A new energy power station automatic inspection and AI fault diagnosis method and system

Abstract

The application discloses a new energy power station automatic inspection and AI fault diagnosis method and system, and the method steps are as follows: an automatic inspection platform carrying a multi-modal sensing unit is controlled to inspect along a preset path, and various data of power station equipment are synchronously collected; a high-fidelity digital twin body fusing a three-dimensional geometric model, an electrical physical model and historical operation data is constructed, collected data are injected to realize synchronous mapping and state updating with a physical power station; an AI model after training is used to diagnose and classify faults through multi-dimensional feature cross-modal fusion analysis, time-space correlation cross verification of equipment; a time series prediction model analyzes the performance degradation trend of components, calculates a health index and predicts a remaining useful life; a fault position is highlighted in the digital twin body, a fault knowledge graph is generated, VR immersive viewing and tracing are supported; an electronic work order containing fault details is automatically generated and pushed to a computerized maintenance management system. The application realizes an intelligent operation and maintenance closed loop, and improves the reliability and efficiency of the power station.

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance technology for new energy power plants, specifically to an automatic inspection and AI fault diagnosis method and system for new energy power plants. Background Technology

[0002] With the transformation of the global energy structure and the advancement of the "dual carbon" goal, the new energy industry, represented by photovoltaics and wind power, has ushered in explosive growth. However, new energy power plants, especially large-scale centralized photovoltaic power plants and wind farms, are usually located in remote areas, harsh environments, and scattered layouts. Their large number of equipment and vast land area have brought unprecedented severe challenges to the traditional operation and maintenance management model.

[0003] Currently, the operation and maintenance of new energy power plants heavily relies on traditional methods, which have the following limitations: First, manual periodic inspections are inefficient, costly, and unsafe. Maintenance personnel must travel long distances and inspect equipment in harsh environments using only their eyes and experience, resulting in high workload, long cycles, and difficulty in detecting potential faults such as internal cracks in photovoltaic modules, early hot spots, and slight performance degradation of inverter components. Second, existing automation technologies are siloed and lack deep integration. Although some advanced power plants have introduced drones and robots for automated inspections and used infrared thermal imagers to collect data, their functions are mostly limited to single data collection and preliminary identification. Inspection data, SCADA system operation data, environmental monitoring data, and other multi-source information are independent, lacking effective spatiotemporal correlation and fusion analysis, making it impossible to accurately locate and analyze the root causes of complex faults. Third, mainstream diagnostic methods are "post-event response," failing to achieve predictive maintenance. Existing systems typically only alarm after equipment failure and shutdown, forcing maintenance personnel to perform reactive repairs, resulting in significant power generation losses and maintenance costs. While some solutions attempt to utilize AI for image recognition, they are mostly limited to judging the current state and lack the ability to analyze the timing trends of equipment performance degradation and predict remaining useful life (RUL), failing to elevate maintenance strategies from prevention to prediction. Finally, the closed loop from diagnosis to execution is broken. Even if a fault is detected, processes such as fault location, report generation, work order dispatch, and maintenance review still require significant manual intervention, resulting in slow response and failing to form an efficient automated closed loop.

[0004] In summary, existing technologies lack a systematic solution that integrates multimodal data fusion, AI-powered intelligent diagnostics, digital twin mapping, predictive maintenance, and automated operation and maintenance (O&M) loops. Therefore, there is an urgent need for an innovative method and system that can break down data silos, achieve holographic perception of the status of new energy power plants, accurate fault diagnosis and forward-looking prediction, and ultimately automatically drive O&M execution, thereby significantly improving the operational reliability, power generation efficiency, and economic benefits of the power plant. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an automatic inspection and AI fault diagnosis method and system for new energy power plants, which solves the pain points of traditional operation and maintenance. By automatically inspecting and collecting multimodal data, building a digital twin, using AI for diagnosis and classification, predicting lifespan, visualizing faults and generating work orders, a closed-loop operation and maintenance system is achieved, thereby improving the reliability and efficiency of power plants.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide the following technical solution: an automatic inspection and AI fault diagnosis method for new energy power plants, comprising the following steps: Step S1: Control the automatic inspection platform equipped with multimodal sensing units to carry out inspections of the new energy power station along a preset path, and simultaneously collect the spatial coordinates of the power station equipment, multimodal sensing data, real-time operation data of the power station, and environmental monitoring data, and assign them a unified timestamp and spatial coordinate label. Step S2: Construct a high-fidelity digital twin of the new energy power station that integrates a three-dimensional geometric model, an electrical and physical model, and historical operating data; inject the collected tagged multi-dimensional data into the digital twin in real time to drive the digital twin to achieve synchronous mapping and status update with the physical power station; Step S3: Based on the multi-dimensional data of the injected digital twin, and using the trained artificial intelligence (AI) model, perform cross-modal fusion analysis and spatiotemporal correlation cross-validation on the image features, acoustic features, operating parameters and environmental parameters of the same device object, and perform fault diagnosis and classification of the device. Step S4: Based on the historical and real-time status data sequences accumulated in the digital twin, analyze the degradation trend of power plant component performance parameters using a time-series prediction model, calculate the health index of the power plant component, and predict its remaining useful life. Step S5: When a fault is diagnosed or a risk is predicted, the three-dimensional location of the faulty device is highlighted in the virtual environment of the digital twin, and a fault knowledge graph is generated; the digital twin is accessed through a virtual reality interactive device to immerse the user in viewing the multi-dimensional historical data of the faulty device, the status of related devices, and the fault root cause chain. Step S6: Based on the diagnostic results of step S3 or the prediction results of step S4, automatically generate an electronic work order containing fault details, location information, maintenance suggestions and priorities, and push the electronic work order to the computerized maintenance management system.

[0007] Furthermore, the construction of a high-fidelity digital twin of a new energy power plant, integrating a three-dimensional geometric model, an electrical-physical model, and historical operating data, specifically involves: Based on the BIM design model, oblique photogrammetry real scene 3D model or CAD drawings of the power station, construct a 3D visualization model that includes the spatial location, geometric shape and topological connection relationship of the power station equipment and facilities. Based on the three-dimensional visualization model, an electrical physical model representing the electrical characteristics of the equipment is embedded; the electrical physical model includes: an equivalent circuit model, a power converter mathematical model, and a line impedance model. Establish a mapping relationship between the virtual objects in the digital twin and the real-time and historical databases of the physical power station; use the collected real-time operation data and environmental data as input to drive the electrical physical model to perform calculations, so that the operation status of the virtual model is synchronized with the physical power station, and continuously store historical state sequence data; By comparing the simulation output data of the digital twin with the actual measurement data of the physical power station, the parameters of the electrical physical model are automatically calibrated and optimized.

[0008] Furthermore, the training method for the aforementioned artificial intelligence (AI) model includes: Annotated multimodal data samples are extracted from the historical database of the digital twin; the multimodal data samples include fault images and acoustic data collected by the multimodal sensing unit and verified by operation and maintenance results, as well as SCADA operation data, environmental data and corresponding fault type labels aligned with spatiotemporal data. The multimodal data samples are standardized, denoised, and augmented; specifically, rotation, cropping, and color adjustment methods are used for image data, and sliding window cutting method is used for time-series data. Construct a deep learning network model suitable for multimodal fusion analysis, which includes at least a convolutional neural network (CNN) branch for image processing, a recurrent neural network (RNN) or Transformer branch for processing temporal data, and an attention mechanism layer for feature fusion; use the training sample set to train the network model end-to-end in a supervised learning manner, and iteratively optimize the model parameters by minimizing the loss function; The trained model is validated using a reserved test sample set until the accuracy and recall metrics reach preset thresholds.

[0009] Furthermore, the cross-modal fusion analysis and spatiotemporal correlation cross-validation of image features, acoustic features, operating parameters, and environmental parameters of the same device object specifically includes: Based on unified timestamps and spatial coordinate labels, image data, acoustic data, operating parameters and environmental parameters within the same time segment and geographic space are bound to the corresponding unique device object ID in the digital twin to form multimodal data sample pairs; The bound multimodal data are then input into the corresponding feature extraction branches, including: Image data is input into a branch of a convolutional neural network (CNN) to extract spatial visual feature vectors. After converting the acoustic data into a spectrogram, it can be input into another CNN branch, or directly into an audio feature extraction network to extract acoustic feature vectors. Input time-series data such as operating parameters and environmental parameters into a recurrent neural network (RNN) or Transformer branch to extract time-series feature vectors. The extracted modal feature vectors are input into a cross-modal attention fusion layer to calculate the correlation weights between different modal features and generate a weighted comprehensive feature vector accordingly. The comprehensive feature vector is input into the final classifier or regressor, which outputs the fault diagnosis result, and then the equipment is diagnosed and classified for faults.

[0010] Furthermore, the fault diagnosis and classification of the equipment specifically includes: The obtained comprehensive feature vector is matched with a predefined fault feature library; the fault feature library stores multimodal feature templates corresponding to different fault modes. Through similarity calculation, a preliminary candidate list containing fault types and their confidence levels is output. Hierarchical fault classification and severity assessment employs a hierarchical classification strategy to classify and assess faults in a refined manner, including the first level of equipment-level classification, the second level of fault mode classification, and the third level of severity grading. The knowledge graph-based reasoning engine is triggered. Based on the device topology connection relationship, historical maintenance records and physical rules, the reasoning engine traverses the fault propagation path and deduces the most likely root cause. Generate structured diagnostic results, which include at least the faulty device ID, fault type, severity level, occurrence time, location information, confidence score, and recommended remedial measures.

[0011] Furthermore, the step of using a time-series prediction model to analyze the degradation trend of key component performance parameters, calculate the health index of key components, and predict their remaining useful life specifically involves: From the historical and real-time status data stored in the digital twin, performance degradation indicators used to characterize the health status of key components are extracted; Based on the performance degradation index, a health index HI ranging from [0,1] is calculated using a weighted fusion or machine learning regression method; where a health index of 1 represents a brand new state and 0 represents a completely failed state. The data sequence of the health index changing over time is input into the time-series prediction model. By learning the degradation trajectory of the health index, the time required for it to reach the preset failure threshold is extrapolated and predicted. This time is the predicted remaining useful life (RUL). The time-series prediction model is a Long Short-Term Memory (LSTM) network, a Transformer, or a ParticleFilter model. Along with the output of the RUL prediction result, the confidence interval or probability distribution of the prediction value is provided to quantify the uncertainty of the prediction.

[0012] Furthermore, the step of highlighting the three-dimensional location of the faulty device in the virtual environment of the digital twin and generating a fault knowledge graph specifically involves: Based on the unique device ID contained in the fault diagnosis results, the corresponding virtual device object is accurately located in the three-dimensional geometric model of the digital twin; by changing the model color, applying pulsed light effects, adding warning icons, or generating bounding boxes, the faulty device is highlighted in the virtual environment and visually contrasted with normal devices. Based on the root cause reasoning results or structured diagnosis results, a knowledge graph centered on the current fault event is dynamically constructed; the nodes of the knowledge graph include at least: faulty equipment entity, fault phenomenon entity, fault mode entity, root cause entity, and affected related equipment entity; the edges between nodes are used to describe the semantic relationships between entities; The generated fault knowledge graph is presented in a graphical form on the user interface, which allows users to interactively click on nodes in the graph to expand and view detailed information, historical fault records, repair solutions, or other related fault cases of that entity.

[0013] Furthermore, the root cause chain of the failure is implemented through the integration of knowledge graph and retrieval enhancement generation RAG technology, specifically as follows: Receive natural language queries from users or diagnostic modules, which describe the currently observed fault phenomena; use generative large models to perform semantic parsing and intent understanding on the queries, and extract key entities and relationships as search conditions; Based on the parsed search criteria, a search is performed in the pre-constructed operation and maintenance knowledge graph; the search traverses the entity nodes and connecting edges related to the faulty equipment and fault mode, and the search scope includes historical similar work orders, expert experience rules, equipment manual chapters, failure mode cases and solutions; The retrieved structured information, text fragments, and case data are combined with the current real-time running data and multimodal feature fusion results to form a contextual information package; this information package is used as the context, together with the instructions optimized by prompt word engineering, to form the input prompts of the large model; Based on the input prompts, the generative large model generates a natural language description that includes the root cause of the failure, the logical chain of reasoning, and suggestions for handling the problem. At the same time, the model output must include the retrieval source information on which its conclusions are based, including the case number, manual entry, or rule ID, so as to make the generated results verifiable and traceable.

[0014] Furthermore, based on the diagnostic results of step S3 or the prediction results of step S4, an electronic work order containing fault details, location information, repair suggestions, and priorities is automatically generated, specifically as follows: Based on the diagnostic results of step S3 or the prediction results of step S4, automatically populate the following fields of the electronic work order: Fault details: including faulty device ID, fault type, severity level, time of occurrence, and reliability score; Location information: including the device's three-dimensional coordinates in the digital twin, its matrix / region number, and the optimal path navigation link to the device; Repair recommendations: Utilize the knowledge graph to retrieve historical successful solutions, required spare parts lists, safe operating procedures, and standard operating procedures (SOPs) that match this failure mode; Priority: The priority of work order processing is dynamically calculated based on the severity level of the fault, the power generation weight of the affected equipment, and the availability of spare parts. The generated structured electronic work orders are automatically pushed to the computerized maintenance management system via the application programming interface (API) for resource scheduling and recording; at the same time, the core information and navigation links of the work orders are pushed to the mobile terminal applications of local maintenance personnel.

[0015] This invention also proposes an automatic inspection and AI fault diagnosis system for new energy power plants, comprising: The multimodal data synchronous acquisition module is used to control the automatic inspection platform equipped with multimodal sensing units to carry out inspections of new energy power plants along a preset path, and synchronously collect the spatial coordinates of power plant equipment, multimodal sensing data, real-time operating data of the power plant monitoring and data acquisition SCADA system, and environmental monitoring data, and assign them a unified timestamp and spatial coordinate label. The digital twin construction and synchronization module is used to construct and maintain a high-fidelity digital twin of a new energy power plant that integrates a three-dimensional geometric model, an electrical and physical model, and historical operating data. It receives tagged multi-dimensional data from the multimodal data synchronization acquisition module and drives the digital twin to achieve synchronous mapping and status updates with the physical power plant. The AI ​​fault diagnosis and classification module is communicatively connected to the digital twin construction and synchronization module. It is used to perform cross-modal fusion analysis and spatiotemporal correlation cross-validation on the image features, acoustic features, operating parameters and environmental parameters of the same device object based on the multi-dimensional data injected into the digital twin and using a trained artificial intelligence AI model, so as to diagnose and classify the device faults. The predictive maintenance analysis module is communicatively connected to the digital twin construction and synchronization module. It is used to analyze the degradation trend of key component performance parameters based on the historical and real-time status data sequences accumulated in the digital twin, calculate the health index of key components, and predict their remaining useful life. The fault visualization and tracing module is communicatively connected to the AI ​​fault diagnosis and classification module and the predictive maintenance analysis module. When a fault is diagnosed or a risk is predicted, it is used to highlight the three-dimensional location of the faulty device in the virtual environment of the digital twin and generate a fault knowledge graph. It is also used to support access through virtual reality interactive devices to immerse users in viewing multi-dimensional historical data of the faulty device, the status of related devices, and the fault root cause chain. The operation and maintenance decision and work order management module is communicatively connected to the AI ​​fault diagnosis and classification module and the predictive maintenance analysis module. It is used to automatically generate electronic work orders containing fault details, location information, maintenance suggestions and priorities based on the diagnosis results or prediction results, and push the electronic work orders to the computerized maintenance management system.

[0016] The beneficial effects of the above-described technical solution of the present invention are as follows: 1. This invention revolutionizes the traditional manual inspection mode by using an automated inspection platform that integrates robots and drones. Equipped with high-precision positioning and multimodal sensing units, it simultaneously collects equipment spatial coordinates, visual / acoustic / environmental data, and SCADA operation data. The data is linked using unified timestamps and spatial coordinate labels, breaking down data silos at the source. A high-fidelity digital twin integrates three-dimensional geometry, electrical and physical models, and historical data to dynamically map the physical state of the power plant, replacing manual visual inspection and significantly improving inspection efficiency while reducing labor costs and the risks associated with operating in harsh environments.

[0017] 2. This invention utilizes an AI model trained on multimodal data to achieve high accuracy and in-depth fault diagnosis. It extracts multi-type features through CNN and RNN / Transformer multi-branch networks, combines this with attention mechanisms for cross-modal fusion analysis, refines fault information through hierarchical classification, and derives complex fault root causes using a knowledge graph inference engine, thus solving the problem of misjudgment inherent in traditional single-modal diagnosis. The spatiotemporal correlation cross-validation of multimodal data and the traceability design of diagnostic results further ensure accurate fault diagnosis, avoiding reactive maintenance and reducing downtime losses and repair waste caused by misjudgments.

[0018] 3. This invention upgrades the operation and maintenance strategy from prevention to prediction through predictive maintenance and automated operation and maintenance closed loop. Based on digital twin data, the time-series predictive model analyzes the performance degradation trend of components, calculates the health index, and predicts the remaining useful life, supporting advance planning of maintenance resources. Simultaneously, it automatically generates electronic work orders containing fault details, location navigation, maintenance suggestions, and dynamic priorities, pushing them to the CMMS and the mobile terminals of operation and maintenance personnel, forming a closed loop of diagnosis, work orders, and scheduling. This significantly improves the reliability and power generation efficiency of power plant operation, reduces operation and maintenance costs, and ensures economic benefits. Attached Figure Description

[0019] Figure 1 This is a flowchart of the automatic inspection and AI fault diagnosis method for new energy power plants of the present invention; Figure 2 This is a block diagram illustrating the principle of the automatic inspection and AI fault diagnosis system for new energy power plants of the present invention. Detailed Implementation

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this invention discloses an automatic inspection and AI fault diagnosis method for new energy power plants, including the following steps: Step S1: Control the automatic inspection platform equipped with multimodal sensing units to carry out inspections of the new energy power station along a preset path, and simultaneously collect the spatial coordinates of the power station equipment, multimodal sensing data, real-time operation data of the power station monitoring and data acquisition system, and environmental monitoring data, and assign them a unified timestamp and spatial coordinate label. Step S2: Construct a high-fidelity digital twin of the new energy power station that integrates the three-dimensional geometric model, the electrical physical model and historical operation data; inject the tagged multi-dimensional data collected in step S1 into the digital twin in real time to drive the digital twin to achieve synchronous mapping and status update with the physical power station; Step S3: Based on the multi-dimensional data of the injected digital twin, and using the trained artificial intelligence (AI) model, perform cross-modal fusion analysis and spatiotemporal correlation cross-validation on the image features, acoustic features, operating parameters and environmental parameters of the same device object, and perform fault diagnosis and classification of the device. Step S4: Based on the historical and real-time status data sequences accumulated in the digital twin, use a time series prediction model to analyze the degradation trend of the performance parameters of key components, calculate the health index of key components, and predict their remaining useful life. Step S5: When a fault is diagnosed or a risk is predicted, the three-dimensional location of the faulty device is highlighted in the virtual environment of the digital twin, and a fault knowledge graph is generated; the digital twin is accessed through a virtual reality interactive device to immerse the user in viewing the multi-dimensional historical data of the faulty device, the status of related devices, and the fault root cause chain. Step S6: Based on the diagnostic results of step S3 or the prediction results of step S4, automatically generate an electronic work order containing fault details, location information, maintenance suggestions and priorities, and push the electronic work order to the computerized maintenance management system.

[0022] In this embodiment, step S1 specifically involves: the inspection platform employing a wheeled / tracked ground robot, a drone (UAV), or a collaborative combination thereof. The platform is equipped with a high-precision GNSS / RTK positioning module, an inertial measurement unit (IMU), and an odometry system to achieve centimeter-level precise positioning and pose awareness of the platform itself in a global coordinate system. The multimodal sensing unit includes at least: Visual sensors: high-resolution visible light cameras, infrared thermal imaging cameras, and ultraviolet imagers. Used to acquire images of equipment appearance, temperature distribution, corona discharge, and other phenomena.

[0023] Acoustic sensors: high-fidelity microphone arrays or ultrasonic sensors. Used to collect ultrasonic signals generated by abnormal noise, vibration sound waves, or partial discharge during equipment operation.

[0024] Environmental sensors: These integrate sensors for temperature and humidity, wind speed and direction, light intensity, and atmospheric pressure, and are used to collect micro-meteorological data for the power plant.

[0025] Data interface unit: Equipped with a 4G / 5G or industrial Wi-Fi communication module, used to establish data connections with power plant monitoring and data acquisition (SCADA) systems, environmental monitoring stations, etc., and subscribe to the required real-time operating data.

[0026] The inspection platform autonomously navigates and controls its motion according to a preset path. When approaching critical equipment (such as inverters, combiner boxes, and transformers), the platform automatically adjusts its position to ensure the sensing units are aligned with the target equipment at the optimal angle and distance, guaranteeing data acquisition quality. During the platform's inspection process, the main control unit (such as an industrial computer or a high-performance edge computing device) synchronously triggers and receives data from all sensors.

[0027] Spatial coordinate labeling: Utilizing the platform's own GNSS / RTK and IMU data, the precise three-dimensional coordinates and attitude data of each sensor's optical center or detection unit in the power station's global coordinate system are calculated in real time through coordinate transformation.

[0028] Timestamp Labels: The main control unit runs a high-precision clock synchronization protocol (such as PTP or NTP) to stamp each frame of image, each audio segment, each environmental data packet, and each piece of real-time running data obtained from the SCADA system with a unified timestamp with microsecond precision.

[0029] Data association: The spatial coordinate labels and timestamp labels mentioned above are bound to the collected raw data (such as image files, audio files, data points) and the device's unique identifier (ID) to form a structured data packet.

[0030] On the platform, edge computing capabilities are used to perform preliminary preprocessing on the collected raw data, such as image compression, audio noise reduction, and data format standardization, to reduce data transmission and storage pressure. The preprocessed tagged data is then streamed in real time to the central server via a wireless network, while also being cached locally to prevent network interruptions.

[0031] In this embodiment, step S2, which constructs a high-fidelity digital twin of a new energy power plant integrating a three-dimensional geometric model, an electrical and physical model, and historical operating data, specifically involves: S21. Based on the BIM design model, oblique photogrammetry real-scene 3D model or CAD drawings of the power station, construct a 3D visualization model that includes the spatial location, geometric shape and topological connection relationship of all power station equipment and facilities (such as photovoltaic arrays, inverters, box-type transformers, combiner boxes and lines). S22. Based on the three-dimensional geometric model, embed an electrical physical model that characterizes the electrical characteristics of the equipment; the physical model includes: equivalent circuit model, power converter mathematical model, and line impedance model; for example, in the photovoltaic power station scenario, the equivalent circuit model is used to describe the single / dual diode equivalent circuit model of the photovoltaic module output, the power converter mathematical model is used to describe the power converter mathematical model of the inverter conversion efficiency and loss, and the line impedance model is used to describe the line impedance model of the electrical connection between string-combiner box-inverter.

[0032] S23. Establish the mapping relationship between the virtual objects in the digital twin and the real-time and historical databases of the physical power station; use the real-time operation data and environmental data collected in step S1 as input to drive the electrical physical model to perform calculations, so that the operation status of the virtual model is synchronized with the physical power station, and continuously store historical state sequence data. S24. By comparing the simulation output data of the digital twin with the actual measurement data of the physical power station, the parameters of the electrical physical model are automatically calibrated and optimized to ensure that its output accuracy meets the preset tolerance requirements.

[0033] Step S3 involves using multi-dimensional data from the injected digital twin, and employing a trained artificial intelligence (AI) model. The training method for the AI ​​model includes: S31. Construct a training sample set: Extract labeled multimodal data samples from the historical database of the digital twin; the samples include fault images and acoustic data collected by the multimodal sensing unit and verified by operation and maintenance results, as well as SCADA operation data, environmental data and corresponding fault type labels aligned with spatiotemporal data. S32. Data Preprocessing and Augmentation: Standardize, denoise, and augment the training samples; for image data, use rotation, cropping, and color adjustment methods, and for time-series data, use sliding window cutting method to expand sample diversity. S33. Model Building and Training: Construct a deep learning network model suitable for multimodal fusion analysis; the model should include at least a convolutional neural network (CNN) branch for image processing, a recurrent neural network (RNN) or Transformer branch for processing time-series data, and an attention mechanism layer for feature fusion; use the training sample set to train the network model end-to-end in a supervised learning manner, and iteratively optimize the model parameters by minimizing the loss function. S34. Model Validation and Deployment: Validate the trained model using the reserved test sample set. Once the accuracy and recall metrics reach the preset thresholds.

[0034] Then, cross-modal fusion analysis and spatiotemporal correlation cross-validation are performed on the image features, acoustic features, operating parameters, and environmental parameters of the same device object, specifically as follows: S35. Based on a unified timestamp and spatial coordinate label, image data, acoustic data, operating parameters and environmental parameters within the same time segment and the same geographic space are bound to the corresponding unique device object ID in the digital twin to form a multimodal data sample pair. S36. Input the bound multimodal data into the corresponding feature extraction branches, including: Image data is input into a branch of a convolutional neural network (CNN) to extract spatial visual feature vectors. After converting the acoustic data into a spectrogram, it can be input into another CNN branch, or directly into an audio feature extraction network to extract acoustic feature vectors. Input time-series data such as operating parameters and environmental parameters into a recurrent neural network (RNN) or Transformer branch to extract time-series feature vectors. S37. Input the extracted modal feature vectors into a cross-modal attention fusion layer, calculate the correlation weights between different modal features, and generate a weighted comprehensive feature vector accordingly; wherein, the temporal feature vectors of the running parameters and / or environmental parameters are used as the query vector Query to guide the attention mechanism to focus on the image and acoustic feature segments most relevant to the current device running state. S38. Input the comprehensive feature vector into the final classifier or regressor and output the fault diagnosis result. Cross-validation is reflected in the fact that the diagnosis result must be logically consistent with the preliminary diagnosis result based on image modality, the preliminary diagnosis result based on acoustic modality, and the SCADA alarm information based on data rules. If a conflict occurs, the expert rules based on knowledge graph will be triggered for arbitration, and the diagnosis conclusion after multimodal cross-validation will be output.

[0035] Finally, the equipment is diagnosed and classified for faults, specifically as follows: S39. Fault Feature Matching and Preliminary Classification: The comprehensive feature vector obtained in step S38 is matched with a predefined fault feature library; the fault feature library stores multimodal feature templates corresponding to different fault modes. Through similarity calculation, a preliminary candidate list containing fault types and their confidence levels is output. S310. Hierarchical Fault Classification and Severity Assessment: A hierarchical classification strategy is used for refined fault classification and assessment. The first level is equipment-level classification, which identifies the specific type of equipment where the fault occurred, such as photovoltaic modules, inverters, or combiner boxes. The second layer is fault mode classification, which determines the specific fault mode. For photovoltaic modules, it further distinguishes between hot spots, cracks, diode faults, shading, and dust accumulation; for inverters, it further distinguishes between IGBT faults, capacitor aging, and MPPT abnormalities. The third level is the severity classification, which divides the severity of faults into multiple levels such as early warning, general, severe, and critical, based on the intensity of the fault characteristics, the development trend, and the degree of impact on power generation efficiency. S311. Root cause reasoning results and verification based on knowledge graph: Trigger the knowledge graph-based reasoning engine. The reasoning engine traverses the fault propagation path based on the device topology connection relationship, historical maintenance records and physical rules to deduce the most likely root cause. S312. Finally, generate a structured diagnostic result, which includes at least the faulty device ID, fault type, severity level, occurrence time, location information, confidence score, and recommended handling measures.

[0036] In this embodiment, step S4 uses a time-series prediction model to analyze the degradation trend of the performance parameters of key components, calculates the health index of key components, and predicts their remaining useful life. Specifically: S41. Extract performance degradation indicators to characterize the health status of key components from the historical and real-time status data stored in the digital twin; for example, in the photovoltaic power plant scenario, for inverters, indicators include conversion efficiency deviation rate, maximum power point tracking (MPPT) deviation value, IGBT module junction temperature fluctuation amplitude, and fault code triggering frequency; for photovoltaic modules, indicators include power attenuation rate, series resistance increase rate, and EL image defect area expansion rate. S42. Based on the performance degradation index, use a weighted fusion or machine learning regression method to calculate a health index (HI) in the range of [0,1] or [0%,100%]; where a health index of 1 (or 100%) represents a brand new state and 0 (or 0%) represents a completely failed state. S43. Input the data sequence of health index changes over time into the time series prediction model; the model learns the degradation trajectory of the health index and extrapolates to predict the time required for it to reach the preset failure threshold, which is the predicted remaining useful life (RUL); the time series prediction model is a long short-term memory network (LSTM), Transformer or ParticleFilter model. S44. While outputting the RUL prediction result, provide the confidence interval or probability distribution of the prediction value to quantify the uncertainty of the prediction.

[0037] In this embodiment, step S5 highlights the three-dimensional location of the faulty device in the virtual environment of the digital twin and generates a fault knowledge graph, specifically as follows: S51. Based on the unique device ID contained in the fault diagnosis results, accurately locate the corresponding virtual device object in the three-dimensional geometric model of the digital twin; by changing the model color, applying pulsed light effects, adding warning icons, or generating bounding boxes, the faulty device is highlighted in the virtual environment and visually contrasted with normal devices. S52. Based on the root cause reasoning results or structured diagnosis results, dynamically construct a knowledge graph centered on the current fault event; the nodes of the knowledge graph shall include at least: faulty equipment entity, fault phenomenon entity, fault mode entity, root cause entity, and affected related equipment entity; the edges between nodes are used to describe the semantic relationships between entities, including "occurred at", "manifested as", "caused by", and "affected to"; S53. The generated fault knowledge graph is presented in a graphical form on the user interface. The user interface supports interactive clicking on nodes in the graph to expand and view the detailed information, historical fault records, maintenance solutions, or other related fault cases of the entity.

[0038] Simultaneously, by accessing the digital twin through virtual reality interactive devices, users can immerse themselves in viewing multi-dimensional historical data of the faulty device, the status of related devices, and the root cause chain of the fault. The root cause chain is achieved through the integration of knowledge graphs and retrieval enhancement generation RAG technology, specifically: S54. Receive a natural language query from the user or diagnostic module. The query describes the currently observed fault phenomenon. Use a generative big data model to perform semantic parsing and intent understanding on the query, and extract key entities and relationships as search conditions. S55. Based on the parsed search conditions, perform a search in the pre-built operation and maintenance knowledge graph; the search traverses the entity nodes and connecting edges related to the faulty equipment and fault mode, and the search scope includes historical similar work orders, expert experience rules, equipment manual chapters, failure mode cases and solutions. S56. Combine the retrieved relevant structured information, text fragments, and case data with the current real-time running data and multimodal feature fusion results into a contextual information package; use the information package as context, together with the instructions optimized by prompt word engineering, to form the input prompts of the large model; S57. The generative large model is based on input prompts to generate natural language descriptions containing the root cause of the failure, the reasoning logic chain, and the handling suggestions. At the same time, the model output must be accompanied by the retrieval source information on which its conclusions are based, including the case number, manual entry or rule ID, so as to realize the verifiability and traceability of the generated results.

[0039] Finally, step S6 automatically generates an electronic work order containing fault details, location information, repair suggestions, and priority based on the diagnostic results of step S3 or the prediction results of step S4, and pushes the electronic work order to the computerized maintenance management system, specifically as follows: S61. Based on the diagnostic results of step S2 or the prediction results of step S3, automatically fill in the following fields of the electronic work order: Fault details: including faulty device ID, fault type, severity level, time of occurrence, and reliability score; Location information: including the device's three-dimensional coordinates in the digital twin, its matrix / region number, and the optimal path navigation link to the device; Repair recommendations: Utilize the knowledge graph to retrieve historical successful solutions, required spare parts lists, safe operating procedures, and standard operating procedures (SOPs) that match this failure mode. Priority: The priority of work order processing is dynamically calculated based on the severity level of the fault, the power generation weight of the affected equipment, and the availability of spare parts. S62. Automatic Work Order Push and Distribution: The generated structured electronic work orders are automatically pushed to the Computerized Maintenance Management System (CMMS) for resource scheduling and recording via the application programming interface (API); at the same time, the core information and navigation links of the work orders are pushed to the mobile terminal application (APP) of the local maintenance personnel.

[0040] like Figure 2 As shown, this invention also proposes an automatic inspection and AI fault diagnosis system for new energy power plants, comprising: The multimodal data synchronous acquisition module 101 is used to control the automatic inspection platform equipped with multimodal sensing units to carry out inspections of new energy power plants along a preset path, and synchronously collect the spatial coordinates of power plant equipment, multimodal sensing data, real-time operating data of the power plant monitoring and data acquisition (SCADA) system, and environmental monitoring data, and assign them a unified timestamp and spatial coordinate label. The digital twin construction and synchronization module 102 is used to construct and maintain a high-fidelity digital twin of a new energy power plant that integrates a three-dimensional geometric model, an electrical and physical model, and historical operating data; it receives tagged multi-dimensional data from the multimodal data synchronization acquisition module 101 and drives the digital twin to achieve synchronous mapping and status updates with the physical power plant. The AI ​​fault diagnosis and classification module 103 is communicatively connected to the digital twin construction and synchronization module 102. It is used to perform cross-modal fusion analysis and spatiotemporal correlation cross-validation on the image features, acoustic features, operating parameters and environmental parameters of the same device object based on the multi-dimensional data injected into the digital twin and using the trained artificial intelligence AI model, so as to diagnose and classify the device faults. The predictive maintenance analysis module 104 is communicatively connected to the digital twin construction and synchronization module 102. It is used to analyze the degradation trend of the performance parameters of key components based on the historical and real-time status data sequences accumulated in the digital twin, and to calculate the health index of the key components and predict their remaining useful life. The fault visualization and tracing module 105 is communicatively connected to the AI ​​fault diagnosis and classification module 103 and the predictive maintenance analysis module 104. When a fault is diagnosed or a risk is predicted, it is used to highlight the three-dimensional location of the faulty equipment in the virtual environment of the digital twin and generate a fault knowledge graph. It is also used to support access through virtual reality (VR) interactive devices to immerse users in viewing multi-dimensional historical data of the faulty equipment, the status of related equipment, and the fault root cause chain. The operation and maintenance decision and work order management module 106 is connected to the AI ​​fault diagnosis and classification module 103 and the predictive maintenance analysis module 104. It is used to automatically generate electronic work orders containing fault details, location information, maintenance suggestions and priorities based on the diagnosis results or prediction results, and push the electronic work orders to the computerized maintenance management system (CMMS).

[0041] The working principle of the automatic inspection and AI fault diagnosis system of this invention is as follows: This invention's automated inspection platform comprises wheeled / tracked ground robots, drones, or combinations thereof, equipped with high-precision positioning and multimodal sensing units. The positioning module (GNSS / RTK, IMU) achieves centimeter-level spatial positioning. The sensing units encompass three core types of equipment: visual (visible light, infrared thermal imaging, ultraviolet cameras), acoustic (microphone arrays, ultrasonic sensors), and environmental (temperature, humidity, wind speed, light intensity). Simultaneously, it connects to the power station's SCADA system via 4G, 5G, or industrial Wi-Fi interfaces to acquire real-time operational data. During data collection, the system uses a high-precision clock synchronization protocol (PTP / NTP) to timestamp all data at the microsecond level. Combining this with the platform's positioning data, it binds each type of sensor data with three-dimensional spatial coordinates in a global coordinate system and a unique device ID, forming structured data packets and fundamentally solving the problem of traditional data silos.

[0042] The digital twin is the core data carrier and simulation platform of this invention. Its construction and synchronization follow the three-step principle of three-dimensional geometric modeling, electrical and physical embedding, and data mapping calibration: (1) Three-dimensional geometric model construction: Based on the power station BIM design model, oblique photography real scene or CAD drawings, restore the spatial position, geometric shape and topological connection relationship of the equipment to form a visualization basic framework; (2) Electrical and physical model embedding: On the basis of the geometric model, implant the physical model that represents the core characteristics of the equipment, such as the single / double diode equivalent circuit model of photovoltaic module, the power converter mathematical model of inverter, and the line impedance model, so that the virtual equipment has physical simulation capability; (3) Data mapping and dynamic calibration: Establish the mapping relationship between the virtual equipment and the real-time / historical database of the physical power station, inject the collected multimodal data into the physical model, and drive the virtual equipment and the physical equipment to synchronize their operating status; at the same time, by comparing the virtual simulation output with the physical measured data, automatically calibrate the physical model parameters to ensure that the simulation accuracy of the digital twin meets the preset tolerance.

[0043] The core of the AI ​​diagnostics of this invention is a multi-branch deep learning model and cross-modal fusion analysis, divided into two stages: model training and fault reasoning. Model training phase: Annotated samples (fault images and acoustic data verified by operation and maintenance, and spatiotemporally aligned SCADA, environmental data and fault labels) are extracted from the digital twin historical database. The images are enhanced by rotation and cropping, and the time series data is cut by sliding window. A multi-branch network is constructed (CNN branch processes image / acoustic spectrograms, RNN / Transformer branch processes time series data), and an attention mechanism layer is introduced to achieve feature fusion. The loss function is minimized through supervised learning until the model accuracy and recall meet the target. In the fault reasoning stage: First, based on timestamps and spatial coordinates, multimodal data of the same device are bound into data sample pairs and input into the corresponding branches to extract features. Then, the correlation weights of each modality feature are calculated through a cross-modal attention layer to generate a comprehensive feature vector. Subsequently, it is matched with a predefined fault feature library, and preliminary results are output through a hierarchical strategy of device-level classification (such as locating the type of faulty device), fault mode classification (such as hot spots / cracks in photovoltaic modules), and severity classification (such as warning / general / severe / critical). For complex faults, the knowledge graph reasoning engine is triggered to deduce the root cause by combining equipment topology and historical maintenance records, and finally generate a structured diagnostic result containing fault ID, type, severity, location, and credibility.

[0044] This invention breaks through the traditional post-event response model. Based on historical and real-time data from digital twins, it achieves predictive maintenance through performance index extraction, health index calculation, and Remaining Useful Life (RUL) prediction. Specifically: Performance degradation index extraction: Core indicators characterizing equipment health are selected from the digital twin database, such as inverter conversion efficiency deviation rate and photovoltaic module power attenuation rate; Health Index (HI) calculation: Weighted fusion or machine learning regression methods are used to normalize multiple indicators into HI values ​​within the [0,1] interval (1 represents a brand-new state, 0 represents complete failure), quantifying the equipment's health status; Remaining Useful Life (RUL) prediction: HI time-series data is input into LSTM, Transformer, or particle filter models. The model learns the HI degradation trajectory and extrapolates the time it takes to reach the failure threshold, which is the RUL; simultaneously, the confidence interval or probability distribution of the RUL is output, quantifying prediction uncertainty and providing a forward-looking basis for operation and maintenance planning.

[0045] When a fault is diagnosed or a risk is predicted, the system achieves fault visualization and root cause tracing through 3D highlighting, knowledge graph tracing, and VR immersive viewing. Based on the faulty device ID, the virtual device is accurately located in the digital twin and highlighted by changing color, pulsed illumination, and adding warning icons to visually distinguish it from normal devices. Simultaneously, a knowledge graph is constructed around the fault event, containing nodes such as "faulty device - fault phenomenon - fault mode - root cause - affected device." Nodes are connected by semantic edges such as "occurred at" and "caused by," allowing users to click on nodes to view detailed data and historical cases. Upon receiving a fault phenomenon query, the system uses a generative large-scale model to parse the semantics and extract search conditions. It retrieves information such as historical work orders, expert rules, and equipment manuals from the operations and maintenance knowledge graph, combining this with real-time data to create context. This drives the large-scale model to generate natural language results containing root causes, reasoning chains, and handling suggestions, along with the search source, ensuring the verifiability and traceability of root cause tracing. VR device access is also supported, allowing operations and maintenance personnel to immerse themselves in viewing historical data of faulty devices and the status of related devices.

[0046] Finally, this invention automatically generates and pushes standardized electronic work orders, forming a closed-loop operation and maintenance system. The work order automatically populates core fields based on diagnostic / predictive results: fault details (including device ID, type, severity, etc.), location information (including digital twin 3D coordinates, area number, and optimal navigation link), maintenance suggestions (combined with historical solutions, spare parts list, and SOP specifications), and dynamic priority (calculated based on fault severity, power generation weight, and spare parts inventory). It is then pushed to the Computerized Maintenance Management System (CMMS) via API, while simultaneously sending the core information and navigation link to the mobile devices of maintenance personnel.

[0047] In summary, this invention, through the deep integration of multiple technologies, breaks through the data silos and passive mode of traditional operation and maintenance, and realizes intelligent operation and maintenance with holographic perception, accurate diagnosis, forward-looking prediction and efficient handling of power plants, significantly improving the reliability and economic benefits of power plant operation.

[0048] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles described in the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A new energy power station automatic inspection and AI fault diagnosis method, characterized in that, Includes the following steps: Step S1: Control the automatic inspection platform equipped with multimodal sensing units to carry out inspections of the new energy power station along a preset path, and simultaneously collect the spatial coordinates of the power station equipment, multimodal sensing data, real-time operation data of the power station, and environmental monitoring data, and assign them a unified timestamp and spatial coordinate label. Step S2: Construct a high-fidelity digital twin of the new energy power station that integrates a three-dimensional geometric model, an electrical and physical model, and historical operating data; inject the collected tagged multi-dimensional data into the digital twin in real time to drive the digital twin to achieve synchronous mapping and status update with the physical power station; Step S3: Based on the multi-dimensional data of the injected digital twin, and using the trained artificial intelligence (AI) model, perform cross-modal fusion analysis and spatiotemporal correlation cross-validation on the image features, acoustic features, operating parameters and environmental parameters of the same device object, and perform fault diagnosis and classification of the device. Step S4: Based on the historical and real-time status data sequences accumulated in the digital twin, analyze the degradation trend of power plant component performance parameters using a time-series prediction model, calculate the health index of the power plant component, and predict its remaining useful life. Step S5: When a fault is diagnosed or a risk is predicted, the three-dimensional location of the faulty device is highlighted in the virtual environment of the digital twin, and a fault knowledge graph is generated; the digital twin is accessed through a virtual reality interactive device to immerse the user in viewing the multi-dimensional historical data of the faulty device, the status of related devices, and the fault root cause chain. Step S6: Based on the diagnostic results of step S3 or the prediction results of step S4, automatically generate an electronic work order containing fault details, location information, maintenance suggestions and priorities, and push the electronic work order to the computerized maintenance management system.

2. The method of claim 1, wherein, The construction of a high-fidelity digital twin of a new energy power plant, integrating a three-dimensional geometric model, an electrical and physical model, and historical operating data, specifically involves: Based on the BIM design model, oblique photogrammetry real scene 3D model or CAD drawings of the power station, construct a 3D visualization model that includes the spatial location, geometric shape and topological connection relationship of the power station equipment and facilities. Based on the three-dimensional visualization model, an electrical physical model representing the electrical characteristics of the equipment is embedded; the electrical physical model includes: an equivalent circuit model, a power converter mathematical model, and a line impedance model. Establish a mapping relationship between the virtual objects in the digital twin and the real-time and historical databases of the physical power station; use the collected real-time operation data and environmental data as input to drive the electrical physical model to perform calculations, so that the operation status of the virtual model is synchronized with the physical power station, and continuously store historical state sequence data; By comparing the simulation output data of the digital twin with the actual measurement data of the physical power station, the parameters of the electrical physical model are automatically calibrated and optimized. 3.The method of claim 1, wherein, The training methods for the artificial intelligence (AI) model include: Annotated multimodal data samples are extracted from the historical database of the digital twin; the multimodal data samples include fault images and acoustic data collected by the multimodal sensing unit and verified by operation and maintenance results, as well as SCADA operation data, environmental data and corresponding fault type labels aligned with spatiotemporal data. The multimodal data samples are standardized, denoised, and augmented; specifically, rotation, cropping, and color adjustment methods are used for image data, and sliding window cutting method is used for time-series data. Construct a deep learning network model suitable for multimodal fusion analysis, which includes at least a convolutional neural network (CNN) branch for image processing, a recurrent neural network (RNN) or Transformer branch for processing temporal data, and an attention mechanism layer for feature fusion; use the training sample set to train the network model end-to-end in a supervised learning manner, and iteratively optimize the model parameters by minimizing the loss function; The trained model is validated using a reserved test sample set until the accuracy and recall metrics reach preset thresholds.

4. The method of claim 1, wherein the method further comprises: The cross-modal fusion analysis and spatiotemporal correlation cross-validation of image features, acoustic features, operating parameters, and environmental parameters of the same device object are specifically as follows: Based on unified timestamps and spatial coordinate labels, image data, acoustic data, operating parameters and environmental parameters within the same time segment and geographic space are bound to the corresponding unique device object ID in the digital twin to form multimodal data sample pairs; The bound multimodal data are then input into the corresponding feature extraction branches, including: Image data is input into a branch of a convolutional neural network (CNN) to extract spatial visual feature vectors. After converting the acoustic data into a spectrogram, it can be input into another CNN branch, or directly into an audio feature extraction network to extract acoustic feature vectors. Input time-series data such as operating parameters and environmental parameters into a recurrent neural network (RNN) or Transformer branch to extract time-series feature vectors. The extracted modal feature vectors are input into a cross-modal attention fusion layer to calculate the correlation weights between different modal features and generate a weighted comprehensive feature vector accordingly. The comprehensive feature vector is input into the final classifier or regressor, which outputs the fault diagnosis result, and then the equipment is diagnosed and classified for faults.

5. The method of claim 4, wherein, The process of diagnosing and classifying equipment faults specifically includes: The obtained comprehensive feature vector is matched with a predefined fault feature library; the fault feature library stores multimodal feature templates corresponding to different fault modes. Through similarity calculation, a preliminary candidate list containing fault types and their confidence levels is output. Hierarchical fault classification and severity assessment employs a hierarchical classification strategy to classify and assess faults in a refined manner, including the first level of equipment-level classification, the second level of fault mode classification, and the third level of severity grading. Trigger the knowledge graph-based inference engine, which traverses the fault propagation path based on device topology connections, historical maintenance records, and physical rules; Generate structured diagnostic results, including at least the faulty device ID, fault type, severity level, occurrence time, location information, reliability score, and recommended remedial measures.

6. The method of claim 1, wherein the method further comprises: The process of using a time-series prediction model to analyze the degradation trend of key component performance parameters, calculate the health index of key components, and predict their remaining useful life is as follows: From the historical and real-time status data stored in the digital twin, performance degradation indicators used to characterize the health status of key components are extracted; Based on the performance degradation index, a health index HI ranging from [0,1] is calculated using a weighted fusion or machine learning regression method; where a health index of 1 represents a brand new state and 0 represents a completely failed state. The data sequence of the health index changing over time is input into the time-series prediction model. By learning the degradation trajectory of the health index, the time required for it to reach the preset failure threshold is extrapolated and predicted. This time is the predicted remaining useful life (RUL). The time-series prediction model is a Long Short-Term Memory (LSTM) network, a Transformer, or a ParticleFilter model. Along with the output of the RUL prediction result, the confidence interval or probability distribution of the predicted value is provided to quantify the uncertainty of the prediction.

7. The method of claim 1, wherein the method further comprises: The three-dimensional location of the faulty device is highlighted in the virtual environment of the digital twin, and a fault knowledge graph is generated, specifically as follows: Based on the unique device ID contained in the fault diagnosis results, the corresponding virtual device object is accurately located in the three-dimensional geometric model of the digital twin; by changing the model color, applying pulsed light effects, adding warning icons, or generating bounding boxes, the faulty device is highlighted in the virtual environment and visually contrasted with normal devices. Based on the root cause reasoning results or structured diagnosis results, a knowledge graph centered on the current fault event is dynamically constructed. The nodes of the knowledge graph include at least: faulty equipment entity, fault phenomenon entity, fault mode entity, root cause entity, and affected related equipment entity; Edges between nodes are used to describe the semantic relationships between entities; The generated fault knowledge graph is presented in a graphical form on the user interface, which allows users to interactively click on nodes in the graph to expand and view detailed information, historical fault records, repair solutions, or other related fault cases of that entity.

8. The method of claim 1, wherein the method further comprises: The root cause chain of the failure is achieved by integrating knowledge graphs and retrieval-enhanced RAG generation technology, specifically as follows: Receive natural language queries from users or diagnostic modules, which describe the currently observed fault phenomena; use generative large models to perform semantic parsing and intent understanding on the queries, and extract key entities and relationships as search conditions; Based on the parsed search criteria, a search is performed in the pre-constructed operation and maintenance knowledge graph; the search traverses the entity nodes and connecting edges related to the faulty equipment and fault mode, and the search scope includes historical similar work orders, expert experience rules, equipment manual chapters, failure mode cases and solutions; The retrieved structured information, text fragments, and case data are combined with the current real-time running data and multimodal feature fusion results into a contextual information package; The information packet is used as context, and together with the instructions optimized by prompt word engineering, it constitutes the input prompts of the large model; Based on the input prompts, the generative large model generates a natural language description that includes the root cause of the failure, the logical chain of reasoning, and the proposed solutions. At the same time, the model output must include the retrieval source information on which its conclusions are based, including the case number, manual entry, or rule ID of the cited case. 9.The method of claim 1, wherein, Based on the diagnostic results of step S3 or the prediction results of step S4, an electronic work order is automatically generated, containing fault details, location information, repair suggestions, and priorities. Specifically: Based on the diagnostic results of step S3 or the prediction results of step S4, automatically populate the following fields of the electronic work order: Fault details: including faulty device ID, fault type, severity level, time of occurrence, and reliability score; Location information: including the device's three-dimensional coordinates in the digital twin, its matrix / region number, and the optimal path navigation link to the device; Repair recommendations: Utilize the knowledge graph to retrieve historical successful solutions, required spare parts lists, safe operating procedures, and standard operating procedures (SOPs) that match this failure mode; Priority: The priority of work order processing is dynamically calculated based on the severity level of the fault, the power generation weight of the affected equipment, and the availability of spare parts. The generated structured electronic work orders are automatically pushed to the computerized maintenance management system via the application programming interface (API) for resource scheduling and recording; at the same time, the core information and navigation links of the work orders are pushed to the mobile terminal applications of local maintenance personnel.

10. A system for implementing the method of automatic inspection and AI fault diagnosis of new energy power stations according to any one of claims 1-9, characterized in that, include: The multimodal data synchronous acquisition module is used to control the automatic inspection platform equipped with multimodal sensing units to carry out inspections of new energy power plants along a preset path, and synchronously collect the spatial coordinates of power plant equipment, multimodal sensing data, real-time operating data of the power plant monitoring and data acquisition SCADA system, and environmental monitoring data, and assign them a unified timestamp and spatial coordinate label. The digital twin construction and synchronization module is used to construct and maintain a high-fidelity digital twin of a new energy power plant that integrates a three-dimensional geometric model, an electrical and physical model, and historical operating data. It receives tagged multi-dimensional data from the multimodal data synchronization acquisition module and drives the digital twin to achieve synchronous mapping and status updates with the physical power plant. The AI ​​fault diagnosis and classification module is communicatively connected to the digital twin construction and synchronization module. It is used to perform cross-modal fusion analysis and spatiotemporal correlation cross-validation on the image features, acoustic features, operating parameters and environmental parameters of the same device object based on the multi-dimensional data injected into the digital twin and using a trained artificial intelligence AI model, so as to diagnose and classify the device faults. The predictive maintenance analysis module is communicatively connected to the digital twin construction and synchronization module. It is used to analyze the degradation trend of key component performance parameters based on the historical and real-time status data sequences accumulated in the digital twin, calculate the health index of key components, and predict their remaining useful life. The fault visualization and tracing module is communicatively connected to the AI ​​fault diagnosis and classification module and the predictive maintenance analysis module. When a fault is diagnosed or a risk is predicted, it is used to highlight the three-dimensional location of the faulty device in the virtual environment of the digital twin and generate a fault knowledge graph. It is also used to support access through virtual reality interactive devices to immerse users in viewing multi-dimensional historical data of the faulty device, the status of related devices, and the fault root cause chain. The operation and maintenance decision and work order management module is communicatively connected to the AI ​​fault diagnosis and classification module and the predictive maintenance analysis module. It is used to automatically generate electronic work orders containing fault details, location information, maintenance suggestions and priorities based on the diagnosis results or prediction results, and push the electronic work orders to the computerized maintenance management system.

Images