New energy power generation equipment health management platform based on large model

Abstract

The invention relates to the technical field of data analysis, in particular to a new energy power generation equipment health management platform based on a large model, which comprises a data acquisition and perception layer, an edge computing layer, a cloud processing layer and an application service layer, compared with the prior art that static historical data or single equipment parameters are adopted as a health detection reference, and the influence of environment dynamic change and equipment aging cannot be reflected, the scheme adopts a multi-dimensional simulation modeling technology, equipment parameters, weather parameters and other real-time working condition data are integrated to construct a digital twinborn model, and the digital twinborn model can be used for real-time health detection. Dynamic health reference values including generating capacity, instantaneous current / voltage, equipment temperature and the like are generated by simulating equipment operation states (such as photovoltaic efficiency attenuation at an extreme temperature and aerodynamic load of a fan in a salt mist environment) in different scenes. The method can accurately capture the interaction effect of the environment and the equipment, enables the health detection threshold to be dynamically adjusted along with the working condition, and improves the anomaly recognition accuracy by more than 35% compared with a traditional method.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a large-scale model-based health management platform for new energy power generation equipment. Background Art

[0002] New energy power generation equipment faces severe challenges in long-term stable operation under complex natural environments and dynamic grid conditions. The existing health management technology system has significant defects in environmental adaptability, real-time performance and diagnostic accuracy.

[0003] Taking offshore wind turbines as an example, traditional health monitoring methods rely on fixed thresholds or historical experience data to set abnormal alarm lines. However, the coupling of multiple factors such as equipment aging, salt spray corrosion, and wave impact leads to prominent problems of sensor drift and performance degradation. Industry measured data shows that in areas where the salt spray concentration is higher than 5 mg per cubic meter, the average monthly drift error of the current sensor can reach 1.5%. Existing calibration technology only supports regular manual maintenance or a simple master-slave switching mechanism, and cannot achieve dynamic error compensation, causing the equipment false alarm rate to remain above 25% for a long time.

[0004] At the data processing level, traditional cloud-based centralized architectures are limited by network bandwidth and transmission latency. For key features such as high-frequency vibration signals and instantaneous current fluctuations, uploading full data results in an effective information loss rate of over 40%. A case study at a Nordic wind farm demonstrated that the 12 kHz high-frequency vibration signal caused by early gearbox wear was not promptly identified due to data compression distortion, ultimately causing a transmission chain rupture accident and indirect losses amounting to millions of euros.

[0005] In terms of fault diagnosis models, current mainstream technologies are based on single physical models or statistical learning algorithms. For example, expert systems rely on vibration spectrum feature matching. The update of their knowledge base lags seriously behind the speed of equipment technology iteration and cannot effectively identify new complex fault modes. The International Renewable Energy Agency report pointed out that 35% of new fault types added in the past three years are not covered by traditional models, and the difference between manual inspection results and algorithm diagnosis conclusions is as high as 30%. However, due to the lack of a closed-loop feedback path, the model iteration cycle is as long as six months, which seriously restricts operation and maintenance efficiency.

[0006] Furthermore, in extreme environmental scenarios, such as sandstorms in desert photovoltaic power plants or biofouling corrosion in tidal turbines, existing sensor calibration technology struggles to dynamically adapt to dramatic changes in environmental parameters. This can lead to distortion of critical monitoring data such as temperature and pressure, which can lead to misjudgments. Industry data shows that unplanned downtime due to sensor drift accounts for 45% of total equipment downtime, severely undermining the economic benefits of renewable energy power generation.

[0007] In this context, there is an urgent need to build a new health management technology system that integrates multi-physics field dynamic simulation, edge intelligent computing, adaptive calibration and closed-loop optimization feedback, so as to systematically solve core problems such as environmental interference suppression, real-time data processing, complex fault identification and model self-evolution. Summary of the Invention

[0008] In order to overcome the problems raised in the above background technology, the present invention proposes a new energy power generation equipment health management platform based on a large model.

[0009] The technical solution of the present invention is: a large-scale model-based new energy power generation equipment health management platform, including: The data acquisition and perception layer is used to collect real-time operating data through multi-source sensors and IoT devices, and integrate external data such as meteorological parameters and power grid status. By collecting real-time operating data through multi-source sensors and IoT devices and integrating external data such as meteorological parameters and power grid status, it can obtain comprehensive and multi-angle operating information of new energy power generation equipment, providing a rich and accurate data foundation for subsequent analysis and diagnosis, and facilitating a more comprehensive understanding of equipment operating conditions. At the edge computing layer, a lightweight anomaly detection model is deployed to screen the data collected by the data acquisition and perception layer and extract key features. This lightweight anomaly detection model is deployed to screen the collected data and extract key features. This allows for preliminary processing at the data source, effectively reducing the amount of data transmitted to the cloud, lowering the network transmission burden, and improving data transmission efficiency. Data screening and feature extraction at the edge simultaneously enable rapid response to device anomalies, timely identification of potential problems, and avoid untimely fault handling due to data transmission delays, thereby improving the real-time performance and reliability of the system. The cloud processing layer is used to build digital twin simulation models, combine external data to predict the operating status of new energy power generation equipment, generate dynamic health benchmark values, and perform health diagnosis and management through large models. Building digital twin simulation models, combining external data to predict the operating status of new energy power generation equipment, and generating dynamic health benchmark values ​​can more accurately assess the health status of equipment, identify potential health risks in advance, implement preventive maintenance of equipment, reduce the occurrence of equipment failures, and extend equipment service life. The application service layer is used to provide equipment health assessment, operation and maintenance work order management, visualization services, and decision support based on the data results of the cloud processing layer. The equipment health assessment service based on the data results of the cloud processing layer can present complex equipment health information to users in an intuitive manner, allowing users to quickly understand the health status of the equipment and facilitate equipment management and maintenance decisions.

[0010] Preferably, the data acquisition and perception layer also includes an intelligent calibration module, which is used to perform intelligent calibration on the sensors in the data acquisition and perception layer. The intelligent calibration module can perform intelligent calibration on the sensors, adjust the working status of the sensors in real time, and reduce data deviations caused by factors such as sensor errors and environmental interference, thereby ensuring that the collected data is more accurate and reliable, and providing a high-quality data foundation for subsequent equipment health management and decision-making. Among them, the intelligent calibration module adopts a master-slave redundant architecture for the core parameter sensors, specifically including: S11: Synchronous data collection: The primary sensor collects data in real time, while the backup sensor samples data synchronously at a fixed interval in its dormant state. Data synchronization between the primary and backup sensors is achieved through timestamps. This active-standby redundant architecture ensures that if one sensor fails or experiences data anomalies, the other sensor can promptly take over data collection, avoiding data collection interruptions and ensuring data continuity and integrity. S12: Decision-making based on the DS evidence theory. First, the data support of the primary and backup sensors is calculated. Then, a joint probability distribution is generated based on the confidence matrix to determine whether the data of the primary and backup sensors meet the consistency threshold. The principle formula for calculating the data support of the primary and backup sensors is: ; in, Indicates the data support of the main sensor and backup sensor, is the measured value of the main sensor, is the measurement value of the backup sensor, and The sensor's range, i.e., the maximum and minimum values ​​it can measure, is determined using the Dempster-Smithsonian evidence theory. The system calculates the data support for both the primary and backup sensors and generates a joint probability distribution based on the confidence matrix to determine whether the data from both sensors meet the consistency threshold. This comprehensive analysis and consistency determination of the two sensor data allows for timely detection of sensor anomalies, improving the system's ability to detect sensor failures and enhancing system reliability. S13: Offset determination and switching strategy. If the offset between the primary and backup sensors exceeds the threshold three times in a row, triggering the automatic switching logic, the backup sensor takes over control and issues a fault warning. If the offset between the primary and backup sensors exceeds the threshold three times in a row, triggering the automatic switching logic, the backup sensor takes over control and issues a fault warning. This automatic switching strategy enables rapid response to sensor failures, minimizing the impact of failures on system operation and improving system stability and availability. S14: After the switchover, the original primary sensor enters calibration mode, verifying its drift characteristics using a standard source input. If it cannot recover, it is marked as faulty and requires replacement. The standard source input is reference value data from the healthy reference library. After the switchover, the original primary sensor enters calibration mode, verifying its drift characteristics using a standard source input. This process helps accurately determine whether the original primary sensor is faulty, as well as the type and severity of the fault. It provides a clear basis for troubleshooting and repair, improving the efficiency and accuracy of equipment maintenance.

[0011] Preferably, when the intelligent calibration module performs intelligent calibration on the sensors in the data acquisition and perception layer, for non-critical parameter sensors, dynamic reference value weights are used to perform error compensation, specifically including: S21: Construct an environmental parameter association matrix, identify non-critical parameters and their sensitive environmental variables, and establish a compensation coefficient matrix through historical data analysis or physical model derivation. A dynamic reference value adjustment strategy is used to discretize environmental variables into a state space, and an adjustment vector group is calculated based on the deviation between the current environmental parameters and the threshold. This dynamic adjustment mechanism can respond to environmental changes in real time and compensate sensor data in a timely manner, avoiding the accumulation of data deviations caused by environmental changes and ensuring data accuracy in different environments. S22: Dynamic adjustment of reference values: discretizes environmental variables into state space, with the action space representing the reference value adjustment amplitude. A decaying learning rate strategy is employed, and the parameter association matrix and the deviation between the current environmental parameters and the threshold are input to calculate the adjustment vector group. A decaying learning rate strategy is employed during the dynamic adjustment of reference values, enabling the algorithm to quickly learn and adapt to the environment in the initial stages. As the learning process progresses, the learning rate is gradually reduced to avoid over-adjustment and oscillation. This strategy can improve the stability and convergence of the algorithm, enabling the system to better adapt to complex and changing environments. S23: Calibration Verification and Update: A sliding window mechanism is introduced to calculate the error variance after the last N adjustments. If the variance continues to increase, the exploration rate is reset, forcing the algorithm to re-explore the environment. The optimized reference value is written to the digital twin model, and the healthy baseline library is regenerated. This mechanism can promptly detect situations where the algorithm may be trapped in a local optimum or the environment has undergone significant changes. By re-exploring the environment, a more appropriate reference value adjustment strategy is found, enhancing the system's adaptability and robustness. Furthermore, the optimized reference value is written to the digital twin model, and the healthy baseline library is regenerated. This enables the healthy baseline library to more accurately reflect the normal operating status of the device in different environments, providing a more reliable basis for device health diagnosis and management. Based on the updated healthy baseline library, the device health status can be more accurately assessed, and potential fault hazards can be detected in advance.

[0012] Preferably, the edge computing layer specifically includes: A11: The fault handling module uses a baseline statistical model to filter abnormal segments and extract statistical features, performs fault detection based on the statistical features, and performs real-time fault processing based on the built-in fault strategy library. The statistical methods used by the baseline statistical model include the 3σ rule and sliding mean analysis. The baseline statistical model is used to filter abnormal segments and extract statistical features using statistical methods such as the 3σ rule and sliding mean analysis, and perform fault detection based on these statistical features. The 3σ rule can effectively identify abnormal points that deviate far from the normal data distribution, and the sliding mean analysis can smooth data fluctuations and capture changes in data trends. The combination of the two can more accurately detect faults during equipment operation and reduce the probability of false alarms and missed alarms. In addition, the built-in fault strategy library allows for rapid real-time fault processing based on the fault strategy library once a fault is detected. This rapid response mechanism can promptly control the scope of fault impact, prevent further deterioration of the fault, reduce equipment downtime and production losses, and improve the operational stability and reliability of new energy power generation equipment. A12: The data filtering module is used to filter the data collected by the data acquisition and perception layer based on the data results of the fault identification model and preset rules, retaining only abnormal data and key feature data for upload to the cloud processing layer. This module effectively reduces unnecessary data transmission, lowers network bandwidth usage, and improves data transmission efficiency. It is particularly suitable for renewable energy power generation equipment, which is often deployed in remote areas with limited network bandwidth. A13: The scheduling optimization module is used to establish a compute node health assessment model and dynamically allocate tasks based on CPU and memory load. This dynamic scheduling method can rationally allocate computing tasks based on the actual load of each compute node, avoiding situations where some compute nodes are overloaded while others are idle, improving computing resource utilization, and ensuring efficient operation of the edge computing layer.

[0013] Preferably, the cloud processing layer specifically includes: A21: The simulation module is used to build a digital twin simulation model. This module simulates new energy power generation equipment based on real-time environmental data and external data to obtain predicted operating status data for the new energy power generation equipment. This module builds a digital twin simulation model that comprehensively considers real-time environmental data (such as wind speed and light intensity) and external data (such as grid dispatch instructions and weather forecasts) to simulate new energy power generation equipment. This comprehensive simulation approach can more accurately simulate the operation of equipment under different operating conditions, obtaining more realistic predicted operating status data, providing a reliable basis for subsequent anomaly identification and fault diagnosis. A22: The anomaly identification module identifies anomalies based on real-time data and predicted operating status data. By comparing and analyzing real-time data with the predicted operating status data obtained by the simulation module, anomalies during equipment operation can be quickly identified. This data comparison-based anomaly identification method has high sensitivity and accuracy, allowing for the timely detection of subtle changes in equipment performance, saving valuable time for subsequent fault diagnosis. A23: The fault diagnosis module is used to diagnose faults in renewable energy power generation equipment based on the data generated by the anomaly identification module. Based on the data provided by the anomaly identification module, the fault diagnosis module can deeply analyze the equipment's operating data and status information to accurately locate the specific location and cause of the fault. This helps operations and maintenance personnel quickly implement effective repair measures, shorten equipment downtime, and improve equipment availability.

[0014] Preferably, the cloud processing layer, when working, specifically includes: S31: Multi-source data fusion and processing. This first receives filtered abnormal data, key feature data, and external environmental parameters from the edge computing layer, receives them through a distributed message queue, cleans and aligns the data, and finally converts the heterogeneous data into a standardized format that can be parsed by the digital twin model. The data received from the edge computing layer may contain noise, missing values, or inconsistent timestamps. Through data cleaning and alignment operations, invalid data can be removed, missing values ​​can be filled, and the data time base can be unified, thereby improving data quality and usability and providing a reliable foundation for subsequent analysis and modeling. S32: Dynamically update the digital twin simulation model. The simulation module establishes a digital twin simulation model based on the equipment design parameters. The cleaned data is injected into the model to simulate the equipment operation and obtain the equipment health baseline value. Establishing a digital twin simulation model based on the equipment design parameters accurately reflects the physical characteristics and operating patterns of the equipment. Injecting the cleaned data into the model to simulate the equipment operation can obtain a more realistic equipment health baseline value, providing an accurate reference for anomaly identification and fault diagnosis. S33: Anomaly Identification: The anomaly identification module calculates the deviation between real-time data and simulated forecast data, uses sliding mean analysis to identify short-term fluctuations, and detects periodic anomalies through wavelet transform. By calculating the deviation between real-time data and simulated forecast data, combining sliding mean analysis to identify short-term fluctuations and wavelet transform to detect periodic anomalies, it can comprehensively analyze data from different angles and improve the accuracy and sensitivity of anomaly detection. Sliding mean analysis can effectively smooth data and capture long-term trend changes in data, while wavelet transform excels at detecting periodic anomalies and local mutations in data. S34: Fault Diagnosis. The fault diagnosis module extracts the statistical and time-frequency characteristics of abnormal segments to form a high-dimensional feature vector. This module then performs fault location and root cause analysis, generating an interpretable report. The extraction of statistical and time-frequency characteristics of abnormal segments forms a high-dimensional feature vector that comprehensively reflects the characteristic information of the fault. Analysis and processing of this high-dimensional feature vector allows for more accurate location and root cause identification of the fault, providing clear guidance for troubleshooting.

[0015] Preferably, the cloud processing layer dynamically updates the digital twin simulation model by: S41: Digital twin simulation model construction, based on the equipment design parameters, uses parametric modeling tools to construct a three-dimensional geometric model of the new energy power generation equipment, and generates a finite element mesh through a meshing algorithm; Based on the equipment design parameters, a parametric modeling tool is used to construct a three-dimensional geometric model of the new energy power generation equipment, which can accurately present the physical structure and form of the equipment. This provides an accurate geometric foundation for subsequent analysis and simulation, making the simulation results closer to the actual operation of the equipment, and generates a finite element mesh through a meshing algorithm, discretizing the continuous physical model into a finite number of units, which is convenient for numerical calculation and analysis. Reasonable meshing can improve calculation accuracy and efficiency, and ensure the reliability of simulation results; S42: Multi-physics field embedding, including mechanical fields, fluid fields, and thermoelectric coupling. The mechanical field uses finite element analysis to solve stress-strain distribution. The fluid field uses computational fluid dynamics to simulate the aerodynamic loads of wind speed and irradiance on renewable energy power generation equipment. Thermoelectric coupling establishes a correlation model between temperature and power generation efficiency for photovoltaic modules. Multi-physics field embedding considers the various physical phenomena and interactions during equipment operation, making the simulation results closer to the actual operating status of the equipment. Compared with single-physics field simulation, multi-physics field coupled simulation can provide more comprehensive and accurate information, providing stronger support for equipment health management and performance optimization. S43: Real-time data-driven and dynamic updates: Data transmitted from the edge computing layer is injected into the model to dynamically update boundary conditions and load status. Models driven by real-time data can more accurately predict the future operating status of equipment, providing timely warnings and decision support for operations and maintenance personnel. For example, when the model detects abnormal changes in certain equipment parameters, it can issue an early alarm, allowing operations and maintenance personnel to take timely measures based on the warning information to avoid equipment failure. S44: Joint simulation and prediction: Using digital twin simulation models to simulate and predict the operating status of renewable energy power generation equipment. Using digital twin simulation models to simulate and predict the operating status of renewable energy power generation equipment can help operation and maintenance personnel understand the performance of equipment under different operating conditions. By simulating and comparing multiple operating strategies, the optimal operating strategy can be selected to improve the power generation efficiency and energy utilization of the equipment. S45: Dynamic health benchmark generation: Monte Carlo simulation is used to generate a probability distribution of the device's health status. Combined with confidence assessment, dynamic benchmark values ​​are output. These output dynamic benchmark values ​​include vibration amplitude thresholds, temperature drift tolerances, and power output parameters. These dynamic benchmark values ​​provide a scientific basis for device health assessment, enabling more accurate judgment of whether the device is operating normally.

[0016] Preferably, the fault diagnosis module extracts the statistical features and time-frequency features of the abnormal segment to form a high-dimensional feature vector, performs fault location and root cause analysis, and generates an interpretable report, specifically including: S51: Feature vector construction: extract the statistical and time-frequency features of the abnormal fragment to form a high-dimensional feature vector. Extracting the statistical and time-frequency features of the abnormal fragment can analyze the fault signal from different angles. Statistical features such as mean, variance, and kurtosis can reflect the overall distribution and fluctuation of the signal; time-frequency features such as wavelet coefficients and short-time Fourier transform coefficients can capture the changes in the signal in time and frequency. Combining these features into a high-dimensional feature vector can more comprehensively characterize the fault characteristics and provide rich information for subsequent diagnosis. S52: Embedding physical knowledge into the fault diagnosis process, combined with the equipment mechanism model, constrains the scope of diagnosis. The equipment mechanism model reflects the physical characteristics and operating laws of the equipment. By matching and analyzing fault characteristics with the mechanism model, some fault hypotheses that do not conform to physical laws can be eliminated, thereby narrowing the scope of fault diagnosis and improving diagnostic efficiency. S53: Large-model semantic analysis uses pre-trained models to parse operation and maintenance logs and historical work order texts, correlating them with fault modes. Using pre-trained models to parse operation and maintenance logs and historical work order texts can mine valuable information from these unstructured text data. Operation and maintenance logs and historical work orders record past equipment operation conditions and troubleshooting experience. Through semantic analysis, key information such as failure modes, causes, and solutions can be extracted, providing a reference for current fault diagnosis. S54: Multimodal fusion reasoning uses an attention mechanism to dynamically weight different modal features and graph neural networks for cross-modal relationship reasoning. Using the attention mechanism to dynamically weight different modal features allows for dynamic adjustments based on the importance of different modal features in fault diagnosis. Fault diagnosis of new energy power generation equipment involves data from multiple modalities, such as vibration signals, temperature signals, and image data. The attention mechanism can automatically learn the importance of different modal features, allowing important features to play a greater role in diagnosis, thereby comprehensively utilizing multi-source information and improving diagnostic accuracy. S55: Fault Location and Root Cause Analysis uses a Bayesian network to calculate fault distribution probabilities and generate an interpretable report. This Bayesian network-based fault distribution probability calculation comprehensively considers the causal relationships and uncertainties among various fault factors, accurately calculating the probability distribution of different fault locations and causes. By selecting the fault location and cause with the highest probability, accurate fault location and root cause analysis can be achieved, providing clear repair guidance for operations and maintenance personnel.

[0017] Preferably, when calculating the fault distribution probability based on the Bayesian network and generating an interpretability report, the following steps are specifically included: S61: Probabilistic output distribution. Input the feature vector into the Bayesian network to output the failure probability of each component, and adjust the fault warning threshold based on the equipment aging coefficient. Inputting the feature vector into the Bayesian network to output the failure probability of each component can quantitatively assess the probability of failure of each equipment component. Bayesian networks are based on probabilistic reasoning and comprehensively consider the causal relationship and uncertainty between various factors, making the calculation results of the failure probability more accurate and reliable. Operation and maintenance personnel can rank the risks of different equipment components according to the size of the failure probability, giving priority to components with a higher failure probability to improve operation and maintenance efficiency. The fault warning threshold is also adjusted based on the equipment aging coefficient to make the warning mechanism more consistent with the actual operating conditions of the equipment. As the equipment is used for a longer time, the probability of failure increases due to equipment aging. By introducing the equipment aging coefficient, the warning threshold can be dynamically adjusted to make the warning more timely and accurate. When the equipment is severely aged, the warning threshold can be appropriately lowered to detect potential fault hazards earlier; when the equipment is newer, the warning threshold can be appropriately increased to avoid false alarms. S62: Visualize fault propagation paths, generate multi-level fault chains, and annotate key node confidence levels. This feature generates multi-level fault chains and annotates key node confidence levels, visually presenting the fault propagation path within the equipment. Operations and maintenance personnel can use this visualization to clearly understand how a fault propagates from one component to another, as well as the likelihood of a fault occurring at each key node. This helps operators gain a deeper understanding of the fault's mechanism, providing more comprehensive information for fault diagnosis and repair. S63: Interpretable report generation: Combined with the large model, a natural language report is generated. This report includes a possible fault cause analysis, recommended fault inspection solutions, and recommended fault repair solutions. This report, combined with the large model, provides comprehensive fault information to operators. The possible fault cause analysis helps operators understand the various possible factors and probabilities that lead to a fault, providing a reference for fault location. The recommended fault inspection solutions guide operators in conducting targeted equipment inspections, improving inspection efficiency. The recommended repair solutions provide operators with specific repair methods and procedures, ensuring smooth maintenance.

[0018] Preferably, the fault diagnosis module is further configured to perform iterative optimization based on actual inspection work orders and actual repair work orders. The specific process is as follows: S71: Abnormal event triggering. After manual inspection is completed, the predicted fault type is compared with the actual fault type. If a discrepancy is found, an abnormal event is triggered. This method can promptly detect deviations between the system diagnosis results and the actual situation, providing clear trigger conditions for subsequent optimization. By promptly detecting errors, problems such as maintenance delays and resource waste caused by misdiagnosis can be avoided, thereby improving operation and maintenance efficiency. S72: Difference type classification. If the actual fault type confirmed by the operator is inconsistent with the system diagnosis, it will be marked as a diagnostic deviation and the cause of the deviation will be recorded. If a fault mode that is not in the Bayesian network is found, it will be marked as a new fault type, triggering the knowledge base expansion process. The cases where the actual fault type confirmed by the operator is inconsistent with the system diagnosis will be marked as diagnostic deviations and new fault types. This classification method can accurately locate the type of problem and provide targeted guidance for subsequent optimization work. For diagnostic deviations, the cause of the deviation can be deeply analyzed and the diagnostic algorithm can be optimized; for new fault types, the knowledge base can be expanded in a timely manner to improve the adaptability and coverage of the system. S73: Feature vector reconstruction: Extract multidimensional features from deviation cases and create a diagnostic deviation feature matrix that compares the original and corrected features of the deviation cases. This method allows for in-depth analysis of the characteristic factors that lead to diagnostic deviations and allows for reconstruction and optimization of feature vectors. The optimized feature vectors more accurately reflect the essential characteristics of the fault, improving the accuracy of fault diagnosis. S74: Dynamic expansion of the knowledge graph. For newly added fault types, the digital twin simulation model is used to simulate the status data of the corresponding fault type, and a new entity relationship is constructed, including the definition of fault node attributes and associated environment variables. For newly added fault types, the digital twin simulation model is used to simulate the status data of the corresponding fault type, and a new entity relationship is constructed, including the definition of fault node attributes and associated environment variables. This method can quickly incorporate relevant knowledge of the newly added fault types into the knowledge graph and enhance the system's knowledge reserve. The dynamic expansion of the knowledge graph enables the system to continuously adapt to new fault conditions, improve the comprehensiveness and accuracy of fault diagnosis, and construct new entity relationships and define fault node attributes and associated environment variables, which helps support more complex fault analysis. By establishing the association between faults and various environmental variables and equipment parameters, we can have a deeper understanding of the fault occurrence mechanism and influencing factors, and provide more comprehensive information for fault diagnosis and maintenance; S75: Model iterative optimization uses Gibbs sampling to dynamically adjust Bayesian network node parameters and evaluates the rationality of the network topology through scoring function structure learning. This approach dynamically optimizes the Bayesian network model based on actual data and feedback, improving the model's diagnostic accuracy. By continuously adjusting node parameters and optimizing the network topology, the model can better fit actual fault data and reduce diagnostic errors.

[0019] Beneficial effects of the present invention: 1. Compared to existing technologies that use static historical data or single device parameters as health detection benchmarks, which have the disadvantage of not reflecting dynamic environmental changes and the impact of equipment aging, this solution uses multi-dimensional simulation modeling technology to integrate device parameters, weather parameters, and other real-time operating condition data to build a digital twin model. By simulating the operating status of equipment in different scenarios (such as photovoltaic efficiency degradation under extreme temperatures and wind turbine aerodynamic loads in salt fog environments), it generates dynamic health benchmark values ​​including power generation, instantaneous current / voltage, and equipment temperature. This method can accurately capture the interactive effects of the environment and equipment, allowing the health detection threshold to be dynamically adjusted according to the operating conditions, and the accuracy of anomaly identification is improved by more than 35% compared to traditional methods. 2. Compared with existing technologies that rely on a single algorithm model for fault diagnosis, which has the disadvantages of high misjudgment rate and delayed identification of new faults, this solution introduces a manual maintenance work order feedback mechanism. After abnormal data is discovered, the fault knowledge base is dynamically optimized by comparing the results of manual on-site verification with the model diagnosis results. If the actual fault type is manually confirmed to be inconsistent with the model diagnosis, the fault label is corrected and distinguishing features are marked. If a new fault is discovered, samples are collected and injected into the model for incremental training. Through this closed-loop iteration, the model can adapt to equipment aging and environmental changes, reducing the false alarm rate by 40% and shortening the new fault identification cycle from an average of 6 months to 2 weeks. 3. Compared to existing technologies that upload all data to the cloud for processing, which suffers from high bandwidth usage and poor real-time performance, this solution deploys lightweight statistical models (such as the 3σ rule and sliding mean analysis) at the edge nodes, extracting features from only abnormal fragments (such as vibration spectrum kurtosis and current fluctuation variance) and compressing and uploading them. This strategy reduces data upload volume by 70% and significantly improves fault response speed through local real-time processing (such as triggering overcurrent protection instructions within 10ms). Under extreme operating conditions, system latency is reduced from 5 seconds in traditional solutions to less than 0.5 seconds. 4. Compared to existing technologies that rely on fixed-cycle calibration or directly use single sensor data, which suffer from high sensor drift false alarm rates and poor environmental adaptability, this solution adopts a primary-secondary redundant architecture for core parameters (such as current and voltage), using DS evidence theory to automatically switch faulty sensors. For non-critical parameters (such as temperature and pressure), an environmental correlation matrix is ​​constructed, and reinforcement learning is used to dynamically update reference value weights (for example, the compensation coefficient increases by 0.15 for every 10% increase in humidity). This mechanism reduces sensor data error by 60% and maintains data confidence levels exceeding 95% even in harsh environments such as salt spray and dust. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Shown is a schematic diagram of the structure of the large-scale model-based new energy power generation equipment health management platform of the present invention; Figure 2 Shown is a schematic diagram of the workflow of the cloud processing layer in the large-scale model-based new energy power generation equipment health management platform of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below with reference to the accompanying drawings and examples.

[0022] See also Figure 1-Figure 2 The present invention provides an embodiment: a large-scale model-based new energy power generation equipment health management platform, comprising: Data collection and perception layer: Used to collect real-time operational data from multiple sensors and IoT devices, and combined with external data including meteorological parameters and power grid status; The data acquisition and perception layer also includes an intelligent calibration module, which is used to perform intelligent calibration on the sensors within the data acquisition and perception layer. The intelligent calibration module adopts a master-slave redundant architecture for the core parameter sensors: specifically, data synchronization acquisition, where the master sensor acquires data in real time, and the backup sensor synchronously samples at a fixed period in a dormant state, and synchronizes the data between the master and backup sensors through timestamps; DS evidence theory decision-making, first calculates the data support of the master and backup sensors, and generates a joint probability distribution based on the confidence matrix to determine whether the data of the master and backup sensors meet the consistency threshold. The principle formula for calculating the data support of the master and backup sensors is: ;in, Indicates the data support of the main sensor and backup sensor, is the measured value of the main sensor, is the measurement value of the backup sensor, and The sensor's range, i.e., the maximum and minimum values ​​the sensor can measure, is used. The offset judgment and switching strategy involves triggering an automatic switching logic if the offsets of the primary and backup sensors exceed the threshold three times in a row. The backup sensor then takes over control and issues a fault warning. After the switchover, the original primary sensor enters calibration mode, verifying its drift characteristics using a standard source input. If the sensor cannot recover, it is marked as faulty and awaiting replacement. The standard source input is reference value data from a healthy benchmark library. For non-critical parameter sensors, dynamic reference value weights are used for error compensation, specifically including: construction of an environmental parameter association matrix, determination of non-critical parameters and their sensitive environmental variables, establishment of a compensation coefficient matrix through historical data analysis or physical model derivation; dynamic adjustment of reference values, discretization of environmental variables into state space, and action space as the reference value adjustment amplitude, and the use of a decaying learning rate strategy, and then inputting the parameter association matrix, the deviation between the current environmental parameters and the threshold, and calculating the adjustment vector group; calibration verification and update, introducing a sliding window mechanism, calculating the error variance after the last N adjustments, and if the variance continues to increase, triggering an exploration rate reset, forcing the algorithm to re-explore the environment, and writing the optimized reference value into the digital twin model to regenerate the health benchmark library. Edge computing layer: Deploy a lightweight anomaly detection model to filter the data collected by the data acquisition and perception layer and extract key features. Specifically, it includes: a fault processing module that uses a baseline statistical model to filter abnormal fragments and extract statistical features, performs fault detection based on statistical features, and performs real-time fault processing based on the built-in fault strategy library. The statistical methods used in the baseline statistical model include the 3σ rule and sliding mean analysis; a data screening module that filters the data collected by the data acquisition and perception layer based on the data results of the fault identification model and preset rules, retaining abnormal data and key feature data and uploading it to the cloud processing layer; a scheduling optimization module that establishes a computing node health assessment model and dynamically allocates tasks based on CPU and memory load; Cloud processing layer: It is used to build a digital twin simulation model, combine external data to predict the working status of new energy power generation equipment, generate dynamic health benchmark values, and perform health diagnosis and management through a large model. Specifically, it includes: a simulation module, which is used to build a digital twin simulation model, simulate the new energy power generation equipment based on real-time environmental data and external data, and obtain the predicted working status data of the new energy power generation equipment; an anomaly recognition module, which is used to identify anomalies based on real-time data and predicted working status data; a fault diagnosis module, which is used to diagnose faults of new energy power generation equipment based on the data results of the anomaly recognition module; The application service layer is used to provide equipment health assessment, operation and maintenance work order management, visualization services, and decision support based on the data results of the cloud processing layer.

[0023] Among them, the cloud processing layer specifically includes: Step 1: Multi-source data fusion and processing. First, the filtered abnormal data, key feature data, and external environment parameters are received from the edge computing layer through a distributed message queue. The data is then cleaned and aligned. Finally, the heterogeneous data is converted into a standardized format that can be parsed by the digital twin model. Step 2: Dynamically update the digital twin simulation model. Use the simulation module to build a digital twin simulation model based on the equipment design parameters, inject the cleaned data into the model, simulate the equipment operation, and obtain the equipment health baseline value. Step 3: Anomaly identification: The anomaly identification module calculates the deviation between real-time data and simulation prediction data, uses sliding mean analysis to identify short-term fluctuations, and detects periodic anomalies through wavelet transform; Step 4: Fault diagnosis: Use the fault diagnosis module to extract the statistical features and time-frequency features of the abnormal segment to form a high-dimensional feature vector, perform fault location and root cause analysis, and generate an interpretable report.

[0024] Among them, the cloud processing layer includes the following when dynamically updating the digital twin simulation model: Step 1: Digital twin simulation model construction: Based on the equipment design parameters, parametric modeling tools are used to construct a 3D geometric model of the new energy power generation equipment, and a finite element mesh is generated through a meshing algorithm; Step 2: Multi-physics field embedding, including mechanical field, fluid field, and thermoelectric coupling. The mechanical field uses finite element analysis to solve the stress-strain distribution. The fluid field uses computational fluid dynamics to simulate the aerodynamic loads of wind speed and irradiance on renewable energy power generation equipment. The thermoelectric coupling establishes a correlation model between temperature and power generation efficiency for photovoltaic modules. Step 3: Real-time data drive and dynamic update: inject data transmitted by the edge computing layer into the model to dynamically update boundary conditions and load status; Step 4: Combined simulation and prediction: Use the digital twin simulation model to simulate and predict the working status of new energy power generation equipment; Step 5: Generate dynamic health benchmarks. Generate a probability distribution of the equipment health status through Monte Carlo simulation, and output dynamic benchmark values ​​based on confidence assessment. The output dynamic benchmark values ​​include vibration amplitude thresholds, temperature drift tolerances, and power output parameters.

[0025] The fault diagnosis module extracts the statistical and time-frequency features of abnormal segments to form a high-dimensional feature vector, performs fault location and root cause analysis, and generates an interpretable report. Specifically, it includes: Step 1: Feature vector construction, extracting the statistical features and time-frequency features of the abnormal segment to form a high-dimensional feature vector; Step 2: Embed physical knowledge, combine it with the equipment mechanism model, and constrain the diagnosis scope; Step 3: Large-model semantic analysis: Use pre-trained models to parse operation and maintenance logs and historical work order texts to correlate failure modes. Step 4: Multimodal fusion reasoning, using the attention mechanism to dynamically weight different modal features and using graph neural networks to reason about cross-modal relationships; Step 5: Fault location and root cause analysis. Calculate the fault distribution probability based on the Bayesian network and generate an interpretable report. This includes: Probabilistic output distribution: input the feature vector into the Bayesian network, output the failure probability of each component, and adjust the fault warning threshold according to the equipment aging coefficient; visualize the fault propagation path, generate a multi-level fault chain, and mark the confidence of key nodes; generate an interpretable report, combine the large model to generate a natural language report, among which the generated natural language report includes a possible analysis of the cause of the fault, recommended fault inspection plan, and recommended fault repair plan.

[0026] The fault diagnosis module is also used to perform iterative optimization based on actual inspection work orders and actual repair work orders. The specific process is as follows: Step 1: Abnormal event triggering. After manual maintenance is completed, the predicted fault type is compared with the actual fault type. If there is a difference, an abnormal event is triggered. Step 2: Classify the difference type. If the actual fault type confirmed by manual verification does not match the system diagnosis, it is marked as a diagnostic deviation and the reason for the deviation is recorded. If a fault mode not included in the Bayesian network is found, it is marked as a new fault type, triggering the knowledge base expansion process. Step 3: Feature vector reconstruction, extracting multidimensional features of the deviation cases and establishing a diagnostic deviation feature matrix that contains the original features of the deviation cases and the comparison of the corrected features; Step 4: Dynamically expand the knowledge graph. For newly added fault types, use the digital twin simulation model to simulate the state data of the corresponding fault type and build new entity relationships, including defining fault node attributes and associated environment variables. Step 5: Iteratively optimize the model, dynamically adjust the Bayesian network node parameters using Gibbs sampling, and evaluate the rationality of the network topology through scoring structure learning.

[0027] Example 1: Intelligent health management of wind turbines Scenario: Offshore wind farms, high salt fog, and high humidity environments, where wind turbine gearboxes and blades are susceptible to corrosion.

[0028] Specific implementation process: Data collection and perception layer: Active / standby redundancy calibration: The gearbox vibration sensor (range 0-200 m / s²) adopts an active / standby architecture. The active sensor monitors in real time, while the standby sensor samples synchronously every 5 minutes (with aligned timestamps). DS evidence theory decision-making: Support is calculated. If the support is less than 0.95 three times in a row, a switchover to the standby sensor is triggered, and a salt spray corrosion alarm is issued. Dynamic reference value compensation: Humidity sensor errors are compensated using Q-learning. When humidity exceeds 90%, the compensation coefficient is increased from 0.2 to 0.35.

[0029] Edge computing layer: Troubleshooting: Sliding mean analysis is used to detect sudden vibration increases (mean value increases by 40% within 3 seconds) and determine bearing abnormalities. Data screening: Only abnormal vibration fragments (frequency 8-12kHz) and environmental parameters are uploaded. Scheduling optimization: Vibration signals are prioritized during high loads, reducing CPU usage from 90% to 65%.

[0030] Cloud processing layer: Digital twin modeling: A 3D model of the wind turbine blades was constructed, embedded with CFD to simulate the impact of salt spray deposition on aerodynamic efficiency, achieving a predicted power curve with a ±3% deviation. A dynamic health benchmark was established: Monte Carlo simulation generated vibration thresholds (normally 5-15 m / s², reduced to 8 m / s² under salt spray corrosion). Fault diagnosis: Vibration spectrum features were extracted and combined with "abnormal noise" text from maintenance logs. A Bayesian network output indicated an 82% probability of gear wear. Model optimization: Manual inspection revealed insufficient lubrication. Gibbs sampling was used to adjust the Bayesian network parameters, reducing the false alarm rate for similar faults from 18% to 7%.

[0031] Application service layer: An AR work order was generated, noting the bearing locations requiring lubrication and recommending a historically successful solution (92% match). A visual large screen displayed the wind turbine health score (reduced from 78 to 65), triggering a maintenance alert.

[0032] Technical effect: The fault warning accuracy rate in salt spray environment is increased to 95%, and the operation and maintenance costs are reduced by 30%.

[0033] Example 2: Hot spot monitoring of photovoltaic power station components Scenario: Desert photovoltaic power stations, with high temperatures and dusty environments, are prone to hot spots on components.

[0034] Specific implementation process: Data collection and perception layer: Non-critical parameter calibration: The backplane temperature sensor is dynamically compensated using the environmental correlation matrix (temperature-irradiance). During sandstorms, the compensation coefficient is adjusted from 0.3 to 0.45. Redundancy check: The active / standby support threshold of the current sensor is set to 0.98, and a switchover occurs if the sensor exceeds the tolerance for two consecutive times.

[0035] Edge computing layer: Anomaly Detection: 3σ rule is used to identify modules with abnormal current flow (daily power generation drop > 30%). Data Filtering: Only infrared image features (temperature > 80°C) and irradiance data of modules with hot spots are uploaded.

[0036] Cloud processing layer: Digital twin simulation: Builds a thermoelectric coupling model of photovoltaic modules to predict hot spot temperature distribution (normal 65°C → extreme 85°C). Dynamic benchmark: Monte Carlo simulation generates temperature tolerances (±5°C threshold with 95% confidence). Multimodal diagnosis: Fusion of infrared image spectra with "local blackening" text in maintenance logs locates a 79% probability of a cold solder joint failure. Knowledge base expansion: Adds diode breakdown failure mode, and GAN generates 500 sets of simulated data for injection into the training set.

[0037] Application service layer: Send hot spot cleaning orders, and the drone will automatically plan the cleaning path. A visual report displays the component efficiency degradation curve and recommends replacing components with poor solder joints.

[0038] Technical effect: The hot spot misjudgment rate is reduced by 40% and the component life is extended by 20%.

[0039] Example 3: Energy Storage Battery Pack Health Management Scenario: Grid-side energy storage power station, where high charging and discharging frequencies lead to battery aging.

[0040] Specific implementation process: Data collection and perception layer: Core parameter calibration: Voltage sensors are redundant (range 0-1000V), with switching enabled when the support level is <0.98 and the difference is >10V. Dynamic compensation: The temperature compensation coefficient for internal resistance is dynamically adjusted through reinforcement learning, with the weight increasing from 0.2 to 0.5 at low temperatures (-20°C).

[0041] Edge computing layer: Fault handling: Extract the second-order derivative characteristics of the charge and discharge curves to identify sudden changes in internal resistance (change rate > 15%). Scheduling optimization: Prioritize balancing instructions to high-load battery clusters.

[0042] Cloud processing layer: Digital twin modeling: The electrochemical-thermal coupling model simulates the risk of lithium plating, and the dynamic reference temperature is set to 45°C.

[0043] Bayesian diagnosis: Combining voltage drop characteristics with the "bulge" text in maintenance records, the output shows a 62% probability of lithium deposition. Parameter optimization: Gibbs sampling adjusts the weights of temperature nodes, increasing the detection rate of low-temperature capacity drop faults from 68% to 91%.

[0044] Application service layer: This triggers a 20% power increase in the liquid cooling system and generates a battery replacement list (for cells with a health score < 60). A visual interface displays the battery cluster aging distribution, highlighting high-risk cells.

[0045] Technical effect: The battery thermal runaway warning rate is increased by 60% and the cycle life is extended by 25%.

[0046] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A health management platform for new energy power generation equipment based on a large model; its characteristics are: include: The data collection and perception layer is used to collect real-time operational data through multi-source sensors and IoT devices, and combine it with external data including meteorological parameters and power grid status; The edge computing layer deploys a lightweight anomaly detection model to filter the data collected by the data acquisition and perception layer and extract key features; The cloud processing layer is used to build a digital twin simulation model, combine external data to predict the operating status of new energy power generation equipment, generate dynamic health benchmark values, and perform health diagnosis and management through large models; The application service layer is used to provide equipment health assessment, operation and maintenance work order management, visualization services, and decision support based on the data results of the cloud processing layer.

2. The large-scale model-based new energy power generation equipment health management platform according to claim 1 is characterized by: The data acquisition and perception layer also includes an intelligent calibration module, which is used to intelligently calibrate the sensors within the data acquisition and perception layer. The intelligent calibration module adopts a master-slave redundant architecture for core parameter sensors, specifically including: S11: Synchronous data collection: the main sensor collects data in real time, and the backup sensor samples synchronously at a fixed period in the dormant state. The data synchronization between the main sensor and the backup sensor is achieved through timestamps. S12: DS evidence theory decision-making, first calculates the data support of the primary sensor and the backup sensor, and generates a joint probability distribution based on the confidence matrix to determine whether the data of the primary sensor and the backup sensor meet the consistency threshold; S13: Offset judgment and switching strategy: If the offset of the main sensor and the backup sensor exceeds the threshold for three consecutive times and the automatic switching logic is triggered, the backup sensor takes over the main control and sends a fault warning; S14: After the switch, the original main sensor enters the calibration mode and verifies its drift characteristics through the standard source input. If it cannot be recovered, it is marked as faulty and needs to be replaced. The standard source input is the reference value data in the healthy benchmark library.

3. The large-scale model-based new energy power generation equipment health management platform according to claim 2 is characterized by: When the intelligent calibration module performs intelligent calibration on sensors in the data acquisition and perception layer, it uses dynamic reference value weights to compensate for errors of non-critical parameter sensors, specifically including: S21: Construct an environmental parameter correlation matrix, determine non-critical parameters and their sensitive environmental variables, and establish a compensation coefficient matrix through historical data analysis or physical model derivation; S22: Dynamic adjustment of reference values. The environmental variables are discretized into a state space, and the action space is the reference value adjustment amplitude. A decaying learning rate strategy is adopted. Then, the parameter association matrix and the deviation between the current environmental parameters and the threshold are input to calculate the adjustment vector group. S23: Calibration verification and update, introduce a sliding window mechanism to calculate the error variance after the last N adjustments. If the variance continues to increase, it triggers the exploration rate reset, forcing the algorithm to re-explore the environment, and write the optimized reference value into the digital twin model to regenerate the health benchmark library.

4. The large-scale model-based new energy power generation equipment health management platform according to claim 3 is characterized by: The edge computing layer specifically includes: A11: The fault processing module uses a baseline statistical model to screen abnormal segments and extract statistical features. It then detects faults based on these features and performs real-time fault processing based on a built-in fault strategy library. The baseline statistical model uses statistical methods such as the 3σ rule and sliding mean analysis. A12: Data screening module, used to screen the data collected by the data acquisition and perception layer based on the data results of the fault identification model and preset rules, retaining abnormal data and key feature data and uploading it to the cloud processing layer; A13: Scheduling optimization module, used to establish a computing node health assessment model and dynamically allocate tasks based on CPU and memory load.

5. The large-scale model-based new energy power generation equipment health management platform according to claim 4 is characterized by: The cloud processing layer specifically includes: A21: Simulation module, used to build a digital twin simulation model, simulate new energy power generation equipment based on real-time environmental data and external data, and obtain predicted operating status data of new energy power generation equipment; A22: Anomaly identification module, used to identify anomalies based on real-time data and predicted working status data; A23: Fault diagnosis module, used to diagnose faults of new energy power generation equipment based on the data results of the abnormality identification module.

6. The large-scale model-based health management platform for new energy power generation equipment according to claim 5 is characterized by: When the cloud processing layer is working, it specifically includes: S31: Multi-source data fusion and processing. First, it receives filtered abnormal data, key feature data, and external environmental parameters from the edge computing layer, receives them through a distributed message queue, cleans and aligns the data, and finally converts the heterogeneous data into a standardized format that can be parsed by the digital twin model. S32: Dynamically update the digital twin simulation model. The simulation module establishes a digital twin simulation model based on the equipment design parameters, and injects the cleaned data into the model to simulate the equipment operation and obtain the equipment health baseline value. S33: Anomaly identification, which uses the anomaly identification module to calculate the deviation between real-time data and simulation prediction data, uses sliding mean analysis to identify short-term fluctuations, and uses wavelet transform to detect periodic anomalies; S34: Fault diagnosis: The fault diagnosis module is used to extract the statistical features and time-frequency features of the abnormal fragments to form a high-dimensional feature vector, and then perform fault location and root cause analysis to generate an interpretable report.

7. The large-scale model-based health management platform for new energy power generation equipment according to claim 6 is characterized by: When the cloud processing layer dynamically updates the digital twin simulation model, it specifically includes: S41: Digital twin simulation model construction, based on equipment design parameters, using parametric modeling tools to build a three-dimensional geometric model of the new energy power generation equipment, and generate a finite element mesh through a meshing algorithm; S42: Multi-physics field embedding, including mechanical field, fluid field, and thermoelectric coupling. The mechanical field uses finite element analysis to solve stress-strain distribution. The fluid field uses computational fluid dynamics to simulate the aerodynamic loads of wind speed and irradiance on renewable energy power generation equipment. The thermoelectric coupling establishes a correlation model between temperature and power generation efficiency for photovoltaic modules. S43: Real-time data-driven and dynamic update: injecting data transmitted by the edge computing layer into the model to dynamically update boundary conditions and load status; S44: Joint simulation and prediction, using digital twin simulation models to simulate and predict the working status of new energy power generation equipment; S45: Dynamic health benchmark generation: Generate the probability distribution of the equipment health status through Monte Carlo simulation, and output the dynamic benchmark value in combination with the confidence assessment. The output dynamic benchmark value includes the vibration amplitude threshold, temperature drift tolerance and power output parameters.

8. The large-scale model-based new energy power generation equipment health management platform according to claim 7 is characterized by: The fault diagnosis module extracts the statistical and time-frequency features of abnormal segments to form a high-dimensional feature vector, performs fault location and root cause analysis, and generates an interpretable report. Specifically, it includes: S51: Feature vector construction, extracting statistical features and time-frequency features of abnormal segments to form a high-dimensional feature vector; S52: Physical knowledge embedding, combined with equipment mechanism model, constrains the diagnosis scope; S53: Large-model semantic analysis, using pre-trained models to parse operation and maintenance logs and historical work order texts and associate failure modes; S54: Multimodal fusion reasoning, using attention mechanism to dynamically weight different modal features and using graph neural network to reason about cross-modal relationships; S55: Fault location and root cause analysis, calculate the fault distribution probability based on Bayesian network, and generate an explainable report.

9. The large-scale model-based health management platform for new energy power generation equipment according to claim 8 is characterized by: When calculating the fault distribution probability based on the Bayesian network and generating an explainability report, the following are specifically included: S61: Probabilistic output distribution, input the feature vector into the Bayesian network, output the failure probability of each component, and adjust the fault warning threshold according to the equipment aging coefficient; S62: Visualize the fault propagation path, generate multi-level fault chains, and annotate the confidence of key nodes; S63: Generate an explainable report, combining the large model to generate a natural language report, wherein the generated natural language report includes a possible analysis of the cause of the fault, a recommended fault inspection plan, and a recommended fault repair plan.

10. The large-scale model-based health management platform for new energy power generation equipment according to claim 9 is characterized in that: The fault diagnosis module is also used to perform iterative optimization based on actual inspection work orders and actual repair work orders. The specific process is as follows: S71: abnormal event triggering. After the manual maintenance is completed, the predicted fault type is compared with the actual fault type. If there is a difference, an abnormal event is triggered. S72: Difference type classification. If the actual fault type confirmed by manual verification is inconsistent with the system diagnosis, it is marked as a diagnostic deviation and the reason for the deviation is recorded. If a fault mode not included in the Bayesian network is found, it is marked as a new fault type, triggering the knowledge base expansion process. S73: Feature vector reconstruction, extracting multidimensional features for the deviation cases, and establishing a diagnostic deviation feature matrix containing the original features of the deviation cases and the comparison of the corrected features; S74: Dynamic expansion of the knowledge graph. For newly added fault types, the digital twin simulation model is used to simulate the state data of the corresponding fault type and construct new entity relationships, including defining fault node attributes and associated environment variables. S75: Iterative model optimization uses Gibbs sampling to dynamically adjust the Bayesian network node parameters, and evaluates the rationality of the network topology through scoring row structure learning.

Images