-wind deviation; and global logical constraint validation (e.g., anomaly-free consistency validation, single anomaly exclusive validation, whether branch results meet scenario thresholds).

[0061] Specifically, taking the causal positive verification between abnormal pitch control performance and blade aerodynamic imbalance as an example, i.e., blade imbalance → abnormal pitch control performance, the corresponding verification conditions include: the blade branch output is aerodynamic imbalance degree ≥ 60 points or pitch angle deviation ≥ 0.5°; the judgment logic is: the power generation branch must detect abnormal operation of the pitch control system and the confidence level is ≥ 70%; the quantitative scoring is: the power generation branch detects pitch control abnormality (10 points), the power generation branch does not detect pitch control abnormality but has other types of abnormalities (5 points), the power generation branch does not detect any abnormality (0 points). If the score is greater than 7 points, the conclusion is: the blade imbalance causes the pitch control abnormality; if the score is less than 5 points, the blade imbalance result is marked as questionable.

[0062] For example, the complementary verification of features between abnormal pitch control in power generation performance and blade aerodynamic imbalance (shared feature: pitch angle deviation) is set with the following verification conditions: the blade branch outputs a pitch angle deviation value of X°, and the power generation branch extracts a pitch angle deviation feature value of Y°. The judgment logic is: if |XY|≤0.2°, it indicates that the features are consistent; otherwise, it is contradictory. The quantitative scoring is: the difference is less than 0.2° (10 points), (0.2°, 0.5°) (5 points), and greater than 0.5° (0 points). The conclusion is: if the score is ≥5 points, the feature is reliable; if the score is <5 points, the pitch angle sensor data needs to be reviewed.

[0063] The system matches the anomaly detection results output by each branch with the established core cross-validation rules. Cross-validation is then performed between the corresponding branches based on the matched core cross-validation rules. The total cross-validation score between each branch is calculated and compared with the preset cross-validation score threshold to determine whether the cross-validation passes. If the cross-validation fails, a detection review prompt is generated, and the detection of each branch is repeated.

[0064] Furthermore, the training process for the hybrid algorithm model includes: selecting the main wind turbine model, acquiring historical operating data of the main turbine model under normal and abnormal conditions in different scenarios (years / wind farm climate), training a shared feature encoder, i.e., training Informer, CNN, and TabNet, with the goal of minimizing the temporal prediction loss and feature fusion loss. After training, the shared encoder is frozen. A joint loss function is designed, and dedicated training data is established for different wind turbine models under different scenario conditions. The multi-task branch decoder is jointly fine-tuned, and grouped by model / operating years / wind farm climate. Different scenario adaptations are performed, and the decoder threshold and weights are optimized. The training termination condition is: total loss convergence (loss decrease of <0.001 for 100 consecutive iterations), and the accuracy of all three detection tasks is ≥85%.

[0065] The formula for the joint loss function is: L = α * L_1 + β * L_2 + γ * L_3, where L_1 is the power generation performance decoder loss (multi-class cross-entropy loss), and the weight α can be set according to the scenario, such as 0.5, which corresponds to the scenario where this loss has the greatest impact; L_2 is the blade aerodynamic decoder loss (regression MSE loss), and the weight β can be set according to the scenario, such as 0.3; L_3 is the yaw decoder loss (regression MSE loss and time series MSE loss), and the weight γ can be set according to the scenario, such as 0.2.

[0066] S4. Based on anomaly detection results and historical case database, combined with fault tree analysis and root cause analysis methods, generate early warning prompts and targeted governance suggestions for anomaly types.

[0067] Specifically, the first step is to use the anomaly detection results from each branch as input, and then use fault tree analysis to decompose and obtain the top event, intermediate events, and bottom events. For example, if the anomaly detection results show an anomaly in the anemometer, then FTA fault tree decomposition is used to obtain the top event: anemometer anomaly; intermediate events: sensor failure, installation deviation, signal transmission anomaly, environmental interference; and bottom events: sensor probe dust / icing, sensor aging, installation angle deviation greater than 5°, loose wiring, lightning strike damage, high humidity corrosion, etc.

[0068] Then, the root cause of the anomaly is determined through root cause analysis. For example, using the 5 Whys method, the corresponding root cause analysis path is: Anemometer output abnormal? → The difference between the wind speed measurement value and the adjacent wind turbine is greater than 3 m / s; Large measurement deviation? → The sensor probe cannot capture the wind speed normally; Probe abnormal? → Either dust accumulation / icing, or probe aging; Why is there dust accumulation / aging? → No regular cleaning and maintenance, or the probe has not been replaced after more than 3 years of operation; No maintenance / no replacement? → The operation and maintenance plan does not cover the regular inspection of the anemometer, or does not associate aging characteristics; Thus, the root cause is found to be: lack of regular cleaning and maintenance, or sensor aging and failure to replace in time.

[0069] Finally, based on the anomaly level priority settings, different levels of early warning information are generated, and different levels of early warning prompts are provided for different anomaly types. The system matches historical case libraries and generates targeted governance suggestions based on wind turbine model, anomaly type, and scenario type, with different anomaly levels, scenario adaptations, and step-by-step governance steps.

[0070] For example, the top-ranked similar case is: CAES1: 2MW model, coastal wind farm (high humidity), 4 years of operation, wind speed deviation 1-1.5m / s (slight anomaly); remedial measures: Step 1: Shut down and clean the probe for accumulated dirt; Step 2: Calibrate the installation angle; Step 3: Replace the aging wiring; remedial effect: wind speed deviation ≤0.5m / s, correlation coefficient ≥0.85, no repeated warnings.

[0071] Furthermore, to further ensure the accuracy of root cause analysis, a pre-constructed knowledge graph of wind turbine anomalies can be built, including: defining equipment entities (wind turbine model, individual unit, core components and auxiliary components), anomaly entities (anomaly types and anomaly levels of the three types of detection branches), analysis entities (FTA events, RCA root causes), scenario entities (operating years, wind farm environment), governance entities (governance measures), and case entities (historical governance cases); defining relationships (fault propagation relationships, scenario adaptation relationships, governance association relationships, collaborative verification relationships, etc.); and adding supplementary attributes (quantitative information of supplementary entities / relationships) and inference rules (FTA rules, RCA rules, and collaborative rules). The anomaly detection results, corresponding features, and scenarios output by each branch are taken as input and transformed into entity instances of a wind turbine anomaly knowledge graph. The top event is obtained through fault tree analysis. Starting from the top event, the path traversal from the top event to the bottom event is completed using the wind turbine anomaly knowledge graph. Then, based on root cause analysis, starting from the bottom event reached, a chain-like reasoning from the bottom event to the root cause is completed using the wind turbine anomaly knowledge graph, tracing back to the root cause. Correlated fault reasoning is performed based on the graph relationships of the wind turbine anomaly knowledge graph, automatically identifying anomaly detection results from other branches that are linked to the current branch's output anomaly detection results. Based on the traced root cause, related fault reasoning results, and scenarios, the historical case library is matched, calculating the graph entity similarity and relationship similarity between the current feature and cases in the case library. The top 3 cases with the highest similarity are selected by total score. Combining the root cause, related faults, scenario adaptation, and the governance measures and effects of the matched cases, targeted governance suggestions are generated based on anomaly level, scenario adaptation, and related governance.

[0072] S5. Acquire SCADA data, laser ranging data and event log data of wind turbines during the governance period, set quantitative evaluation indicators, obtain evaluation results of targeted governance suggestions and provide feedback, so as to dynamically update the parameters of the hybrid algorithm model.

[0073] First, general quantitative evaluation indicators, specific evaluation indicators for each anomaly type, and evaluation indicators for the remediation process are established. Specifically, the general quantitative evaluation indicators include: performance recovery rate, anomaly elimination rate (mean anomaly score after remediation / mean anomaly score before remediation), operational stability of key parameters (standard deviation of key parameters after remediation / standard deviation before remediation), and economic benefit improvement rate (difference in daily power generation before and after remediation / daily power generation before remediation). The specific evaluation indicators for each anomaly type include: specific indicators for power generation performance anomalies (wind speed deviation, crosswind power attenuation rate, pitch angle coordination deviation, temperature rise slope, and temperature sensitivity difference, etc.), specific indicators for blade start-up imbalance anomalies (imbalance score after remediation, maximum difference in three-blade pitch angle, etc.), and specific indicators for yaw wind deviation anomalies (fixed wind deviation, mean dynamic deviation, etc.). The difference between the evaluation and the corresponding parameter target threshold is calculated to quantify the score. The evaluation indicators for the remediation process include: measure matching rate (number of measures actually implemented / number of recommended measures) and remediation timeliness rate (actual execution time / recommended execution time).

[0074] Then, the weighted calculation of the evaluation scores obtained according to various quantitative evaluation indicators yields the final evaluation score for targeted governance recommendations.

[0075] Finally, the anomaly detection results, root causes, governance measures and results, and evaluation results of each branch are archived to generate new cases and fed back to the historical case library. Based on the multi-source data corresponding to the governance results in the new cases with final evaluation scores greater than the preset score threshold, the parameters of the hybrid algorithm model are trained and updated. The model parameters of the three branches in the hybrid algorithm can be incrementally trained to achieve the update purpose.

[0076] like Figure 3 As shown in the figure, this embodiment discloses a multi-source data fusion wind turbine anomaly detection development platform. This platform adopts a B / S architecture and a front-end / back-end separation design. The front-end is developed based on Vue + HTML5 + Node.js, and the back-end uses the Spring Boot 2 framework, supporting cross-platform deployment such as Linux / Windows. Specifically, it includes a data middle platform 100 and an anomaly detection and governance platform 200. The data middle platform 100 includes a multi-source data acquisition and storage unit 110, and the anomaly detection and governance platform 200 includes a multi-source data processing unit 211, an anomaly detection unit 212, a governance suggestion unit 213, and a feedback optimization unit 214.

[0077] The multi-source data acquisition and storage unit 110 is used to acquire and store multi-source data such as SCADA data, laser ranging data and event recording data of the wind turbine. The multi-source data processing unit 211 is used to clean, align and standardize the collected data to form a multi-source dataset in a unified format. The anomaly detection unit 212 is used to perform detection and analysis on multi-source datasets using a hybrid algorithm model, and output anomaly detection results including anomaly type and location information. The hybrid algorithm model extracts features from the input multi-source dataset and incorporates them into a global feature pool. It uses a shared feature encoder to complete the extraction of global general feature vectors, and jointly uses a multi-task branch decoder to complete the detection and output of anomaly detection results for each task branch. It verifies the anomaly detection results output by each branch based on cross-validation rules. The decoder of each task branch completes the hybrid adaptation of the detection algorithm according to the corresponding anomaly type preset hybrid strategy. The training process of the hybrid algorithm model includes: selecting the main wind turbine model, obtaining historical operating data of the main turbine model under normal and abnormal conditions in different scenarios, and training the shared feature encoder; designing a joint loss function, establishing exclusive training data for different wind turbine models under different scenario conditions, and jointly fine-tuning the multi-task branch decoder. The governance suggestion unit 213 is used to generate targeted governance suggestions for different anomaly types based on the anomaly detection results, combined with fault tree analysis and root cause analysis methods. The feedback optimization unit 214 is used to acquire SCADA data, laser ranging data and event record data of wind turbines during the governance period, set quantitative evaluation indicators, obtain evaluation results of targeted governance suggestions and provide feedback, so as to dynamically update the parameters of the hybrid algorithm model.

[0078] In addition, the development platform includes a visualization module 220, which provides display formats such as line charts, bar charts, heatmaps, and 3D model diagrams, supporting data drill-down, zooming, and panning operations. A ledger management module 230 stores basic information and operating records of wind turbine equipment, supporting data import, export, and dynamic updates. A permission management module 240, based on JWT technology, assigns operation permissions according to roles (administrator / maintenance personnel / viewer) to control the scope of data access.

[0079] This application also discloses a computer-readable storage medium.

[0080] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the wind turbine anomaly detection method with multi-source data fusion described above. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0081] This application also discloses a computer device.

[0082] Specifically, the computer device includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed to perform the aforementioned multi-source data fusion method for wind turbine anomaly detection.

[0083] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for wind turbine anomaly detection based on multi-source data fusion, characterized in that, include: Collect SCADA data, laser ranging data, and event log data from wind turbine generators; The collected data is cleaned, aligned, and standardized to form a multi-source dataset in a unified format. A hybrid algorithm model is used to detect and analyze multi-source datasets, and output anomaly detection results including anomaly type and location information. The hybrid algorithm model extracts features from the input multi-source dataset and incorporates them into a global feature pool. It uses a shared feature encoder to complete the extraction of global general feature vectors, and combines a multi-task branch decoder to complete the detection and output of anomaly detection results for each of the multi-task branches. The anomaly detection results output by each branch are verified based on cross-validation rules. Furthermore, the decoder of each task branch completes the hybrid adaptation of the detection algorithm according to the preset hybrid strategy for the corresponding anomaly type; The training process of the hybrid algorithm model includes: selecting the main wind turbine model, obtaining historical operating data of the main turbine model under normal and abnormal conditions in different scenarios, and training a shared feature encoder; designing a joint loss function, establishing exclusive training data for different wind turbine models under different scenario conditions, and jointly fine-tuning the multi-task branch decoder. The joint multi-task branch decoder completes the anomaly detection results for each of the multiple task branches, including: For the task branch decoder of abnormal power generation performance, the CNN algorithm and the XGBoost algorithm are adapted. The global general feature vector output by the shared encoder, the general shared original features extracted from the global feature pool, and the power generation performance abnormality detection-specific derived features are concatenated and input into the CNN algorithm model to obtain the preliminary estimate. Then, the preliminary estimate is added and input into the XGBoost algorithm model. The XGBoost algorithm is used to perform a multi-classification task and output the various abnormal classes and their confidence scores. For the blade aerodynamic imbalance anomaly type task branch decoder, the random forest and gradient boosting tree algorithms are adapted. The global general feature vector output by the shared encoder, the general shared original features extracted from the global feature pool, and the blade aerodynamic imbalance anomaly type-specific derived features are concatenated and input into the random forest algorithm model to obtain the preliminary estimate; then the preliminary estimate is input into the gradient boosting tree algorithm model to output the post-estimate; and finally, the weighted fusion is used to obtain the final anomaly detection result. For the yaw deviation anomaly type task branch decoder, the random forest algorithm model and the LSTM algorithm model are adapted. The global general feature vector output by the shared encoder and the static feature tensor from the yaw deviation anomaly type-specific derived features extracted from the global feature pool are concatenated and input into the random forest algorithm model to output a fixed initial deviation value. The global general feature vector output by the shared encoder and the temporal feature tensor from the general shared original features extracted from the global feature pool and the yaw deviation anomaly type-specific derived features are concatenated and input into the LSTM algorithm model to output a dynamic deviation distribution. Based on the output fixed initial deviation value and dynamic deviation distribution, fusion verification is performed, and the verified deviation value is output. Based on anomaly detection results and historical case databases, combined with fault tree analysis and root cause analysis methods, early warning prompts and targeted governance suggestions are generated for anomaly types. During the governance period, SCADA data, laser ranging data and event log data of wind turbines are obtained, quantitative evaluation indicators are set, evaluation results of targeted governance suggestions are obtained and fed back, so as to dynamically update the parameters of the hybrid algorithm model.

2. The multi-source data fusion based wind turbine anomaly detection method of claim 1, wherein, The hybrid algorithm model extracts features from the input multi-source dataset and incorporates them into a global feature pool, including: A three-level classification feature extraction strategy is set up, including: environmental investigation features, core performance features, and auxiliary features; For each level of classification features, common shared original features are pre-extracted, and then derivative features are extracted for the exclusive derivative features constructed for each anomaly type. The exclusive derivative features constructed for each anomaly type include: a first derivative feature characterizing the association between the corresponding detection target and environmental features of the anomaly type, a second derivative feature characterizing the correlation between core performance and the corresponding anomaly degree of the anomaly type, and a third derivative feature characterizing the association between auxiliary features and core performance. The auxiliary features include pitch features and large component temperature features. Features with variance less than the variance threshold, importance less than the importance threshold, and correlation with core performance less than the correlation threshold are removed. The extracted general shared original features and various anomaly detection-specific derived features are included in the feature pool and divided into three-level classification sub-feature pools. Each of the three-level classification sub-feature pools contains the corresponding general shared original features and various anomaly detection-specific derived features, and is labeled with normalization markers, scene adaptation markers, environmental interference amount, core performance deviation amount, and auxiliary correlation value.

3. The multi-source data fusion based wind turbine anomaly detection method of claim 1, wherein, The process of extracting the global universal feature vector using a shared feature encoder includes: Informer, CNN, and TabNet were selected as the running models for the shared feature encoder. The steps for extracting the global general feature vector using Informer, CNN, and TabNet include: The Informer model is used to process the temporal features in multi-source datasets and extract a general temporal feature vector, including predicted future temporal feature values. A CNN model is used to process static or structured features in multi-source datasets, and spatial general feature vectors at different scales are extracted by setting different convolutional layers. The TabNet model is used to concatenate the extracted temporal and spatial general features, and the feature selection mask of TabNet is enabled. Pre-weighted features are applied to different scenarios and detection tasks. The temporal and spatial general features are then fused through an attention mechanism to achieve adaptive feature selection and fusion, and the final general feature vector is output.

4. The multi-source data fusion wind turbine anomaly detection method of claim 1, wherein, The verification of anomaly detection results for each branch output based on cross-validation rules includes: An anomaly correlation graph is constructed based on historical anomaly detection results of different anomaly types to obtain the logical basis for cross-validation. Define the core dimensions of cross-validation, including result consistency verification, feature complementarity verification, and logical constraint verification; Based on the obtained cross-validation logic foundation and core dimensions of cross-validation, core cross-validation rules are formulated. Each rule includes validation conditions, judgment logic, quantitative score, and conclusion. The core cross-validation rules include: causal forward validation, reverse validation, and feature complementarity validation between power generation performance pitch anomaly and blade aerodynamic imbalance; causal forward validation, reverse validation, and feature complementarity validation between power generation performance yaw anomaly and yaw wind deviation; conflict-free validation and trend consistency validation between blade aerodynamic imbalance and yaw wind deviation; and global logic constraint validation. The system matches the anomaly detection results output by each branch with the established core cross-validation rules. Cross-validation is then performed between the corresponding branches based on the matched core cross-validation rules. The total cross-validation score between each branch is calculated and compared with the preset cross-validation score threshold to determine whether the cross-validation passes. If the cross-validation fails, a detection review prompt is generated.

5. The multi-source data fusion wind turbine anomaly detection method of claim 1, wherein, The method of generating targeted remediation suggestions for different anomaly types based on anomaly detection results and historical case databases, combined with fault tree analysis and root cause analysis, includes: Using the anomaly detection results output from each branch as input, the system obtains the top, intermediate, and bottom events through fault tree analysis; then, it determines the root cause of the anomaly through root cause analysis; it matches the historical case library and generates targeted governance suggestions based on the wind turbine model, wind turbine anomaly type, and wind turbine scenario type, with different levels of anomaly, scenario-specific adaptation, and step-by-step governance.

6. The multi-source data fusion wind turbine anomaly detection method of claim 5, wherein, The process of acquiring SCADA data, laser ranging data, and event log data of wind turbines during the governance period, setting quantitative evaluation indicators, obtaining evaluation results of targeted governance suggestions and providing feedback, and dynamically updating the parameters of the hybrid algorithm model includes: Set up general quantitative assessment indicators, specific assessment indicators for each anomaly type, and assessment indicators for the governance process; The weighted calculation of the evaluation scores obtained according to various quantitative evaluation indicators yields the final evaluation score for targeted governance recommendations. The anomaly detection results, root causes, governance measures, results, and evaluation results of each branch are archived to generate new cases and fed back to the historical case library. The multi-source data corresponding to the governance results in the new cases with final evaluation scores greater than the preset score threshold are used as incremental data to train and update the parameters of the hybrid algorithm model.

7. A multi-source data fusion wind turbine anomaly detection development platform, characterized in that, include: The data platform and the anomaly detection and governance platform; the data platform includes a multi-source data acquisition and storage unit, and the anomaly detection and governance platform includes a multi-source data processing unit, anomaly detection unit, governance suggestion unit, and feedback optimization unit; The multi-source data acquisition and storage unit is used to acquire SCADA data, laser ranging data and event recording data of the wind turbine. The multi-source data processing unit is used to clean, align, and standardize the collected data to form a multi-source dataset in a unified format. The anomaly detection unit is used to perform detection and analysis on multi-source datasets using a hybrid algorithm model, and output anomaly detection results including anomaly type and location information. The hybrid algorithm model extracts features from the input multi-source dataset and incorporates them into a global feature pool. It uses a shared feature encoder to complete the extraction of global general feature vectors, and jointly uses a multi-task branch decoder to complete the detection and output of anomaly detection results for each of the multi-task branches. The anomaly detection results output by each branch are verified based on cross-validation rules. Furthermore, the decoder of each task branch completes the hybrid adaptation of the detection algorithm according to the preset hybrid strategy for the corresponding anomaly type; The training process of the hybrid algorithm model includes: selecting the main wind turbine model, acquiring historical operating data of the main turbine model under normal and abnormal conditions in different scenarios, and training a shared feature encoder; designing a joint loss function, establishing dedicated training data for different wind turbine models under different scenarios, and jointly fine-tuning the multi-task branch decoder; the joint multi-task branch decoder completing the anomaly detection results of each task branch includes: adapting the CNN algorithm and XGBoost algorithm to the task branch decoder for power generation performance anomaly type, concatenating the global general feature vector output by the shared encoder with the general shared original features extracted from the global feature pool and the power generation performance anomaly detection-specific derived features and inputting them into the CNN algorithm model to obtain preliminary estimates; then adding the preliminary estimates and inputting them into the XGBoost algorithm model, performing multi-classification tasks through the XGBoost algorithm, and outputting various anomaly classes and their confidence scores; adapting the random forest and gradient boosting tree algorithms to the task branch decoder for blade aerodynamic imbalance anomaly type, concatenating the global general feature vector output by the shared encoder with the general shared original features extracted from the global feature pool and the power generation performance anomaly detection-specific derived features and inputting them into the CNN algorithm model to obtain preliminary estimates; then adding the preliminary estimates and inputting them into the XGBoost algorithm model, performing multi-classification tasks through the XGBoost algorithm, and outputting various anomaly classes and their confidence scores; and ... The global general feature vector output by the shared encoder, along with the general shared original features extracted from the global feature pool and the blade aerodynamic imbalance anomaly-specific derived features, are concatenated and input into the random forest algorithm model to obtain a preliminary estimate. This preliminary estimate is then input into the gradient boosting tree algorithm model to output a post-estimate. Weighted fusion is performed to obtain the final anomaly detection result. For the yaw-wind deviation anomaly type task branch decoder, the random forest algorithm model and the LSTM algorithm model are adapted. The global general feature vector output by the shared encoder and the static feature tensor from the yaw-wind deviation anomaly-specific derived features extracted from the global feature pool are concatenated and input into the random forest algorithm model to output a fixed initial deviation value. The global general feature vector output by the shared encoder and the temporal feature tensor from the general shared original features extracted from the global feature pool and the yaw-wind deviation anomaly-specific derived features are concatenated and input into the LSTM algorithm model to output a dynamic deviation distribution. Based on the output fixed initial deviation value and dynamic deviation distribution, fusion verification is performed, and the verified deviation value is output. The governance suggestion unit is used to generate targeted governance suggestions for different anomaly types based on the anomaly detection results, combined with fault tree analysis and root cause analysis methods. The feedback optimization unit is used to acquire SCADA data, laser ranging data and event record data of wind turbines during the governance period, set quantitative evaluation indicators, obtain evaluation results of targeted governance suggestions and provide feedback, so as to dynamically update the parameters of the hybrid algorithm model.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 6.

9. A computer device, comprising: The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 6.

Images