Equipment health management method and system for wind generating set

Abstract

The invention provides an equipment health management method and system for a wind generating set. The method belongs to the technical field of wind power generation, and comprises the following steps: acquiring multi-source data of a wind generating set, preprocessing the acquired multi-source data, and integrating the preprocessed multi-source data into a uniform data format by using a data fusion technology; based on domain knowledge and data characteristics, key features reflecting the health state of the equipment are extracted, and a feature subset with the highest prediction value for the health state of the equipment is screened out through a feature selection algorithm. Key components of the wind generating set are monitored in real time through a high-precision sensor network, and comprehensive collection of data is realized in combination with multi-source information such as a historical fault database and meteorological data.

Description

Technical Field

[0001] The invention provides a device health management method and system for a wind generator set, belonging to the technical field of wind power generation. Background Art

[0002] Traditional wind turbine equipment health management mainly relies on regular manual inspection and maintenance, which is not only inefficient but also difficult to capture early signs of equipment failure. In recent years, although some predictive maintenance methods based on data analysis have been proposed, most of these methods are limited to the application of a single model, lacking effective integration of different data sources and comprehensive capture of complex failure modes. In addition, existing methods often ignore the impact of changes in the equipment operating environment on the health status of the equipment, resulting in inaccurate prediction results. Summary of the invention

[0003] The present invention provides a method and system for equipment health management of a wind turbine generator set, which is used to solve the problems mentioned in the above background technology:

[0004] The present invention provides a method for managing the health of a wind turbine generator set, the method comprising:

[0005] S1. Acquire multi-source data of wind turbines, pre-process the acquired multi-source data, and integrate the pre-processed multi-source data into a unified data format using data fusion technology;

[0006] S2. Extract key features that reflect the health status of the equipment based on domain knowledge and data characteristics, and select the feature subset with the most predictive value for the health status of the equipment through feature selection algorithm;

[0007] S3. Design an integrated learning framework, which includes multiple base learners, each of which is trained using a different feature subset, and combines the prediction results of multiple base learners through a voting mechanism to form a final prediction model;

[0008] S4. Based on real-time environmental data and equipment operating status, a dynamic risk assessment model is constructed. The health risk level of the equipment under different environmental conditions is evaluated in real time according to the dynamic risk assessment model. The parameters of the risk assessment model are continuously updated in combination with historical failure data and expert knowledge.

[0009] S5. Input the real-time monitoring data into the integrated learning model to obtain the prediction results of the equipment health status. According to the dynamic risk assessment results, the health threshold is set. When the prediction result is lower than the threshold, the early warning mechanism is triggered.

[0010] S6. Dynamically adjust the maintenance plan based on the health status monitoring and early warning results to achieve predictive maintenance, and continuously optimize the maintenance strategy based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals.

[0011] The present invention proposes an equipment health management system for a wind turbine generator set, the system comprising:

[0012] Data acquisition module: acquires multi-source data of wind turbines, pre-processes the acquired multi-source data, and integrates the pre-processed multi-source data into a unified data format using data fusion technology;

[0013] Feature extraction module: Based on domain knowledge and data characteristics, it extracts key features that reflect the health status of the equipment, and uses feature selection algorithms to select the feature subset with the most predictive value for the health status of the equipment;

[0014] Model formation module: Design an integrated learning framework, which includes multiple base learners, each of which is trained using a different feature subset, and combines the prediction results of multiple base learners through a voting mechanism to form a final prediction model;

[0015] Parameter update module: Based on real-time environmental data and equipment operating status, a dynamic risk assessment model is constructed. The health risk level of the equipment under different environmental conditions is evaluated in real time according to the dynamic risk assessment model. The parameters of the risk assessment model are continuously updated in combination with historical fault data and expert knowledge.

[0016] Mechanism trigger module: inputs real-time monitoring data into the integrated learning model to obtain the prediction results of the equipment health status, sets the health threshold according to the dynamic risk assessment results, and triggers the early warning mechanism when the prediction result is lower than the threshold;

[0017] Strategy optimization module: Dynamically adjust the maintenance plan according to the health status monitoring and early warning results to achieve predictive maintenance, and continuously optimize the maintenance strategy based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals.

[0018] The beneficial effects of the present invention are as follows: through real-time monitoring of key components of wind turbines by a high-precision sensor network, and combined with multi-source information such as historical fault databases and meteorological data, comprehensive data collection is achieved. The data is cleaned and denoised by using a rule-based cleaning method and an adaptive filtering algorithm, thereby improving data quality. The data of different scales are converted into a unified format by the Z-score standardization and normalization method, which is convenient for subsequent analysis. The high-level features are extracted by a deep learning model and the data dimensions are reduced by principal component analysis, which effectively reduces the complexity of data processing; based on physical models and statistical learning methods, key features reflecting the health status of the equipment are extracted, such as vibration signal spectrum features, temperature trend features, etc. Highly correlated features are identified by correlation analysis or mutual information methods, and feature subsets are further screened by feature selection algorithms such as random forests to ensure that the selected features are representative and can accurately predict the health status of the equipment; an integrated learning framework containing multiple base learners is constructed, each base learner is trained using a different feature subset, and the prediction results of each base learner are combined through a voting mechanism to form a final prediction model. This design improves the stability and generalization ability of the model and reduces the errors that may be caused by a single model. A dynamic risk assessment model is built based on real-time environmental data and equipment operating status to evaluate the health risk level of the equipment under different environmental conditions in real time. Fuzzy logic or evidence theory is introduced to deal with the uncertainty problem in risk assessment, and the model parameters are regularly updated in combination with historical fault data and expert knowledge base. When the predicted result is lower than the set health threshold, the early warning mechanism is automatically triggered, and the alarm information is sent to relevant personnel or systems to ensure that timely measures are taken to prevent the occurrence of failures; the maintenance plan is dynamically adjusted according to the health status monitoring and early warning results, and preventive maintenance measures are implemented. A maintenance strategy optimization model is built based on the reinforcement learning algorithm, and the optimal maintenance strategy is automatically found considering the maintenance cost and equipment reliability goals. The maintenance strategy is further optimized through the multi-objective optimization algorithm to meet the balanced solution of multiple goals. Maintenance feedback data is continuously collected for the training and optimization of the reinforcement learning model, and the maintenance strategy is continuously optimized to improve overall efficiency and reduce costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a step diagram of the method of the present invention;

[0020] Figure 2 This is a system module diagram of the present invention. DETAILED DESCRIPTION

[0021] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0022] One embodiment of the present invention, as Figure 1As shown, a device health management method for a wind turbine generator set, the method comprising:

[0023] S1. Acquire multi-source data of wind turbine generator sets, wherein the multi-source data includes operation data (such as vibration, temperature, power, etc.), historical fault records, environmental data (such as wind speed, wind direction, temperature, humidity, etc.) and maintenance logs, and pre-process the acquired multi-source data, wherein the pre-processing includes cleaning, denoising, standardization and normalization of the data; and use data fusion technology to integrate the pre-processed multi-source data into a unified data format;

[0024] S2. Extract key features that reflect the health status of the equipment based on domain knowledge and data characteristics, and select the feature subset with the most predictive value for the health status of the equipment through feature selection algorithm;

[0025] S3. Design an integrated learning framework, which includes multiple base learners (such as decision trees, support vector machines, neural networks, etc.), each base learner is trained using a different feature subset, and the prediction results of multiple base learners are combined through a voting mechanism to form a final prediction model;

[0026] S4. Based on real-time environmental data and equipment operating status, a dynamic risk assessment model is constructed. The health risk level of the equipment under different environmental conditions is evaluated in real time according to the dynamic risk assessment model. The parameters of the risk assessment model are continuously updated in combination with historical failure data and expert knowledge.

[0027] S5. Input the real-time monitoring data into the integrated learning model to obtain the prediction results of the equipment health status. According to the dynamic risk assessment results, the health threshold is set. When the prediction result is lower than the threshold, the early warning mechanism is triggered.

[0028] S6. Dynamically adjust the maintenance plan based on the health status monitoring and early warning results to achieve predictive maintenance, and continuously optimize the maintenance strategy based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals.

[0029] The working principle of the above technical solution is as follows: obtain data from multiple sources from wind turbines, including real-time operating data (such as vibration, temperature, power, etc.), historical fault records, environmental data (such as wind speed, wind direction, temperature, humidity, etc.) and maintenance logs; pre-process these multi-source data, including data cleaning (removing invalid or erroneous data), denoising (reducing random fluctuations in data), standardization (scaling data to the same scale) and normalization (converting data to values ​​within a specific range) to improve data quality and consistency; use data fusion technology to integrate the pre-processed multi-source data into a unified Data format is convenient for subsequent analysis and modeling; based on domain knowledge and data characteristics, key features that can reflect the health status of the equipment are identified; through feature selection algorithms, the feature subset with the most predictive value for the health status of the equipment is selected from a large number of features to reduce the complexity of the model and improve the prediction accuracy; an integrated learning framework is designed, which contains multiple base learners, such as decision trees, support vector machines, neural networks, etc.; each base learner is trained with different feature subsets to capture different information in the data; the prediction results of multiple base learners are combined through a voting mechanism to form the final prediction model. This integrated learning method can improve the stability and accuracy of the model; the model parameters are optimized using cross-validation and grid search techniques to further improve the performance of the model; a dynamic risk assessment model is constructed based on real-time environmental data and equipment operating status; the model can evaluate the health risk level of the equipment under different environmental conditions in real time; combined with historical fault data and expert knowledge, the parameters of the risk assessment model are continuously updated to adapt to changes in equipment status; real-time monitoring data is input into the integrated learning model to obtain the prediction results of the equipment health status; according to the dynamic risk assessment results, the health threshold is set. When the prediction result is lower than the threshold, the early warning mechanism is triggered to remind the operator to pay attention to the health status of the equipment; according to the health status monitoring and early warning results, the maintenance plan is dynamically adjusted to achieve predictive maintenance; based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals, the maintenance strategy is continuously optimized. The reinforcement learning algorithm can continuously adjust the strategy based on the feedback during the maintenance process to achieve the best maintenance effect.

[0030] The effects of the above technical solutions are as follows: by acquiring multi-source data and performing preprocessing, including cleaning, denoising, standardization and normalization, and integrating data using data fusion technology, a more comprehensive and accurate information basis is provided for the model; key features are extracted based on domain knowledge and data characteristics, and the feature subsets with the most predictive value are screened out through feature selection algorithms, further improving the prediction accuracy of the model; an integrated learning framework is designed, which includes multiple base learners, each of which is trained using a different feature subset, and the prediction results are combined through a voting mechanism to reduce the risk of a single model and improve the robustness and stability of the model; model parameters are optimized using cross-validation and grid search techniques to ensure that the model can perform well under different conditions; a dynamic risk assessment model is constructed that can evaluate the health risk level of equipment under different environmental conditions in real time, providing a scientific basis for early warning and maintenance decisions; historical fault data and expert knowledge are combined to ensure that the model can be used to predict the health risk level of equipment under different environmental conditions. The parameters of the risk assessment model are continuously updated so that the model can adapt to changes in equipment status and environmental conditions. According to the health status monitoring and early warning results, the maintenance plan is dynamically adjusted to achieve predictive maintenance, avoiding unnecessary downtime and maintenance costs. The maintenance strategy is optimized based on the reinforcement learning algorithm, combining the maintenance cost and equipment reliability goals to achieve the best balance between maintenance efficiency and equipment performance. Through real-time monitoring and early warning, abnormal conditions in the health status of equipment are discovered in time, and corresponding maintenance measures are taken, which effectively reduces the failure rate and extends the service life of the equipment. The entire health management method integrates multiple links such as data preprocessing, feature extraction, integrated learning, dynamic risk assessment, early warning mechanism and maintenance strategy optimization, realizing the intelligent and automated health management of wind turbines. The application of data-driven and machine learning algorithms is strengthened, which improves the intelligence level of wind turbine health management and provides strong support for the digital transformation of the wind power industry.

[0031] In one embodiment of the present invention, the S1 includes:

[0032] S11. Based on a high-precision sensor network, key components of the wind turbine generator set are covered, including gearboxes, generators, bearings and blades, and key parameters of key components are monitored in real time; the key parameters include vibration, temperature, oil pressure, current and power;

[0033] S12. Establish a historical fault database to record in detail the type, time of occurrence, cause, repair measures and cost of each fault, as well as the changes in operating data before and after the fault;

[0034] S13, integrating meteorological data sources, including local meteorological station data, satellite cloud images and numerical weather forecasts, and obtaining external environmental parameters, including wind speed, wind direction, temperature, humidity and air pressure; systematizing maintenance logs, recording relevant information of each maintenance activity, including time, personnel, operation content, replaced parts and maintenance costs, and associating them with specific wind turbine generator sets;

[0035] S14, using a rule-based cleaning method (such as setting a threshold to filter outliers) to clean the data, and applying an adaptive filtering algorithm (such as Kalman filtering, minimum mean square error filtering) to denoise the data;

[0036] S15. According to the distribution characteristics of the data, the data are converted to the same scale through Z-score standardization, and for specific features or feature combinations, normalization methods (such as logarithmic transformation, Box-Cox transformation) are used to map the data to a specific range;

[0037] S16. Use deep learning models (such as autoencoders and convolutional neural networks) to fuse data and extract high-level features. At the same time, apply principal component analysis (PCA) to reduce data dimensions.

[0038] The working principle of the above technical solution is: through a high-precision sensor network, the key components of the wind turbine (such as gearboxes, generators, bearings, blades) are monitored in real time. These sensors can capture real-time data of key parameters (such as vibration, temperature, oil pressure, current, power), which are the basis for subsequent analysis and prediction; establish a historical fault database to record in detail the type, time of occurrence, cause, maintenance measures and cost of each fault, as well as the changes in operating data before and after the fault. This step helps to understand the fault mode, identify the precursors of faults, and provide valuable historical data for subsequent data analysis and model training; integrate multiple meteorological data sources (such as local weather station data, satellite cloud maps, numerical weather forecasts) to obtain external environmental parameters (wind speed, wind direction, temperature, humidity, air pressure). At the same time, the maintenance log is systematically processed to record the detailed information of each maintenance activity and associate it with a specific wind turbine. This information helps to analyze the impact of environmental factors on the equipment status and the changes in equipment performance caused by maintenance activities; rule-based cleaning methods (such as setting thresholds to filter outliers) are used to identify and remove outliers or invalid data in the data. At the same time, adaptive filtering algorithms (such as Kalman filtering and minimum mean square error filtering) are used to denoise the data to reduce random fluctuations and noise in the data and improve data quality; according to the distribution characteristics of the data, the data is converted to the same scale through the Z-score standardization method for subsequent analysis. For specific features or feature combinations, normalization methods (such as logarithmic transformation and Box-Cox transformation) are used to map the data to a specific range to eliminate the dimensional differences between different features and improve the convergence speed and prediction accuracy of the model; the data is fused through deep learning models (such as autoencoders and convolutional neural networks) to extract high-level features. These high-level features can more accurately reflect the relationship between equipment status and environmental factors. At the same time, the principal component analysis (PCA) method is used to reduce data dimensions, remove redundant information, and improve the generalization ability of the model.

[0039] The effects of the above technical solutions are as follows: real-time monitoring of key components of wind turbines through a high-precision sensor network can promptly detect small changes in equipment status and provide accurate data support for early warning and fault prediction; real-time monitoring of key parameters (such as vibration, temperature, oil pressure, current, power) helps capture early signals of equipment failure and reduce the probability of failure; establishing a historical fault database to record in detail the type, time of occurrence, cause, maintenance measures and cost of each failure, which helps to analyze failure modes, identify precursors to failures, and provide valuable historical data for subsequent equipment maintenance and health management; by analyzing the changes in operating data before and after the failure, the development process of equipment failure can be further understood and the accuracy of fault prediction can be improved; integrating meteorological data sources to obtain external environmental parameters (wind speed, wind direction, temperature, humidity, air pressure) helps to analyze the impact of environmental factors on equipment status and improve the adaptability and stability of equipment in different environments; through the systematic recording of maintenance logs, external environmental parameters can be associated with maintenance activities to formulate more scientific maintenance plans. The data is cleaned and denoised using rule-based cleaning methods and adaptive filtering algorithms, which can effectively remove outliers and noise from the data and improve the quality and reliability of the data. Data standardization and normalization can eliminate the dimensional differences between different features and provide a consistent data basis for subsequent data analysis and model training. Data fusion and high-level feature extraction through deep learning models can more accurately reflect the relationship between equipment status and environmental factors and improve the prediction performance of the model. The application of principal component analysis (PCA) reduces data dimensions and removes redundant information, which helps to reduce the complexity of the model and improve the operating efficiency and generalization ability of the model. The entire technical solution integrates real-time monitoring, historical fault analysis, environmental adaptability, data quality improvement, feature extraction and dimensionality reduction, and other links to realize the intelligent and automated health management of wind turbines. The application of these technologies helps to improve the accuracy and efficiency of equipment health management, reduce operation and maintenance costs, extend the service life of equipment, and provide strong support for the sustainable development of the wind power industry.

[0040] In one embodiment of the present invention, the S16 includes:

[0041] Design multiple specific autoencoder models for monitoring data of key components of wind turbines (such as vibration, temperature, oil pressure, etc.);

[0042] Each autoencoder is responsible for learning and reconstructing features from a single type of sensor data to capture the inherent structure and regularity of its own dataset;

[0043] The hidden layer outputs of these autoencoders are used as feature vectors and input into an integrated autoencoder. This integrated model achieves deep fusion of cross-type data by jointly learning the representation of all feature vectors and extracting more comprehensive and abstract high-level features.

[0044] Convolutional neural network (CNN) is used to extract features from images or time series data such as blade surface defect detection or gearbox internal structure analysis. Through multi-layer convolution and pooling operations, CNN can automatically learn low-level features such as edges, textures, and shapes in images, as well as higher-level spatial structures and time-series dependency features.

[0045] Ensemble learning methods such as random forest and gradient boosting tree are used to evaluate the importance of each fused feature in predicting wind turbine faults. The feature subset that contributes most to fault prediction is selected by calculating the feature importance score.

[0046] The interaction between features is analyzed using a method based on SHAP value, and the fused feature data is standardized.

[0047] Through PCA analysis, the high-dimensional feature space is projected into a low-dimensional space composed of principal components, and the minimum number of principal components whose cumulative contribution rate reaches a preset threshold (such as 95%) is selected.

[0048] The working principle of the above technical solution is as follows: for the monitoring data (vibration, temperature, oil pressure, etc.) of the key components of the wind turbine generator set (such as gearbox, generator, bearing, blade), multiple specific autoencoder models are designed. Each autoencoder is responsible for processing a single type of sensor data, reconstructing the input data through unsupervised learning, thereby capturing the inherent structure and regularity of the respective data set; the hidden layer output of the autoencoder is used as a feature vector, which contains the low-dimensional representation learned from the original data. These feature vectors can more effectively reflect the essential characteristics of the data and provide a basis for subsequent data fusion and fault prediction; the hidden layer outputs of all specific autoencoders are used as input to design an integrated autoencoder. The integrated model realizes deep fusion of cross-type data by jointly learning the representation of all feature vectors. This step can extract more comprehensive and abstract high-level features, which can better reflect the complex relationship between equipment status and environmental factors; for image or time series data such as blade surface defect detection or gearbox internal structure analysis, convolutional neural network (CNN) is used for feature extraction. Through multi-layer convolution and pooling operations, CNN automatically learns low-level features such as edges, textures, and shapes in images, as well as higher-level spatial structures and temporal dependency features. These features are of great significance for identifying equipment failures and abnormal states; Ensemble learning methods such as random forests and gradient boosting trees are used to evaluate the importance of each fused feature in predicting wind turbine failures. By calculating the importance score of the features, the feature subset that contributes most to fault prediction is screened out. This step helps to reduce redundant features and improve the prediction performance and generalization ability of the model; The interaction between features is analyzed using a method based on SHAP (Shapley Additive Ex Planations) values. SHAP values ​​can explain the contribution of each feature to the model prediction results, thereby revealing the complex relationship between features. This step helps to understand the importance of features and their interaction mechanisms; The fused feature data is standardized to ensure that different features are numerically comparable. This step helps to improve the convergence speed and prediction accuracy of the model; Through principal component analysis (PCA), the high-dimensional feature space is projected into a low-dimensional space composed of principal components. PCA can remove redundant information between features and retain the most representative feature components; it selects the minimum number of principal components whose cumulative contribution rate reaches a preset threshold (such as 95%). This step helps to reduce the complexity of the model and improve its operating efficiency and generalization ability.

[0049] The effects of the above technical solution are as follows: by designing multiple specific autoencoder models, each autoencoder can learn and reconstruct features from a single type of sensor data to capture the inherent structure and regularity of each data set; the integrated autoencoder realizes the deep fusion of cross-type data and extracts more comprehensive and abstract high-level features, which helps to more accurately reflect the operating status of the wind turbine generator set; a convolutional neural network (CNN) is used to extract features from image or time series data such as blade surface defect detection or gearbox internal structure analysis; CNN can automatically learn low-level features such as edges, textures and shapes in images, as well as higher-level spatial structures and time series dependency features, to improve the accuracy of fault detection; ensemble learning methods such as random forests and gradient boosting trees are used to evaluate the importance of each fused feature for predicting wind turbine faults; by calculating the importance score of the feature, the feature subset that contributes most to fault prediction is screened out, which helps to simplify the model and improve prediction efficiency; the interaction between features is analyzed using a method based on SHAP values, which helps to deeply understand Understand the mechanism of fault occurrence; standardize the fused feature data to improve the stability and reliability of the model; project the high-dimensional feature space into a low-dimensional space composed of principal components through PCA analysis; select the minimum number of principal components whose cumulative contribution rate reaches the preset threshold (such as 95%) to reduce the complexity of the model and improve the computational efficiency; ensemble learning methods such as random forests can improve the generalization ability and robustness of the model by integrating multiple weak learners; help to reduce the dependence on a single model and the computational cost while maintaining the model performance; this technical solution can realize real-time monitoring and early warning of wind turbines, timely detect potential faults and take corresponding maintenance measures; help to reduce downtime and maintenance costs caused by faults, and improve the operating efficiency and reliability of wind turbines; by combining deep learning, ensemble learning and PCA dimensionality reduction and other technical means, this technical solution provides strong support for the intelligent operation and maintenance of wind turbines; help to realize the automation and intelligence level of operation and maintenance work, reduce operation and maintenance costs and improve operation and maintenance efficiency.

[0050] In one embodiment of the present invention, the S2 includes:

[0051] S21. Extracting first related features based on the physical model, where the first related features include spectrum features of the vibration signal (such as peak frequency and sideband width) and trend features of the temperature signal (such as heating rate and temperature fluctuation range);

[0052] S22. Extracting second related features by a statistical learning method, where the second related features include statistical features of the data (such as mean, variance, skewness, kurtosis) and time series features (such as autocorrelation coefficient and partial autocorrelation coefficient);

[0053] S23. Use deep learning technology to automatically extract high-dimensional features through autoencoders or convolutional neural networks to capture complex patterns in data;

[0054] S24. Use correlation analysis (such as Pearson correlation coefficient, Spearman rank correlation coefficient) or mutual information method to identify features that are highly correlated with the health status of the equipment; based on the model, use random forest to further screen feature subsets according to the importance score of the features.

[0055] The working principle of the above technical solution is: to extract features directly related to the equipment status through physical models. For example, for vibration signals, signal processing technology can be used to extract spectral features such as peak frequency and sideband width, which can reflect the vibration status and potential failure modes of the equipment. At the same time, for temperature signals, trend features such as heating rate and temperature fluctuation range can be extracted to evaluate the thermal state and stability of the equipment; statistical principles are used to extract the intrinsic characteristics of the data. By calculating the statistical characteristics of the data such as mean, variance, skewness, kurtosis, etc., the distribution and discreteness of the data can be understood. In addition, for time series data, features such as autocorrelation coefficient and partial autocorrelation coefficient can be extracted to capture the law and trend of data changes over time. These features help to identify the changing trends and anomalies of the equipment status; use the powerful data processing capabilities of deep learning models (such as autoencoders or convolutional neural networks) to automatically extract useful features from high-dimensional data. Deep learning models can capture complex patterns and associations in data by learning the intrinsic representation and hierarchical structure of data. These automatically extracted features are usually more expressive and generalizable than manually designed features, which helps to improve the accuracy of equipment health status monitoring and fault diagnosis; correlation analysis methods (such as Pearson correlation coefficient and Spearman rank correlation coefficient) are used to evaluate the degree of association between features and equipment health status. By calculating the correlation coefficient between features and target variables (such as equipment fault status), features that are highly correlated with equipment health status can be identified; based on the correlation analysis, machine learning models such as random forests are used to further screen feature subsets. The random forest model can sort features according to their importance scores, thereby selecting the features that contribute most to equipment health status monitoring and fault diagnosis. This process helps to reduce redundant features and improve the efficiency and accuracy of the model.

[0056] The effects of the above technical solution are as follows: by deeply analyzing the working principle and physical characteristics of the equipment, the first relevant features directly related to the health status of the equipment can be accurately extracted, such as the spectral characteristics of the vibration signal and the trend characteristics of the temperature signal; these features have clear physical meanings and can directly reflect the operating status and potential failure modes of the equipment; using the principles and methods of statistics, statistical features and time series features are extracted from the data, such as mean, variance, skewness, kurtosis, autocorrelation coefficient, partial autocorrelation coefficient, etc.; these features can reveal the inherent laws and trends of the data and provide strong support for subsequent fault diagnosis and health management; through deep learning models such as autoencoders or convolutional neural networks, useful features can be automatically extracted from high-dimensional data to capture complex patterns and associations in the data; deep learning models have strong data processing and generalization capabilities, and can extract features that are more expressive than manually designed features; using correlation analysis (such as Pearson correlation coefficient, Spearman rank correlation coefficient) or mutual information method, it is possible to accurately identify features that are highly related to the health status of the equipment. features that are highly correlated with the target variable; these methods can quantify the degree of association between features and target variables, providing a scientific basis for subsequent feature selection; using machine learning models such as random forests, feature subsets can be further screened according to the importance scores of features; the random forest model can evaluate the contribution of each feature to the model's predictive performance, thereby selecting the features that contribute most to equipment health status monitoring and fault diagnosis; this process helps to reduce redundant features and improve the efficiency and accuracy of the model; through comprehensive and accurate feature extraction and efficient and accurate feature selection, a more accurate fault diagnosis model can be constructed; these models can timely discover the potential failure modes of equipment and improve the accuracy and reliability of fault diagnosis; based on the extracted and selected features, a more scientific and reasonable health management strategy can be formulated; for example, targeted maintenance plans and preventive measures can be formulated according to the operating status and potential failure modes of the equipment to extend the service life of the equipment and improve operating efficiency; through accurate fault diagnosis and optimized health management strategies, unnecessary downtime and maintenance costs can be reduced.

[0057] In one embodiment of the present invention, the S24 includes:

[0058] S241, constructing a multi-dimensional correlation matrix based on all extracted features (including the features obtained in steps S21, S22, and S23);

[0059] S242. Convert the multi-dimensional correlation matrix into a feature network, where nodes represent features and edges represent the strength of correlation between features; use graph theory algorithms (such as PageRank and HITS algorithms) to analyze the centrality and influence of nodes in the network and identify features that play a key role in the network;

[0060] S243. Based on the historical health status records of the equipment (such as data before and after the failure), use sensitivity analysis techniques (such as Sobol's sensitivity index) to evaluate the sensitivity of each feature to changes in the health status of the equipment;

[0061] S244. Recalculate and analyze the feature correlation within each time window through dynamic time window technology. Based on the analysis results, comprehensively evaluate the correlation, centrality, sensitivity and time dynamics of each feature to select the most representative feature subset that is highly correlated with the health status of the equipment.

[0062] S245. Use feature importance scores (such as the Gini index or information gain in random forests) to rank features and further optimize feature subsets.

[0063] The working principle of the above technical solution is as follows: based on all the features extracted in steps S21, S22, and S23, a multidimensional correlation matrix containing multiple statistics is constructed. This matrix not only contains the traditional Pearson correlation coefficient and Spearman rank correlation coefficient, but also introduces advanced statistics such as mutual information and maximum information coefficient (MIC); the multidimensional correlation matrix is ​​converted into a feature network, in which nodes represent features and edges represent the strength of correlation between features. Then, graph theory algorithms (such as PageRank and HITS algorithms) are used to analyze the centrality and influence of nodes in the network; for the historical health status records of the equipment, sensitivity analysis techniques (such as Sobol' sensitivity index) are used to evaluate the sensitivity of each feature to changes in the health status of the equipment. By simulating small changes in the status of the equipment, the change range of the feature value is observed; the dynamic time window technology is used to recalculate and analyze the feature correlation within each time window. Based on the analysis results, the correlation, centrality, sensitivity and time dynamics of each feature are comprehensively evaluated; the selected feature subsets are sorted using feature importance scoring methods (such as the Gini index or information gain in random forests). For example, suppose that the health status of a wind turbine is being monitored. Through steps S21, S22, and S23, multiple features including vibration signal spectrum features, temperature signal trend features, data statistical features, and high-dimensional features automatically extracted by deep learning were extracted. In step S24, a multi-dimensional correlation matrix was first constructed, and it was found that there was a strong nonlinear correlation between some vibration signal spectrum features and temperature signal trend features. Then, the correlation matrix was converted into a feature network, and the PageRank algorithm was used to identify features with high centrality in the network. These features are often closely related to the failure mode of wind turbines. Then, the Sobol' sensitivity index was used to perform sensitivity analysis on the features, and it was found that some features were highly sensitive to changes in the health status of wind turbines. Through dynamic time window analysis, feature subsets that showed high correlation and sensitivity in different time periods were further screened out; finally, the Gini index in random forest was used to sort the feature subsets to ensure that the final selected features can not only fully reflect the health status of wind turbines, but also have high prediction accuracy.

[0064] The effects of the above technical solutions are as follows: by integrating advanced statistics such as Pearson correlation coefficient, Spearman rank correlation coefficient, mutual information, maximum information coefficient (MIC), etc., the S24 technical solution can comprehensively measure the linear and nonlinear correlations between features; this comprehensive correlation analysis helps to reveal the complex relationship between features and provide a scientific basis for subsequent feature selection; by converting the multi-dimensional correlation matrix into a feature network, and using graph theory algorithms to analyze the centrality and influence of nodes in the network, it is possible to identify features that play a key role in the network; these key features are often closely related to the health status of the equipment and are of great significance for fault diagnosis and health monitoring; by adopting sensitivity analysis techniques (such as Sobol's sensitivity index), the S24 technical solution can evaluate the sensitivity of each feature to changes in the health status of the equipment; this analysis helps to screen out the features that are most sensitive to changes in the health status of the equipment, thereby improving the accuracy of feature selection; using dynamic time window technology, the S24 technical solution can recalculate and analyze the feature correlations within each time window; this dynamic analysis helps to capture the characteristics of the equipment. The S24 technical solution can further screen out the feature subsets that are highly correlated with the health status of the equipment and are the most representative by analyzing the changing trends of the features over time; by using the feature importance score (such as the Gini index or information gain in the random forest), the S24 technical solution can sort the screened feature subsets; this sorting helps to further optimize the feature subsets and select the features that contribute most to the model prediction performance, thereby improving the accuracy and generalization ability of the model; the optimized feature subsets can more accurately reflect the health status of the equipment, help to detect potential faults in advance and take corresponding maintenance measures; this can not only reduce the losses caused by equipment failures, but also improve the reliability and service life of the equipment; the S24 technical solution does not rely on specific equipment or scenarios, and can be widely used in fault diagnosis and health monitoring of various industrial equipment; this adaptability and flexibility make the S24 technical solution have broad application prospects and market demand; the S24 technical solution can be seamlessly integrated with existing fault diagnosis and health monitoring systems, making it convenient for users to upgrade and expand; this integration and expansion capability helps to reduce users' implementation costs and time costs.

[0065] In one embodiment of the present invention, the S242 includes:

[0066] Each element in the multidimensional correlation matrix is ​​regarded as a correlation coefficient between features, and a feature network is constructed using the correlation coefficient, in which each node represents a feature, and the edges between nodes are drawn according to the absolute value of the correlation coefficient, and the weight of the edge is the absolute value of the correlation coefficient;

[0067] After the feature network is constructed, the centrality and influence of the nodes in the network are analyzed through the PageRank algorithm and HITS algorithm based on graph theory; the PageRank algorithm simulates the random walk process in the network and calculates the probability of each node being visited, thereby evaluating the importance of the node. The HITS algorithm iteratively calculates the Hub value and Authority value of the node to identify the hub nodes and authoritative nodes in the network. The combination of these two algorithms can more comprehensively reveal the key nodes in the feature network, that is, those features that play an important role in the network.

[0068] By calculating graph theory indicators such as path length and clustering coefficient between features, local and global connection patterns between features are obtained, and the feature network is visualized using visualization tools;

[0069] After identifying the key features, formulate a feature screening strategy based on the network characteristics of the features (such as centrality, influence, connectivity, etc.); including:

[0070] According to the results of PageRank and HITS algorithms, the features that occupy an important position in the network are screened out;

[0071] Combined with information such as correlation coefficient and path length between features, we can further screen out those features that are closely related to key features and play a bridging role in the network;

[0072] By comprehensively evaluating the network characteristics and actual application requirements of the features, a concise and effective feature subset is determined.

[0073] The working principle of the above technical solution is: collect and process feature data from different data sources, calculate the correlation coefficients between them (such as Pearson correlation coefficient, Spearman rank correlation coefficient, etc.), and form a multi-dimensional correlation matrix; each element in the matrix represents the strength of association between two features, and its absolute value reflects the closeness of the relationship between the features; use the correlation matrix to construct a feature network, in which each node corresponds to a feature, and the edges between nodes are drawn according to the absolute value of the correlation coefficient, and the weight of the edge is the absolute value of the correlation coefficient; the network constructed in this way can intuitively show the association between features and provide a basis for subsequent graph analysis; the PageRank algorithm is applied to simulate the random walk process in the network, and the probability of each node being visited is calculated to evaluate the importance of the node. The higher the PageRank value of the node, the greater its influence in the network; at the same time, the HITS algorithm is used to iteratively calculate the Hub value and Authority value of the node to identify the hub nodes and authoritative nodes in the network. Hub nodes point to multiple authoritative nodes, while authoritative nodes are pointed to by multiple Hub nodes; combining the results of PageRank and HITS algorithms can more comprehensively reveal the key nodes in the feature network; calculating graph theory indicators such as path length and clustering coefficient between features to obtain local and global connection patterns between features; using visualization tools (such as Gephi, NetworkX, etc.) to visualize the feature network, intuitively presenting the correlation between features and the topological structure of the network; based on the results of PageRank and HITS algorithms, screen out features that occupy an important position in the network; combining information such as correlation coefficients and path lengths between features, further screen out features that are closely related to key features and act as bridges in the network; by comprehensively evaluating the network characteristics of features and actual application requirements (such as feature interpretability, computational cost, contribution to the prediction model, etc.), determine a concise and effective feature subset.

[0074] The effect of the above technical solution is: by constructing a feature network through a multi-dimensional correlation matrix, and using PageRank and HITS algorithms to analyze the centrality and influence of nodes in the network, the key features that occupy an important position in the network can be accurately identified. These features are usually highly correlated with the health status of the equipment and have a high predictive value; compared with traditional feature selection methods, this solution quickly screens out key features through graph theory algorithms, greatly reducing the computational complexity and time cost of feature selection; by displaying the feature network through visualization tools, you can intuitively see the correlation between features and the topological structure of the network. This helps users understand the working principle of the model and improves the interpretability of the model; the selected feature subset is both concise and effective, which can reduce the sensitivity of the model to noisy data and improve the robustness and generalization ability of the model; the streamlined feature subset means that in the subsequent data processing and model training process, the amount of data to be processed is greatly reduced, thereby optimizing the utilization of computing resources; reducing the number of features can reduce the complexity and computing cost of model training, especially when dealing with large-scale data sets, this optimization is particularly important; the identification of key features can provide strong support for equipment fault warning and prediction, helping users to detect potential problems in a timely manner and take corresponding measures; by continuously monitoring the changes in key features, real-time evaluation and management of equipment health status can be achieved, improving the reliability and service life of equipment; the solution combines knowledge from multiple fields such as mathematical statistics, graph analysis, visualization technology and machine learning, and promotes cross-disciplinary integration and innovation.

[0075] In one embodiment of the present invention, S3 includes:

[0076] S31. Based on data characteristics and problem complexity, select various types of base learners, such as decision trees, random forests, gradient boosting machines, support vector machines, neural networks (such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs). Design specific input formats and preprocessing steps for base learners based on different types of features (such as time series features, statistical features, and high-dimensional features).

[0077] S32, using integration strategies (such as bagging, boosting, etc.), combining the prediction results of multiple base learners to improve the stability and generalization ability of the model;

[0078] S33, using cross-validation techniques to evaluate the performance of the integrated model, and automatically adjusting model parameters through grid search to find the optimal configuration;

[0079] S34. Based on a diversity enhancement mechanism, the diversity enhancement mechanism includes feature subset perturbation and model structure perturbation to increase the differences between base learners.

[0080] The working principle of the above technical solution is: select multiple types of base learners according to the data characteristics and problem complexity. For example, for time series data, you can choose recurrent neural network (RNN) or long short-term memory network (LSTM); for statistical features and high-dimensional features, you can choose decision trees, random forests or support vector machines, etc.; for different types of features, design specific input formats and preprocessing steps for base learners. For example, for time series data, smoothing, detrending or seasonal decomposition may be required; for statistical features, standardization or normalization may be required; for high-dimensional features, feature selection or dimensionality reduction may be required.Suppose there is a set of data sets containing time series features, statistical features, and high-dimensional features for predicting the remaining useful life (RUL) of equipment; for time series features, we can choose LSTM as the base learner and design the input format to be fixed-length sequence data, while performing smoothing and seasonal decomposition to eliminate noise and periodic effects; for statistical features, we can choose random forest as the base learner and perform standardization to ensure that the feature values ​​are on the same scale; for high-dimensional features, we can first use principal component analysis (PCA) for dimensionality reduction, and then choose support vector machine (SVM) as the base learner; using integration strategies (such as Bagging, Boost ing, etc.), combining the prediction results of multiple base learners to improve the stability and generalization ability of the model; the Bagging strategy randomly extracts multiple subsets from the original data set to train different base learners, and takes the average or majority vote of the prediction results of these base learners as the final prediction; the Boosting strategy gradually adjusts the sample weights and base learner weights so that subsequent base learners can focus on samples that the previous base learners failed to classify correctly, thereby improving the performance of the overall model. For example, in the above-mentioned device RUL prediction example, the Bagging strategy can be used to combine the prediction results of multiple random forest base learners; or, the Boosting strategy can be used , first use a simple decision tree as the initial base learner, then adjust the sample weight according to the prediction error, train the next base learner, and repeat this process until the preset number of base learners is reached or the prediction error converges; use cross-validation techniques (such as k-fold cross-validation) to evaluate the performance of the ensemble model; automatically adjust model parameters (such as the number, type, depth, learning rate, etc. of base learners) through grid search to find the optimal configuration; cross-validation can ensure the stability and generalization ability of the model on different data sets, while grid search can systematically explore the parameter space and find the best parameter combination; for example, in the above example of device RUL prediction, 5-fold cross-validation can be used to evaluate the ensemble performance of the model; then, grid search can be used to adjust parameters such as the number of random forest base learners, maximum depth, minimum number of samples, etc. to find the optimal configuration; based on diversity enhancement mechanisms (such as feature subset perturbation, model structure perturbation, etc.), the differences between base learners are increased; feature subset perturbation introduces diversity by using different feature subsets on different base learners; model structure perturbation introduces diversity by changing the structure of the base learners (such as changing the depth of the decision tree, changing the number of layers and neurons of the neural network, etc.); increasing the differences between base learners helps to improve the stability and generalization ability of the integrated model, because different base learners may capture the characteristics of the data in different aspects.

[0081] The effects of the above technical solutions are: flexible selection of various types of base learners, such as decision trees, random forests, gradient boosting machines, support vector machines and neural networks, according to data characteristics and problem complexity; this diversity enables S3 to cope with different types of data and problems, improving the adaptability and flexibility of the model; specific input formats and preprocessing steps are designed for different types of features (such as time series features, statistical features, high-dimensional features); this targeted preprocessing helps to extract effective information from features and improve model performance. ; Adopting integration strategies (such as Bagging, Boosting, etc.) and combining the prediction results of multiple base learners can improve the stability and generalization ability of the model; the integration strategy can reduce the overfitting or underfitting problems that may exist in a single base learner and improve the prediction accuracy of the overall model; using cross-validation technology to evaluate the performance of the integrated model to ensure the stability and reliability of the model on different data sets; automatically adjusting model parameters through grid search can find the optimal configuration and further improve the performance of the model; based on the diversity enhancement mechanism (including feature subset perturbation and model structure perturbation), the differences between base learners are increased; this diversity helps to improve the robustness of the integrated model, because different base learners may capture the characteristics of the data in different aspects, thereby jointly improving the performance of the overall model; by increasing the base learners The above can reduce the model's dependence on specific data and improve the model's generalization ability; this means that the model can maintain good prediction performance when facing new data or unknown data; by integrating multiple base learners and adopting cross-validation technology, the risk of model overfitting is effectively reduced; this makes S3 more reliable in practical applications and can predict unknown data more accurately; although S3 adopts an integration strategy, the model's interpretability can be improved by analyzing the prediction results and contributions of each base learner; this helps users understand the model's decision-making process and adjust and optimize it according to actual needs; the S3 technical solution has good scalability and integration, and can be combined with other machine learning algorithms or models to form a more complex hybrid model; this enables S3 to cope with more complex problems and challenges and meet the diverse needs in practical applications.

[0082] In one embodiment of the present invention, the S4 includes:

[0083] S41. Based on real-time environmental data and equipment operating status, a risk assessment index system is constructed, wherein the risk assessment index system includes environmental adaptability indicators (such as wind speed adaptability and temperature adaptability), operating efficiency indicators (such as power loss rate and energy efficiency ratio), and failure probability indicators (such as failure prediction probability based on historical data);

[0084] S42. Use machine learning algorithms (such as Bayesian networks, Markov chains, and random forests) to build a dynamic risk assessment model, and evaluate the health risk level of the equipment under different environmental conditions in real time based on the constructed dynamic risk assessment model;

[0085] S43. Introduce fuzzy logic or evidence theory to deal with uncertainty issues in risk assessment; combine historical failure data and expert knowledge base to regularly update the parameters and rules of the risk assessment model;

[0086] S44. Through the online learning mechanism, the model is updated online using newly collected real-time data, and the model is adapted and continuously optimized.

[0087] S45. Based on the model performance monitoring and feedback mechanism, the prediction accuracy and stability of the model are evaluated in real time, and model degradation problems are discovered and handled.

[0088] The working principle of the above technical solution is as follows: based on real-time environmental data and equipment operating status, a risk assessment indicator system including environmental adaptability indicators (such as wind speed adaptability, temperature adaptability), operating efficiency indicators (such as power loss rate, energy efficiency ratio) and failure probability indicators (such as failure prediction probability based on historical data) is constructed; by collecting and analyzing real-time data, key indicators related to equipment health risks are extracted to provide a basis for subsequent risk assessment; machine learning algorithms are used to learn and analyze historical data and real-time data, and patterns and laws in the data are extracted to achieve real-time assessment of equipment health risks. The model can dynamically adjust the assessment results according to changes in different environmental conditions; through methods such as fuzzy logic or evidence theory, the uncertainty in risk assessment is quantified to improve the accuracy of the assessment. At the same time, combined with historical data and expert knowledge, the parameters and rules of the model are continuously updated to ensure that the model can accurately reflect changes in equipment status and environmental conditions; the online learning mechanism can capture user behavior and data changes in real time, thereby achieving timely updates of the model. This helps the model adapt to new data features and patterns, improve the accuracy and timeliness of the assessment; and timely discover model degradation problems by monitoring indicators such as the prediction accuracy and stability of the model. Once a problem is found, appropriate measures can be taken to repair and optimize it to ensure the continued effectiveness and accuracy of the model; by integrating a variety of advanced technical means, real-time and accurate assessment of equipment health risks is achieved. At the same time, through online learning and continuous optimization mechanisms, the accuracy and timeliness of the model are ensured. This helps companies and institutions better predict and manage risks, thereby improving the reliability and safety of equipment.

[0089] The effects of the above technical solutions are as follows: by integrating environmental adaptability indicators (such as wind speed adaptability, temperature adaptability), operating efficiency indicators (such as power loss rate, energy efficiency ratio) and failure probability indicators (such as failure prediction probability based on historical data), the S4 technical solution can comprehensively and accurately reflect the health status and risk level of the equipment; the selection and construction of these indicators are based on real-time environmental data and equipment operating status, ensuring the real-time and accuracy of risk assessment; advanced machine learning algorithms such as Bayesian networks, Markov chains, and random forests are used to build dynamic risk assessment models; these algorithms can learn and extract useful features and laws from large amounts of data, thereby accurately assessing the health risks of equipment; by updating model parameters and rules in real time, the model can continuously adapt to changes in equipment status and environmental conditions, improving the accuracy and real-time nature of risk assessment; fuzzy logic or evidence theory is introduced to deal with uncertainty issues in risk assessment; these methods can quantify uncertainty factors and incorporate them into risk assessment The model is then updated online to improve the robustness and reliability of the model; the parameters and rules of the risk assessment model are updated regularly in combination with historical failure data and expert knowledge base; this helps ensure that the model can accurately reflect changes in equipment status and environmental conditions, and improve the accuracy and reliability of risk assessment; the model is updated and optimized online using newly collected real-time data through an online learning mechanism; this enables the model to adapt to changing equipment status and environmental conditions, and improve the real-time and accuracy of risk assessment; based on the model performance monitoring and feedback mechanism, the prediction accuracy and stability of the model are evaluated in real time; by discovering and handling model degradation problems, the continued effectiveness and reliability of the model can be ensured; the risk assessment results provided by the S4 technical solution can provide auxiliary decision support for equipment management and maintenance; by identifying high-risk equipment and potential failure points, it helps to take preventive and maintenance measures in advance, reducing equipment failure rates and repair costs; based on the risk assessment results, resource allocation can be optimized and limited resources can be used where they are most needed.

[0090] In one embodiment of the present invention, the S43 includes:

[0091] Classify the uncertainty factors in real-time environmental data and equipment operating status, including random uncertainty (such as random fluctuations in wind speed and temperature), cognitive uncertainty (such as measurement errors caused by sensor accuracy limitations), and fuzzy uncertainty (such as the fuzziness of equipment performance degradation limits); and quantify the degree of each type of uncertainty through statistical methods;

[0092] Based on fuzzy sets and membership functions, fuzzy processing is performed on environmental adaptability indicators (such as "wind speed adaptability" is subdivided into fuzzy sets such as "low wind speed sensitivity", "medium wind speed adaptation", and "high wind speed tolerance") and operating efficiency indicators (such as "power loss rate" is divided into "minor loss", "moderate loss", "serious loss", etc.); through the fuzzy rule base, combined with expert experience and historical data, fuzzy reasoning rules for equipment health status under different conditions are defined; for example, "if the wind speed is in the high wind speed sensitive area and the temperature exceeds the appropriate range, the probability of equipment failure increases";

[0093] Apply the Dempster-Shafer Evidence Theory (DST) to integrate information from different sources (e.g., multiple sensors, historical data, expert judgment) to form a belief distribution about the health status of the equipment;

[0094] Combine historical fault data and expert knowledge base to regularly update membership functions, rule bases in fuzzy logic systems, and basic belief assignment (BBA) in evidence theory;

[0095] Based on the introduction of incremental learning methods, the basic trust assignment of fuzzy logic rules and evidence theory can be fine-tuned based on newly collected real-time data instead of completely retraining.

[0096] The working principle of the above technical solution is: uncertainty factors are classified into: random uncertainty: this type of uncertainty comes from random fluctuations in natural phenomena or equipment status, such as random changes in wind speed and temperature;

[0097] Epistemic uncertainty: The uncertainty in the collected data due to the accuracy limitations of the measurement equipment or sensors is called epistemic uncertainty.

[0098] Fuzzy uncertainty: The fuzziness of the equipment performance degradation boundary and the difficulty in defining certain indicators (such as wind speed adaptability and power loss rate) with precise numerical values ​​are all fuzzy uncertainties.

[0099] Quantify the uncertainty factors: Quantify the degree of the above-mentioned uncertainties through statistical methods, such as probability theory and mathematical statistics. This helps to more accurately consider these uncertainties in subsequent risk assessments.

[0100] The environmental adaptability index (such as wind speed adaptability) and the operating efficiency index (such as power loss rate) are fuzzified and divided into different fuzzy sets (such as low wind speed sensitivity, medium wind speed adaptation, high wind speed tolerance; slight loss, moderate loss, severe loss, etc.); the membership function is defined for each fuzzy set to describe the degree to which the index value belongs to the set; the fuzzy reasoning rules for the health status of the equipment under different conditions are defined by combining expert experience and historical data. For example, "if the wind speed is in the high wind speed sensitive area and the temperature exceeds the suitable range, the probability of equipment failure increases"; these rules are used to infer the health status of the equipment based on real-time data and environmental conditions; DST is used to integrate information from different sources (such as multiple sensors, historical data, expert judgment) to form a more reliable belief distribution about the health status of the equipment; by calculating the basic belief allocation (BBA), DST can comprehensively consider the conflicts and consistency between various evidences, thereby obtaining a more robust evaluation result; combined with historical failure data and expert knowledge base, the membership function, rule base and BBA in the fuzzy logic system are regularly updated in the evidence theory. This helps maintain the accuracy and timeliness of the system; the introduction of incremental learning methods enables the basic trust assignment of fuzzy logic rules and evidence theory to be fine-tuned based on newly collected real-time data; incremental learning avoids the need to completely retrain the model, thereby improving the adaptability and responsiveness of the system.

[0101] The effects of the above technical solution are as follows: by classifying and quantifying the uncertainty factors in real-time environmental data and equipment operating status, the solution can more accurately reflect the complexity and dynamics in equipment health risk assessment. This helps to reduce the assessment errors caused by uncertainty factors and improve the accuracy of risk assessment; by fuzzifying environmental adaptability indicators and operating efficiency indicators through fuzzy sets and membership functions, and combining fuzzy rule bases for reasoning, the solution can handle complex problems such as the fuzziness of equipment performance degradation boundaries. This fuzzy processing method enhances the adaptability and robustness of the system, enabling it to better cope with various uncertainties and ambiguities; DST can integrate information from different sources, such as multiple sensors, historical data, and expert judgments, to form a more reliable belief distribution about the health status of the equipment. This helps to reduce information islands and one-sidedness and improve the reliability and comprehensiveness of the assessment; combined with historical fault data and expert knowledge bases, membership functions, rule bases in fuzzy logic systems, and basic trust allocation (BBA) in evidence theory are regularly updated. This regular update mechanism ensures that the system can keep up with changes in equipment status and environmental conditions, and maintain the accuracy and timeliness of the assessment; the introduction of incremental learning methods enables the basic trust allocation of fuzzy logic rules and evidence theory to be fine-tuned based on newly collected real-time data rather than completely retrained. This not only improves the response speed of the system, but also reduces the update cost, and realizes the continuous optimization and intelligent upgrade of the system; the solution provides comprehensive evaluation results and decision support for equipment management and maintenance by integrating multi-source information, handling uncertain factors, and realizing fuzzy reasoning and evidence theory integration. This helps enterprises and institutions better predict and manage equipment health risks, formulate scientific and reasonable maintenance plans and emergency plans, reduce failure rates and repair costs, and improve equipment reliability and safety.

[0102] In one embodiment of the present invention, S5 includes:

[0103] S51, inputting the real-time monitoring data into the integrated learning model to obtain the real-time prediction result of the health status of the equipment, wherein the real-time prediction result includes the fault type, fault degree and remaining service life;

[0104] S52. According to the dynamic risk assessment results, the health threshold is dynamically adjusted, and the warning thresholds corresponding to different risk levels are considered to perform graded warnings;

[0105] S53. When the prediction result is lower than the set health threshold, the early warning mechanism is automatically triggered to send an alarm message to relevant personnel or systems; the alarm message includes the fault type, location, severity and recommended measures;

[0106] S54. Design an early warning response process, which includes preliminary diagnosis, emergency maintenance preparation, and resource scheduling; and establish an emergency response team to respond quickly and take effective measures after receiving the early warning;

[0107] S55. Based on the early warning effect evaluation mechanism, obtain the maintenance records and data feedback after the early warning, evaluate the accuracy and effectiveness of the early warning mechanism, and continuously optimize the early warning strategy and process based on the evaluation results.

[0108] The working principle of the above technical solution is as follows: collect real-time monitoring data of the equipment, which may include vibration signals, temperature, pressure and other sensor information; use pre-trained integrated learning models (such as random forests, gradient boosting trees, etc.) to process the input data to obtain the real-time health status prediction results of the equipment; the prediction results include fault type (such as bearing fault, gear fault, etc.), fault severity (minor, moderate, severe) and remaining useful life (RUL, Remaining Useful Life); based on real-time monitoring data and historical data, evaluate the current risk status of the equipment, considering the possibility and consequences of the failure; according to the risk assessment results, dynamically adjust the health threshold to ensure the sensitivity and accuracy of the early warning mechanism; set early warning thresholds corresponding to different risk levels to achieve graded early warning, so as to take different response measures for different levels of risks; When the prediction result is lower than the set health threshold, the early warning mechanism is automatically triggered; the alarm information is sent to relevant personnel or systems, and the alarm information includes the fault type, location, severity and recommended measures for rapid response and processing; a clear early warning response process is designed, including preliminary diagnosis (confirming the fault type and degree through remote monitoring or on-site inspection), emergency maintenance preparation (preparing necessary maintenance tools and spare parts) and resource scheduling (arranging maintenance personnel and equipment to arrive at the site); a professional emergency response team is established, and the team members should have rich maintenance experience and rapid response capabilities to ensure that effective measures can be taken quickly after receiving the early warning; based on the maintenance records and data feedback after the early warning, the accuracy and effectiveness of the early warning mechanism are evaluated; according to the evaluation results, the early warning strategy and process are continuously optimized to improve the sensitivity and accuracy of the early warning mechanism, reduce false alarms and missed alarms, and ensure the safe and stable operation of the equipment. For example, suppose that the fan equipment of a factory has an abnormal vibration signal during operation, and the real-time monitoring data is input into the integrated learning model. The model predicts that the fan may have a bearing fault, and the fault degree is medium, and the remaining service life is 30 days. According to the dynamic risk assessment results, the system dynamically adjusts the health threshold and triggers an intermediate warning. The alarm information was sent to the plant maintenance team, who responded quickly, conducted a preliminary diagnosis, and prepared the necessary repair tools and spare parts. After emergency repairs, the fault was eliminated, ensuring the safe and stable operation of the wind turbine. At the same time, based on the maintenance records and data feedback of this warning, the system optimized the warning mechanism to improve the accuracy and efficiency of future warnings.

[0109] The effects of the above technical solutions are: obtaining real-time prediction results of the health status of the equipment, including the type of fault, degree of fault and remaining service life; dynamically adjusting the health threshold according to the dynamic risk assessment results; considering the warning thresholds corresponding to different risk levels, and conducting graded warnings; automatically triggering the warning mechanism when the prediction result is lower than the set health threshold; sending alarm information to relevant personnel or systems, and the alarm information includes the type of fault, location, severity and recommended measures; designing an early warning response process, including preliminary diagnosis, emergency maintenance preparation and resource scheduling; establishing an emergency response team to respond quickly and take effective measures after receiving the early warning; obtaining maintenance records and data feedback after the early warning based on the early warning effect evaluation mechanism; evaluating the accuracy and effectiveness of the early warning mechanism, and continuously optimizing the early warning strategy and process based on the evaluation results; through real-time monitoring and Prediction can timely discover potential equipment failures to avoid the occurrence or expansion of failures; hierarchical warning and automated warning mechanisms can quickly notify relevant personnel or systems, shorten response time, and improve maintenance efficiency; through dynamic risk assessment and warning threshold adjustment, the health status of equipment can be judged more accurately to reduce false alarms and missed alarms; timely repair and maintenance can extend the service life of equipment and reduce the occurrence rate of failures; based on the warning effect evaluation mechanism, the warning strategy and process can be continuously optimized to improve the accuracy and effectiveness of warnings; by collecting and analyzing maintenance records and data feedback, the maintenance needs and rules of equipment can be more accurately understood, and more reasonable maintenance plans can be formulated; real-time monitoring and early warning mechanisms can timely discover and deal with safety hazards and prevent accidents; the establishment of an emergency response team can quickly respond to emergencies and reduce accident losses.

[0110] In one embodiment of the present invention, the S52 includes:

[0111] Based on the risk level (such as low risk, medium risk, and high risk) output by the dynamic risk assessment model, a dynamic mapping relationship with the equipment health status threshold is established;

[0112] Through adaptive algorithms, based on multi-dimensional information, including historical fault data, equipment type, and operating environment, health thresholds at different risk levels are dynamically adjusted;

[0113] Within each risk level, the warning threshold is further refined, for example, setting three levels: primary warning, intermediate warning, and emergency warning; each level corresponds to different fault severity and urgency;

[0114] Use machine learning algorithms (such as support vector machines and neural networks) to learn historical fault data and automatically optimize the threshold settings for each warning level;

[0115] Incorporate environmental factors into the adjustment of early warning thresholds; based on the environmental adaptability model, monitor the changes in environmental factors in real time, predict the changing trend of the equipment health status through real-time environmental data, and adjust the early warning thresholds accordingly.

[0116] The working principle of the above technical solution is as follows: based on the output of the dynamic risk assessment model, determine the risk level of the equipment (such as low risk, medium risk, and high risk); establish a dynamic mapping relationship between the risk level and the equipment health status threshold. As the risk level increases, the health threshold decreases accordingly to warn of potential faults in advance; use adaptive algorithms to dynamically adjust the health thresholds under different risk levels based on multi-dimensional information such as historical fault data, equipment type, and operating environment; the adaptive algorithm can automatically adapt to changes in equipment status and environmental conditions to ensure the accuracy and timeliness of the early warning mechanism; within each risk level, further refine the early warning threshold, such as setting three levels of primary warning, intermediate warning, and emergency warning; each warning level corresponds to different fault severity and urgency, which helps relevant personnel or systems take corresponding countermeasures according to different situations; learn historical fault data through machine learning algorithms (such as support vector machines and neural networks) to automatically optimize the threshold settings of each warning level. This helps to improve the accuracy and generalization of the early warning mechanism; incorporate environmental factors into the adjustment of early warning thresholds by real-time monitoring of changes in environmental factors such as temperature, humidity, pressure, etc.; based on the environmental adaptability model, use real-time environmental data to predict the changing trend of equipment health status; based on the prediction results, dynamically adjust the early warning threshold to adapt to the impact of environmental changes on equipment health status.

[0117] The effect of the above technical solution is: by establishing a dynamic mapping relationship between the risk level output by the dynamic risk assessment model and the equipment health status threshold, it is ensured that as the risk level increases, the health threshold can be reduced accordingly, thereby early warning of potential failures. This dynamic adjustment mechanism improves the accuracy and timeliness of the warning, which helps to avoid or reduce the losses caused by failures; the adaptive algorithm is used to dynamically adjust the health threshold according to multi-dimensional information (such as historical fault data, equipment type, and operating environment), so that the system can automatically adapt to changes in equipment status and environmental conditions. This adaptability and flexibility helps to ensure that the early warning mechanism remains effective under different circumstances; further refine the warning threshold within each risk level, such as setting three levels of primary warning, intermediate warning, and emergency warning, which helps relevant personnel or systems take different response measures according to different levels of warning. This refinement improves the response efficiency and enables repair and maintenance work to be carried out more targeted; historical fault data is learned through machine learning algorithms (such as support vector machines and neural networks) to automatically optimize the threshold settings of each warning level. This learning mechanism can continuously improve the system's predictive ability and accuracy, thereby optimizing the early warning strategy and reducing false alarms and missed alarms. It incorporates environmental factors into the adjustment of early warning thresholds, monitors changes in environmental factors in real time, and predicts the changing trend of equipment health status based on the environmental adaptability model, which helps to more comprehensively consider the impact of the equipment operating environment on the equipment health status. This comprehensiveness improves the accuracy and practicality of the early warning mechanism. Through accurate early warnings and timely responses, the allocation of maintenance resources can be optimized, and unnecessary downtime and maintenance costs can be reduced. At the same time, by optimizing the early warning strategy through machine learning, the additional costs caused by false alarms and missed alarms can be reduced.

[0118] In one embodiment of the present invention, S6 includes:

[0119] S61. Dynamically adjust maintenance plans based on health status monitoring and early warning results, prioritize maintenance tasks for high-risk equipment, and implement preventive maintenance measures, such as regular inspections, component replacement, and parameter adjustments;

[0120] S62. Based on the preset maintenance cost-benefit analysis model, evaluate the cost-benefit ratio of different maintenance strategies, and provide decision support for optimizing maintenance strategies based on the evaluation results;

[0121] S63, constructing a maintenance strategy optimization model based on a reinforcement learning algorithm, and automatically finding an optimal maintenance strategy based on multiple objective functions, the multiple objective functions including maintenance cost, equipment reliability, and downtime;

[0122] S64, further optimizing the maintenance strategy through a multi-objective optimization algorithm (such as a genetic algorithm, a particle swarm optimization algorithm) to find a balanced solution that meets multiple objectives;

[0123] S65. Continuously collect maintenance feedback data for the training and optimization of the reinforcement learning model, and based on the maintenance strategy evaluation mechanism, regularly evaluate the effectiveness and cost-effectiveness of the maintenance strategy and continuously optimize the maintenance strategy.

[0124] The working principle of the above technical solution is as follows: according to the health status monitoring and early warning results, the maintenance needs of the equipment are evaluated in real time; the maintenance tasks of high-risk equipment are arranged in priority to ensure the safe operation of key equipment; preventive maintenance measures are implemented, such as regular inspections, component replacement, parameter adjustment, etc., to reduce the possibility of failures; based on the preset maintenance cost-benefit analysis model, the costs and benefits of different maintenance strategies are evaluated; by comparing the costs and benefits of different strategies, decision support is provided for optimizing maintenance strategies; for example, assuming there are two maintenance strategies A and B, strategy A has a lower cost but relatively lower benefits, while strategy B has a higher cost but significant benefits. Through cost-benefit analysis, it can be decided to choose strategy B with higher benefits if the budget allows. A maintenance strategy optimization model is constructed using a reinforcement learning algorithm; multiple objective functions are defined, such as maintenance cost, equipment reliability, downtime, etc., as reward or penalty signals for reinforcement learning; the optimal maintenance strategy that meets multiple objectives is automatically found through continuous trial and error and learning; for example, the reinforcement learning model can simulate the operating status of the equipment under different maintenance strategies, and adjust the strategy based on the simulation results to maximize long-term rewards (such as equipment reliability) and minimize costs (such as maintenance costs and downtime); on the basis of the reinforcement learning algorithm, multi-objective optimization algorithms (such as genetic algorithms and particle swarm optimization algorithms) are introduced; these algorithms can find a balance solution between multiple objectives and further optimize the maintenance strategy; through iterative search and selection, the optimal or approximately optimal solution that meets all objectives is found; maintenance feedback data is continuously collected, including maintenance cost, equipment failure rate, downtime, etc.; these data are used to train and optimize the reinforcement learning model to improve the model's predictive ability and accuracy; based on the maintenance strategy evaluation mechanism, the effectiveness and cost-effectiveness of the maintenance strategy are regularly evaluated; based on the evaluation results, the maintenance strategy is continuously optimized to adapt to changes in equipment status and environmental conditions.

[0125] The effects of the above technical solution are as follows: According to the health status monitoring and early warning results: By real-time monitoring of the health status of the equipment, potential faults can be discovered in time, and the maintenance plan can be dynamically adjusted according to the early warning results. This dynamic adjustment mechanism ensures that maintenance tasks can be arranged for high-risk equipment first, thereby effectively preventing the occurrence of faults and improving the reliability and safety of the equipment; through preventive maintenance measures such as regular inspections, component replacement, and parameter adjustment, the service life of the equipment can be further extended, the occurrence of sudden failures can be reduced, and the overall performance of the equipment can be improved; based on the preset maintenance cost-benefit analysis model: the model can comprehensively consider the costs and benefits of different maintenance strategies and provide decision support for optimizing maintenance strategies. By comparing the cost-effectiveness ratio of different strategies, the most cost-effective maintenance plan can be selected, thereby reducing maintenance costs and improving economic benefits; by constructing a maintenance strategy optimization model and introducing multiple objective functions (such as maintenance cost, equipment reliability, downtime, etc.), the reinforcement learning algorithm can automatically find the optimal maintenance strategy that meets these goals. This automated optimization process greatly improves the optimization efficiency of the maintenance strategy and reduces human intervention and decision-making errors; on the basis of the reinforcement learning algorithm, a multi-objective optimization algorithm (such as genetic algorithm, particle swarm optimization algorithm, etc.) is introduced to further optimize the maintenance strategy. These algorithms can find a balance between multiple objectives to ensure that the maintenance strategy is optimal or near-optimal in terms of equipment reliability, maintenance cost reduction, and downtime reduction. By collecting various data during the maintenance process (such as maintenance costs, equipment failure rates, downtime, etc.), they can provide rich information for the training and optimization of reinforcement learning models. These data help the model to more accurately understand the operating status and maintenance needs of the equipment, thereby improving the optimization effect of the maintenance strategy. The effectiveness and cost-effectiveness of the maintenance strategy are regularly evaluated, and the maintenance strategy is adjusted and optimized based on the evaluation results. This continuous improvement mechanism ensures that the maintenance strategy can adapt to changes in equipment status and environmental conditions and maintain its effectiveness and economy.

[0126] One embodiment of the present invention, as Figure 2 As shown, an equipment health management system for a wind turbine generator set, the system comprising:

[0127] Data acquisition module: acquires multi-source data of wind turbines, including operation data (such as vibration, temperature, power, etc.), historical fault records, environmental data (such as wind speed, wind direction, temperature, humidity, etc.) and maintenance logs, and pre-processes the acquired multi-source data, including cleaning, denoising, standardization and normalization of the data; uses data fusion technology to integrate the pre-processed multi-source data into a unified data format;

[0128] Feature extraction module: Based on domain knowledge and data characteristics, it extracts key features that reflect the health status of the equipment, and uses feature selection algorithms to select the feature subset with the most predictive value for the health status of the equipment;

[0129] Model formation module: Design an integrated learning framework, which includes multiple base learners (such as decision trees, support vector machines, neural networks, etc.), each of which is trained using a different feature subset, and the prediction results of multiple base learners are combined through a voting mechanism to form a final prediction model; and use cross-validation and grid search techniques to optimize model parameters;

[0130] Parameter update module: Based on real-time environmental data and equipment operating status, a dynamic risk assessment model is constructed. The health risk level of the equipment under different environmental conditions is evaluated in real time according to the dynamic risk assessment model. The parameters of the risk assessment model are continuously updated in combination with historical fault data and expert knowledge.

[0131] Mechanism trigger module: inputs real-time monitoring data into the integrated learning model to obtain the prediction results of the equipment health status, sets the health threshold according to the dynamic risk assessment results, and triggers the early warning mechanism when the prediction result is lower than the threshold;

[0132] Strategy optimization module: Dynamically adjust the maintenance plan according to the health status monitoring and early warning results to achieve predictive maintenance, and continuously optimize the maintenance strategy based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals.

[0133] The working principle of the above technical solution is as follows: obtain data from multiple sources from wind turbines, including real-time operating data (such as vibration, temperature, power, etc.), historical fault records, environmental data (such as wind speed, wind direction, temperature, humidity, etc.) and maintenance logs; pre-process these multi-source data, including data cleaning (removing invalid or erroneous data), denoising (reducing random fluctuations in data), standardization (scaling data to the same scale) and normalization (converting data to values ​​within a specific range) to improve data quality and consistency; use data fusion technology to integrate the pre-processed multi-source data into a unified Data format is convenient for subsequent analysis and modeling; based on domain knowledge and data characteristics, key features that can reflect the health status of the equipment are identified; through feature selection algorithms, the feature subset with the most predictive value for the health status of the equipment is selected from a large number of features to reduce the complexity of the model and improve the prediction accuracy; an integrated learning framework is designed, which contains multiple base learners, such as decision trees, support vector machines, neural networks, etc.; each base learner is trained with different feature subsets to capture different information in the data; the prediction results of multiple base learners are combined through a voting mechanism to form the final prediction model. This integrated learning method can improve the stability and accuracy of the model; the model parameters are optimized using cross-validation and grid search techniques to further improve the performance of the model; a dynamic risk assessment model is constructed based on real-time environmental data and equipment operating status; the model can evaluate the health risk level of the equipment under different environmental conditions in real time; combined with historical fault data and expert knowledge, the parameters of the risk assessment model are continuously updated to adapt to changes in equipment status; real-time monitoring data is input into the integrated learning model to obtain the prediction results of the equipment health status; according to the dynamic risk assessment results, the health threshold is set. When the prediction result is lower than the threshold, the early warning mechanism is triggered to remind the operator to pay attention to the health status of the equipment; according to the health status monitoring and early warning results, the maintenance plan is dynamically adjusted to achieve predictive maintenance; based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals, the maintenance strategy is continuously optimized. The reinforcement learning algorithm can continuously adjust the strategy based on the feedback during the maintenance process to achieve the best maintenance effect.

[0134] The effects of the above technical solutions are as follows: by acquiring multi-source data and performing preprocessing, including cleaning, denoising, standardization and normalization, and integrating data using data fusion technology, a more comprehensive and accurate information basis is provided for the model; key features are extracted based on domain knowledge and data characteristics, and the feature subsets with the most predictive value are screened out through feature selection algorithms, further improving the prediction accuracy of the model; an integrated learning framework is designed, which includes multiple base learners, each of which is trained using a different feature subset, and the prediction results are combined through a voting mechanism to reduce the risk of a single model and improve the robustness and stability of the model; model parameters are optimized using cross-validation and grid search techniques to ensure that the model can perform well under different conditions; a dynamic risk assessment model is constructed that can evaluate the health risk level of equipment under different environmental conditions in real time, providing a scientific basis for early warning and maintenance decisions; historical fault data and expert knowledge are combined to ensure that the model can be used to predict the health risk level of equipment under different environmental conditions. The parameters of the risk assessment model are continuously updated so that the model can adapt to changes in equipment status and environmental conditions. According to the health status monitoring and early warning results, the maintenance plan is dynamically adjusted to achieve predictive maintenance, avoiding unnecessary downtime and maintenance costs. The maintenance strategy is optimized based on the reinforcement learning algorithm, combining the maintenance cost and equipment reliability goals to achieve the best balance between maintenance efficiency and equipment performance. Through real-time monitoring and early warning, abnormal conditions in the health status of equipment are discovered in time, and corresponding maintenance measures are taken, which effectively reduces the failure rate and extends the service life of the equipment. The entire health management method integrates multiple links such as data preprocessing, feature extraction, integrated learning, dynamic risk assessment, early warning mechanism and maintenance strategy optimization, realizing the intelligent and automated health management of wind turbines. The application of data-driven and machine learning algorithms is strengthened, which improves the intelligence level of wind turbine health management and provides strong support for the digital transformation of the wind power industry.

[0135] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for equipment health management of a wind turbine generator set, characterized in that: The method comprises: S1. Acquire multi-source data of wind turbines, pre-process the acquired multi-source data, and integrate the pre-processed multi-source data into a unified data format using data fusion technology; S2. Extract key features that reflect the health status of the equipment based on domain knowledge and data characteristics, and select the feature subset with the most predictive value for the health status of the equipment through feature selection algorithm; S3. Design an integrated learning framework, which includes multiple base learners, each of which is trained using a different feature subset, and combines the prediction results of multiple base learners through a voting mechanism to form a final prediction model; S4. Based on real-time environmental data and equipment operating status, a dynamic risk assessment model is constructed. The health risk level of the equipment under different environmental conditions is evaluated in real time according to the dynamic risk assessment model. The parameters of the risk assessment model are continuously updated in combination with historical failure data and expert knowledge. S5. Input the real-time monitoring data into the integrated learning model to obtain the prediction results of the equipment health status. According to the dynamic risk assessment results, the health threshold is set. When the prediction result is lower than the threshold, the early warning mechanism is triggered. S6. Dynamically adjust the maintenance plan based on the health status monitoring and early warning results to achieve predictive maintenance, and continuously optimize the maintenance strategy based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals.

2. According to claim 1, a method for managing the health of a wind turbine generator set, characterized in that: Said S1 comprises: S11. Based on a high-precision sensor network, key components of wind turbines are covered and key parameters of key parts are monitored in real time; S12. Establish a historical fault database to record in detail the type, time of occurrence, cause, repair measures and cost of each fault, as well as the changes in operating data before and after the fault; S13, integrate meteorological data sources and obtain external environmental parameters, systematize maintenance logs, record relevant information of each maintenance activity, and associate it with specific wind turbines; S14, using a rule-based cleaning method to clean the data, and applying an adaptive filtering algorithm to denoise the data; S15. According to the distribution characteristics of the data, the data are converted to the same scale through Z-score standardization, and the data are mapped to a specific range by normalization method for specific features or feature combinations; S16. The data is fused through deep learning models to extract high-level features; at the same time, principal component analysis is applied to reduce data dimensions.

3. According to claim 1, a method for managing the health of a wind turbine generator set, characterized in that: The S2 comprises: S21. Extracting first related features based on the physical model, where the first related features include frequency spectrum features of the vibration signal and trend features of the temperature signal; S22. Extracting second related features by using a statistical learning method, where the second related features include statistical features and time series features of the data; S23. Use deep learning technology to automatically extract high-dimensional features through autoencoders or convolutional neural networks to capture complex patterns in data; S24. Use correlation analysis or mutual information method to identify features that are highly correlated with the health status of the equipment; based on the model, use random forest to further screen feature subsets according to the importance score of the features.

4. According to claim 3, a method for managing the health of a wind turbine generator set is characterized in that: The S24 comprises: Based on all the extracted features, a multi-dimensional correlation matrix is ​​constructed; Convert the multi-dimensional correlation matrix into a feature network, where nodes represent features and edges represent the strength of correlation between features; use graph theory algorithms to analyze the centrality and influence of nodes in the network and identify features that play a key role in the network; Based on the historical health status records of the equipment, sensitivity analysis techniques are used to evaluate the sensitivity of each feature to changes in the health status of the equipment; Through dynamic time window technology, the feature correlation within each time window is recalculated and analyzed. Based on the analysis results, the correlation, centrality, sensitivity and time dynamics of each feature are comprehensively evaluated to screen out the most representative feature subset that is highly correlated with the health status of the equipment. The features are ranked using feature importance scores to further optimize the feature subset.

5. The device health management method for a wind turbine generator set according to claim 1, characterized in that: The S3 includes: S31. Based on data characteristics and problem complexity, select various types of base learners, and design specific input formats and preprocessing steps for base learners according to different types of features; S32, adopting an integration strategy to combine the prediction results of multiple base learners to improve the stability and generalization ability of the model; S33, using cross-validation techniques to evaluate the performance of the integrated model, and automatically adjusting model parameters through grid search to find the optimal configuration; S34. Based on a diversity enhancement mechanism, the diversity enhancement mechanism includes feature subset perturbation and model structure perturbation to increase the differences between base learners.

6. The equipment health management method for a wind turbine generator set according to claim 1, characterized in that: The S4 comprises: S41. Based on the real-time environmental data and the equipment operation status, a risk assessment index system is constructed, wherein the risk assessment index system includes an environmental adaptability index, an operation efficiency index, and a failure probability index; S42. Use machine learning algorithms to build a dynamic risk assessment model, and evaluate the health risk level of the equipment under different environmental conditions in real time based on the built dynamic risk assessment model; S43. Introduce fuzzy logic or evidence theory to deal with uncertainty issues in risk assessment; combine historical failure data and expert knowledge base to regularly update the parameters and rules of the risk assessment model; S44, through the online learning mechanism, the model is updated online using the newly collected real-time data, and the model is adaptively and continuously optimized; S45. Based on the model performance monitoring and feedback mechanism, the prediction accuracy and stability of the model are evaluated in real time, and model degradation problems are discovered and handled.

7. A method for managing equipment health of a wind turbine generator set according to claim 6, characterized in that: The S43 includes: Classify the uncertainty factors in real-time environmental data and equipment operating status, and quantify the degree of each type of uncertainty through statistical methods; Based on fuzzy sets and membership functions, fuzzy processing is performed on environmental adaptability indicators and operating efficiency indicators. Through the fuzzy rule base, combined with expert experience and historical data, fuzzy reasoning rules for equipment health status under different conditions are defined; Apply Dempster-Shafer evidence theory to integrate information from different sources to form a belief distribution about the health status of the equipment; Combine historical fault data and expert knowledge base to regularly update membership functions, rule bases in fuzzy logic systems, and basic trust allocation in evidence theory; Based on the introduction of incremental learning methods, the basic trust assignment of fuzzy logic rules and evidence theory can be fine-tuned based on newly collected real-time data.

8. The device health management method for a wind turbine generator set according to claim 1, characterized in that: The S5 comprises: S51, inputting the real-time monitoring data into the integrated learning model to obtain the real-time prediction result of the health status of the equipment, wherein the real-time prediction result includes the fault type, fault degree and remaining service life; S52. According to the dynamic risk assessment results, the health threshold is dynamically adjusted, and the warning thresholds corresponding to different risk levels are considered to perform graded warnings; S53. When the prediction result is lower than the set health threshold, the early warning mechanism is automatically triggered and an alarm message is sent to relevant personnel or systems; S54. Design an early warning response process, which includes preliminary diagnosis, emergency maintenance preparation, and resource scheduling; and establish an emergency response team to respond quickly and take effective measures after receiving the early warning; S55. Based on the early warning effect evaluation mechanism, obtain the maintenance records and data feedback after the early warning, evaluate the accuracy and effectiveness of the early warning mechanism, and continuously optimize the early warning strategy and process based on the evaluation results.

9. The device health management method for a wind turbine generator set according to claim 1, characterized in that: The S6 comprises: S61. Dynamically adjust maintenance plans and implement preventive maintenance measures based on health status monitoring and early warning results; S62. Based on the preset maintenance cost-benefit analysis model, evaluate the cost-benefit ratio of different maintenance strategies, and provide decision support for optimizing maintenance strategies based on the evaluation results; S63. Based on the reinforcement learning algorithm, a maintenance strategy optimization model is constructed, and based on multiple objective functions, the optimal maintenance strategy is automatically found; S64, further optimizing the maintenance strategy through a multi-objective optimization algorithm to find a balanced solution that meets multiple objectives; S65. Continuously collect maintenance feedback data for the training and optimization of the reinforcement learning model, and based on the maintenance strategy evaluation mechanism, regularly evaluate the effectiveness and cost-effectiveness of the maintenance strategy and continuously optimize the maintenance strategy.

10. An equipment health management system for a wind turbine generator set, characterized in that: The system comprises: Data acquisition module: acquires multi-source data of wind turbines, pre-processes the acquired multi-source data, and integrates the pre-processed multi-source data into a unified data format using data fusion technology; Feature extraction module: Based on domain knowledge and data characteristics, it extracts key features that reflect the health status of the equipment, and uses feature selection algorithms to select the feature subset with the most predictive value for the health status of the equipment; Model formation module: Design an integrated learning framework, which includes multiple base learners, each of which is trained using a different feature subset, and combines the prediction results of multiple base learners through a voting mechanism to form a final prediction model; Parameter update module: Based on real-time environmental data and equipment operating status, a dynamic risk assessment model is constructed. The health risk level of the equipment under different environmental conditions is evaluated in real time according to the dynamic risk assessment model. The parameters of the risk assessment model are continuously updated in combination with historical fault data and expert knowledge. Mechanism trigger module: inputs real-time monitoring data into the integrated learning model to obtain the prediction results of the equipment health status, sets the health threshold according to the dynamic risk assessment results, and triggers the early warning mechanism when the prediction result is lower than the threshold; Strategy optimization module: Dynamically adjust the maintenance plan according to the health status monitoring and early warning results to achieve predictive maintenance, and continuously optimize the maintenance strategy based on the reinforcement learning algorithm, combined with the maintenance cost and equipment reliability goals.

Images