An unmanned intelligent operation and maintenance system for a new energy station based on multi-source data fusion

Abstract

The application belongs to the technical field of new energy intelligent operation and maintenance, and discloses a new energy station unattended intelligent operation and maintenance system based on multi-source data fusion, which integrates various types of sensing terminals such as visible light, infrared and voiceprint through an acquisition module, comprehensively covers scenes such as wind turbine generators, booster stations and power transmission lines, and breaks the limitation of a single data source of a traditional system by matching a space-time synchronous calibration and hierarchical preprocessing mechanism; the fusion module adopts an edge and cloud collaborative architecture, can dynamically switch the fusion strategy according to actual needs, can also realize cross-modal semantic alignment, and can optimize the acquisition parameters through a bidirectional feedback mechanism, so that the multi-source data form effective complementation; the identification module constructs a small sample learning algorithm system, uses multi-modal collaborative sample enhancement technology and meta-learning driven adaptation capability, and can quickly identify new defects only with a small amount of samples.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for new energy, specifically a new energy power station unmanned intelligent operation and maintenance system based on multi-source data fusion. Background Technology

[0002] With the large-scale development of the new energy industry, wind farms, photovoltaic power stations, and other facilities are mostly located in remote areas. Equipment is dispersed and the operating environment is complex and variable, making traditional operation and maintenance models inadequate for the industry's demand for efficient operation. Current operation and maintenance practices mainly face the following technical problems:

[0003] Most existing operation and maintenance systems rely on a single data source for collection and analysis. Even if some systems are equipped with video surveillance, sensors and other equipment, they have failed to establish a deep integration system of multi-source data such as optical, infrared, acoustic, and environmental parameters. Different types of data have inconsistent formats and information is isolated from each other, resulting in an incomplete perception of equipment status. This can easily lead to missed or misjudged faults and make it difficult to support a comprehensive assessment of equipment status in unattended scenarios.

[0004] Traditional systems often use fixed thresholds for alarms, which can only identify faults that have already occurred and cannot capture fault trends during the gradual change of equipment. At the same time, the operation and maintenance decision-making process relies heavily on human experience and lacks intelligent diagnosis and optimization scheduling capabilities based on data fusion. This leads to untimely fault handling, unreasonable allocation of operation and maintenance resources, and makes it difficult to achieve the industry goal of early detection, early handling, and precise operation and maintenance. Summary of the Invention

[0005] The purpose of this invention is to provide an unmanned intelligent operation and maintenance system for new energy power stations based on multi-source data fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an unattended intelligent operation and maintenance system for new energy power stations based on multi-source data fusion, comprising a data acquisition module, a fusion module, an identification module, a decision-making module, a scheduling module, a risk linkage module, and a knowledge graph module.

[0007] Preferably, the acquisition module integrates acquisition terminals including visible light cameras, infrared thermal imaging equipment, acoustic fingerprint sensors, temperature and humidity sensors, gas sensors, drones, rail-mounted robots, and intelligent inspection terminals, covering the entire scene of wind turbines, substations, and transmission lines, and has the functions of spatiotemporal synchronous calibration, fusion demand reverse-driven acquisition, and hierarchical preprocessing.

[0008] The spatiotemporal synchronization calibration mechanism adopts a dual synchronization strategy combining the PTP precision time protocol and hardware pulse triggering to establish a unified time reference for each acquisition terminal. For devices that do not support hardware synchronization, software-level timestamp interpolation and resampling compensation are used to ensure that the alignment error of image, soundprint, and sensor data in the time dimension does not exceed ±5ms. At the same time, spatial coordinate binding of the acquired data is achieved by combining the site's GIS geographic information data. The reverse-driven acquisition based on fusion requirements is achieved through a dynamic optimization unit for acquisition parameters. This unit receives feature difference data fed back by the fusion module. When the contribution of a specific type of data to defect identification increases, the acquisition accuracy and frequency of the corresponding data are automatically increased.

[0009] The hierarchical preprocessing mechanism includes a data quality grading unit and a lightweight noise reduction unit. The data quality grading unit classifies the standardized data into three levels, A, B, and C, through noise detection and outlier removal algorithms. Level A data is directly connected to the fusion module, while Level B and C data are connected after lightweight noise reduction processing at the edge. The acquisition module maintains a working condition-driven and protocol-unified mode, acquiring data every 15 minutes under normal working conditions, and increasing to full acquisition at the 1-second level when under high load or near the fault threshold, uniformly compatible with the IEC61850 standard, and outputting standardized data to the fusion module.

[0010] The data quality grading unit uses a combination of 3σ and isolated forest algorithms to achieve noise detection and outlier removal: For numerical data from sensors such as temperature, humidity, and gas concentration, the 3σ principle is used to identify outliers that deviate from the mean by 3σ, with σ ranging from 1.2 to 1.5 and dynamically adjusted according to the data type; For unstructured data such as images and voiceprints, the isolated forest algorithm is used to detect noise and outlier samples, with the number of decision trees set to 100, the sample subset size set to 256, and the outlier threshold set to 0.7.

[0011] The lightweight noise reduction unit uses median filtering with a window size of 3×3 for B-level data and wavelet thresholding for C-level data, selecting the db4 wavelet basis, decomposing into 3 layers, and using a soft threshold λ=0.02×data standard deviation to ensure that the data retains key features after noise reduction.

[0012] High-load operating condition criteria: wind turbine load rate ≥ 85%, substation main transformer load rate ≥ 80%, transmission line load rate ≥ 90% of rated load;

[0013] Criteria for determining near-fault threshold conditions: Equipment operating parameters reach the preset warning threshold range; for example, bearing temperature ≥90℃, insulation resistance ≤100MΩ, and gas sensor detection concentration ≥80% of the alarm threshold.

[0014] The aforementioned thresholds for load rate, temperature, and resistance can be calibrated using historical fault data from the knowledge graph module, based on different equipment models at different sites. The calibration coefficient is 0.95 to 1.05.

[0015] Preferably, the fusion module receives standardized multi-source data output by the acquisition module and establishes a collaborative architecture of lightweight edge fusion and deep cloud fusion, including three major mechanisms: dynamic fusion strategy switching, cross-modal alignment enhancement, and bidirectional feedback of fusion results.

[0016] The dynamic fusion strategy switching selects the fusion level based on the quality level of the collected data and the requirements for defect identification. In high-fault scenarios, an edge-side feature-level fusion priority strategy is adopted, and the weight of key data features is strengthened through an attention mechanism. During the stable operation phase of the equipment, an edge-decision-level preliminary fusion combined with a cloud-based deep fusion mode is adopted. Cross-modal alignment enhancement is achieved through a cross-modal alignment algorithm. In view of the semantic differences between image, voiceprint, and sensor data, semantic associations of different modal data are established through feature embedding association to improve the effectiveness of fused features.

[0017] Feature embedding employs a Transformer-based cross-modal embedding model with 6 encoder layers, 8 multi-head attention heads, and 512-dimensional features. The specific steps are as follows:

[0018] ① CNN features are extracted from visible light, infrared and other image data. The Layer 4 output of ResNet50 is used with a dimension of 2048. MFCC features (13-dimensional coefficients + first-order difference + second-order difference, a total of 39 dimensions) are extracted from the voiceprint data. After standardization of the sensor data, a 64-dimensional feature vector is formed.

[0019] ② By using an embedding model, various features are mapped to a 512-dimensional unified semantic space;

[0020] ③ Cosine similarity is used to calculate the correlation between features of different modalities. The similarity threshold is set to 0.65. Feature pairs with correlation higher than the threshold are bound into a fusion feature group to achieve cross-modal semantic alignment.

[0021] Specific adaptation rules for data quality levels and fusion strategies:

[0022] When performing deep fusion of Grade A data (which is free of noise and outliers after hierarchical preprocessing by the acquisition module), a feature splicing and attention-weighted fusion mode is adopted to prioritize the preservation of original feature details of each modality.

[0023] For B-level data, an "feature consistency verification" step is added during feature-level fusion at the edge side to remove cross-modal feature conflicts. The conflict judgment criterion is that the correlation between different modal features is less than 0.3.

[0024] After noise reduction processing, only core features are extracted from the C-level data at the edge, such as the temperature anomaly range of infrared data and the main frequency anomaly segment of voiceprint data. When fusion in the cloud, their weight ratio is reduced to no more than 0.2.

[0025] Specific steps for edge-side feature-level fusion:

[0026] ① Feature extraction is performed on various types of single-modal data, including texture and shape features from image data, spectral features from voiceprint data, and temporal features from sensor data;

[0027] ②Use normalization to unify feature scale;

[0028] ③ Generate edge-side fusion feature vectors based on weighted summation of cross-modal correlation.

[0029] The two-way feedback mechanism of the fusion results enables collaboration with the acquisition module, reverses the deep fusion results to optimize its acquisition strategy, outputs multimodal fusion feature data to the recognition module, and forms a collaborative closed loop between the edge, cloud and acquisition.

[0030] Preferably, the identification module receives multimodal fusion feature data from the fusion module, constructs a dedicated library containing few-shot learning algorithms, and integrates multimodal collaborative sample enhancement, meta-learning-driven model adaptation, and cross-module bidirectional optimization mechanisms.

[0031] Multimodal collaborative sample enhancement utilizes cross-modal data correlation to generate complementary samples. Based on the correlation between infrared thermal imaging defect regions, visible light texture features, and abnormal audioprint spectra, multimodal collaborative small samples are generated through an improved GAN network to improve sample expansion efficiency. Meta-learning-driven model adaptation is achieved through model-independent meta-learning algorithms in the small sample learning algorithm library. The model-independent meta-learning algorithm is pre-trained on multiple similar defect recognition tasks and can quickly adjust model parameters with a small number of newly added defect samples to adapt to the needs of new types of defect recognition.

[0032] The dedicated library for few-shot learning algorithms includes three core algorithms and applicable scenarios:

[0033] ① The model-independent meta-learning algorithm is suitable for scenarios where the similarity between the new defect and the existing defect type is ≥60%, and the adaptation speed is fast, requiring only 5~10 samples to complete the adaptation.

[0034] ② The prototype network algorithm is suitable for scenarios with low defect feature dimensions, such as simple rust, leakage and other appearance defects. It achieves recognition by calculating the distance between the new sample and the prototype feature, and has high computational efficiency.

[0035] ③ The matching network algorithm is suitable for defect scenarios with strong correlation between multimodal features, such as faults with superimposed infrared thermal anomalies and voiceprint anomalies. Similarity matching improves recognition accuracy. Algorithm selection mechanism: The recognition module automatically calculates the feature dimensions and similarity of newly added defect samples, triggering the corresponding algorithm call.

[0036] Specific steps for meta-learning pre-training and adaptation:

[0037] ① Collect multimodal samples of common defects such as oil leakage and metal corrosion as pre-training data, with 50 to 100 samples of each class, and construct 500 task sets. Each task set contains 10 support set samples + 20 query set samples.

[0038] ② The pre-training process uses the MAML algorithm, with an inner learning rate of 0.01, an outer learning rate of 0.001, 10,000 iterations, SGD as the optimizer, and Xavier uniform distribution for model parameter initialization;

[0039] ③ New defect adaptation: Input 5-10 new defect samples as support set, adjust the model parameters through 10 inner layer updates, and complete the adaptation through 1 outer layer update. After adaptation, the model recognition accuracy is ≥85%.

[0040] ④ The network backbone of the model-independent meta-learning algorithm uses MobileNetV3-Small, with an input size of 224×224, a classification head of 2 fully connected layers, and an output dimension equal to the number of defect categories.

[0041] The specific structure of the improved GAN network:

[0042] ① The generator adopts the U-Net architecture. The encoder contains 4 convolutional blocks, each containing 2 3×3 convolutional layers + BatchNorm + ReLU activation function, with a stride of 2; the decoder contains 4 deconvolutional blocks, each containing 1 2×2 deconvolutional layer + BatchNorm + ReLU activation function; the output has 3 channels, corresponding to infrared, visible light, and voiceprint feature maps;

[0043] ② The discriminator adopts the PatchGAN architecture, which contains 5 convolutional blocks, 3×3 convolutional layers + LeakyReLU activation function, stride 2; the output is a 32×32 Patch similarity matrix;

[0044] ③ The loss function adopts WGAN-GP loss + cross-modal consistency loss; training parameters: batchsize=16, number of iterations=50000, initial learning rate=0.0002, Adam optimizer is used, β1=0.5, β2=0.999, and the learning rate is decayed to 0.9 times the original value every 1000 iterations.

[0045] Cross-module bidirectional optimization feeds back the defect identification results to the fusion module, optimizes the weight allocation of data for each modality, and outputs the equipment defect identification results and operational anomaly information.

[0046] The weight allocation for each modality was calculated using a combination of entropy weighting and analytic hierarchy process.

[0047] ① Calculate the objective weight using the entropy weight method: Based on the information entropy of the defect identification results, E_i = -ΣP_ij × lnP_ij, where P_ij is the contribution rate of the i-th modal data to the identification of the j-th type of defect, and the objective weight w_obj_i = (1-E_i) / Σ(1-E_i).

[0048] ② Analytic Hierarchy Process (AHP) for calculating subjective weights: Construct a three-layer structure: target layer (defect identification accuracy) - criterion layer (modal effectiveness) - solution layer (each modality), with consistency test index CR < 0.1 and subjective weight w_sub_i;

[0049] ③ The combined weight w_i = 0.6 × w_obj_i + 0.4 × w_sub_i, the weight update cycle is 1 hour, and the weight ratio of infrared (initial weight 0.3), visible light (0.25), soundprint (0.2), and sensor (0.25) is dynamically adjusted.

[0050] Preferably, the decision-making module receives the equipment defect identification results and operational anomaly information output by the identification module, and combines the raw data from the acquisition module and the fusion features from the fusion module to construct a two-way interactive architecture of physical entities, virtual models, and operation and maintenance decisions, integrating multi-physics joint simulation, virtual-real linkage decision-making pre-play, and three-dimensional enhanced verification and knowledge linkage mechanisms.

[0051] Multiphysics co-simulation integrates a multiphysics co-simulation engine, including fluid mechanics, structural mechanics, and thermodynamics, on the basis of a full-element digital twin. It can simulate the operating status and fault evolution process of equipment under complex working conditions. Virtual-real linkage decision-making pre-simulation executes virtual pre-simulation of candidate operation and maintenance schemes based on multi-source fusion data and digital twin models, evaluates the implementation effect, resource consumption and safety risks of the schemes, and automatically selects the optimal scheme. The implementation process and results of the operation and maintenance schemes are synchronized to the digital twin model in real time for the purpose of correcting model parameters.

[0052] Multiphysics co-simulation engine parameter settings:

[0053] ① The fluid dynamics simulation uses the k-ε turbulence model (Cμ=0.09, C1=1.44, C2=1.92), with a mesh size of 5~20mm, a time step of 0.01s, and an iteration convergence residual ≤1×10⁻⁶. -4 ;

[0054] ② The structural mechanics simulation uses the elasticity equation, with elastic modulus E=206GPa, Poisson's ratio μ=0.3, density ρ=7850kg / m³, and the mesh type is tetrahedral mesh. The constraint condition is set as full constraint at the fixed end of the equipment.

[0055] ③ The thermodynamic simulation uses the Fourier heat conduction equation, with thermal conductivity λ=50W / (m·K), specific heat capacity c=460J / (kg·K), and the boundary condition is ambient temperature. The measured data from the acquisition module are dynamically input over time.

[0056] ④ The coupling method adopts unidirectional coupling, fluid mechanics → structural mechanics → thermodynamics, the coupling step size is 0.1s, the simulation result output frequency is 1Hz, the deviation threshold between the simulation result and the measured data is set to 5%, and the model parameters are corrected when the deviation exceeds the threshold.

[0057] Multiphysics coupling data interaction interface: Adopting a standardized data exchange format, the pressure field and velocity field data output from the fluid mechanics simulation are transmitted to the structural mechanics simulation module in real time through the interface as the load input for structural stress calculation; the deformation data output from the structural mechanics simulation is transmitted to the thermodynamics simulation module to correct the heat conduction boundary conditions; the data interaction delay between modules is ≤10ms, the interface is compatible with the output formats of ANSYS and ABAQUS simulation software, and supports real-time data reading, writing and backtracking.

[0058] The three-dimensional enhanced verification mode adopts a three-dimensional architecture based on multi-physics simulation deviation, measured data difference, and historical case matching degree. It obtains historical fault case support through linkage with the knowledge graph module, incorporates the simulation cases of the digital twin model into the knowledge graph, and finally outputs graded fault diagnosis results and optimal operation and maintenance solutions.

[0059] Preferably, the scheduling module receives the graded fault results and optimal operation and maintenance plan output by the decision module, and is driven by digital twin simulation, integrating virtual pre-simulation optimization of scheduling plan, four-dimensional priority evaluation and full-process closed-loop feedback mechanism.

[0060] The virtual pre-simulation optimization of the scheduling scheme is initiated before the improved ant colony algorithm route planning. The implementation path, time consumption, resource consumption, and potential safety risks of different scheduling schemes are simulated through a digital twin model. The route planning and resource allocation are adjusted based on the simulation results to ensure the feasibility and optimality of the scheme. The four-dimensional priority evaluation model includes four dimensions: fault level, equipment impact range, operation and maintenance cost, and safety risk level. Critical faults and high safety risk tasks are double-toped, and resources are allocated in a priority manner.

[0061] Four-dimensional priority quantization and weight calculation:

[0062] ① Fault level quantification: Critical (5 points), Severe (3 points), and General (1 point), based on the degree of impact of the defect on equipment operation. For example, blade cracking is critical, and slight corrosion is general.

[0063] ② Quantification of equipment impact range: single equipment (1 point), local area (3 points), entire site (5 points), calculate the number of affected nodes based on the equipment association topology;

[0064] ③ Quantification of operation and maintenance costs: Low (1 point, cost < 50,000 yuan), Medium (3 points, 50,000 to 200,000 yuan), High (5 points, > 200,000 yuan), based on the accounting of manpower, spare parts, and downtime losses;

[0065] ④ Safety risk level quantification: low (1 point), medium (3 points), high (5 points), and extremely high (7 points), based on the probability of personnel injury risk and equipment explosion / fire risk. A probability >30% is considered extremely high.

[0066] ⑤ The entropy weight method is used to determine the weight of each dimension: fault level (0.35), equipment impact range (0.25), operation and maintenance cost (0.2), and safety risk level (0.2). Priority score = Σ (dimensional score × dimension weight). A score ≥ 4 is a first-level task (priority scheduling), 2~4 is a second-level task, and < 2 is a third-level task.

[0067] The end-to-end closed-loop feedback enables the full-link flow of operation and maintenance data. Operation and maintenance personnel upload the progress and results of handling through smart terminals, generate closed-loop reports, and feed them back to the decision-making module and knowledge graph module in real time. These reports are used to correct the parameters of the digital twin model, enrich the operation and maintenance case library, and simultaneously optimize the scheduling algorithm parameter settings. The scheduling task information is also synchronized to the risk linkage module.

[0068] Preferably, the risk linkage module links the real-time data of the acquisition module, the defect results of the identification module, and the task information of the scheduling module to build a dynamic risk-adaptive linkage mechanism, integrating dynamic risk level adaptation, full-process security linkage of operation and maintenance, and digital twin backtracking of security events.

[0069] The dynamic risk level adaptation is based on real-time data from the acquisition module and defect identification results to dynamically assess the site's safety risk level. Different risk levels correspond to different safety monitoring frequencies and linkage rule thresholds. The full-process safety linkage of operation and maintenance integrates safety prevention and control into all aspects of operation and maintenance. When receiving operation and maintenance tasks pushed by the scheduling module, it simultaneously sends safety risk warnings, protection requirements, and emergency response plans for the task area. During the operation, it monitors personnel location and environmental parameters in real time. When safety hazards occur, it immediately triggers linkage measures including sound and light reminders and access control locking, and dynamically adjusts the priority of operation and maintenance tasks and feeds them back to the scheduling module.

[0070] The security incident twin retrospective records security alarm events and handling processes into a digital twin model, forming a security incident retrospective archive. It uses virtual simulation to recreate the event evolution process, analyze the causes of risks, and optimize security linkage rules. At the same time, it synchronizes security incident cases to the knowledge graph module.

[0071] Preferably, the knowledge graph module integrates historical data from the acquisition module, fusion features from the fusion module, defect cases from the identification module, simulation cases from the decision-making module, operation and maintenance records from the scheduling module, and security event archives from the risk linkage module to construct a dynamic iterative architecture driven by multi-source data, integrating multi-source data fusion and updating, cross-module knowledge collaboration, and dynamic optimization matching algorithms.

[0072] Multi-source data fusion and updating automatically mines the relationships between historical operational data, fusion feature data, defect cases, simulation cases, operation and maintenance records, and security event archives of each module through reinforcement learning algorithms, and updates the five-dimensional architecture of the knowledge graph, including equipment, defects, causes, solutions, and risks. Cross-module knowledge collaboration provides historical fault correlation data and simulation boundary conditions for the decision-making module, similar defect cases and feature matching basis for the identification module, historical statistical data on operation and maintenance costs and security risks for the scheduling module, and security event correlation and prevention rules for the risk linkage module.

[0073] The dynamic optimization matching algorithm continuously optimizes itself through incremental learning, dynamically adjusting the case matching weights according to changes in the operation and maintenance scenario, improving the accuracy of matching the optimal solution, and forming a cycle of knowledge accumulation, collaborative application, and iterative optimization.

[0074] The beneficial effects of this invention are as follows:

[0075] 1. This invention integrates multiple types of sensing terminals, such as visible light, infrared, and acoustic fingerprint sensors, through a data acquisition module, comprehensively covering scenarios such as wind turbines, substations, and transmission lines. Coupled with spatiotemporal synchronization calibration and hierarchical preprocessing mechanisms, it breaks through the limitations of traditional systems relying on a single data source. The fusion module adopts an edge-cloud collaborative architecture, dynamically switching fusion strategies according to actual needs and achieving cross-modal semantic alignment. Furthermore, a bidirectional feedback mechanism optimizes acquisition parameters, enabling effective complementarity among multiple data sources. This improves data integrity and relevance, comprehensively capturing details of equipment operating status, significantly reducing the probability of missed or incorrect fault detection, and providing data support for comprehensive equipment status assessment in unattended scenarios.

[0076] 2. The identification module of this invention constructs a small-sample learning algorithm system, which utilizes multimodal collaborative sample enhancement technology and meta-learning-driven adaptation capabilities to quickly identify new defects with only a small number of samples. The decision module relies on full-element digital twins and multi-physics joint simulation technology to simulate the fault evolution process and predict potential risks in advance, rather than just identifying faults that have already occurred. At the same time, by virtually pre-simulating the operation and maintenance plan, the optimal handling plan is selected, which gets rid of the lag and blindness of traditional manual decision-making, realizes early detection and early handling of faults, and reduces equipment downtime losses and operation and maintenance costs.

[0077] 3. The scheduling module of this invention achieves optimal allocation of operation and maintenance resources through four-dimensional priority evaluation and improved ant colony algorithm; the risk linkage module can dynamically adapt to risk levels, integrate safety prevention and control into all aspects of operation and maintenance, and ensure operational safety; the knowledge graph module integrates data from all modules to construct a five-dimensional architecture of equipment, defects, causes, solutions, and risks, and continuously optimizes algorithm parameters, case matching weights, and linkage rules through cross-module knowledge collaboration and dynamic iteration; it gets rid of the dependence on human experience in traditional operation and maintenance, making resource scheduling more accurate and safety prevention and control more comprehensive. Attached Figure Description

[0078] Figure 1 This is a flowchart of the overall system of the present invention;

[0079] Figure 2 This is a flowchart of the multi-source data acquisition and fusion process of the present invention.

[0080] Figure 3 This is a flowchart of the defect identification and intelligent decision-making process of the present invention;

[0081] Figure 4 This is a flowchart of the closed-loop process of full-link operation and maintenance scheduling and knowledge iteration in this invention. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] like Figures 1 to 4 As shown in the figure, this embodiment of the invention provides an unmanned intelligent operation and maintenance system for new energy power stations based on multi-source data fusion, including a data acquisition module, a fusion module, an identification module, a decision-making module, a scheduling module, a risk linkage module, and a knowledge graph module. The specific implementation of each module is as follows:

[0084] The acquisition module integrates acquisition terminals such as visible light cameras, infrared thermal imaging equipment, acoustic sensors, temperature and humidity sensors, gas sensors, drones, rail-mounted robots, and intelligent inspection terminals, covering all scenarios of wind turbines, substations, and transmission lines. It has spatiotemporal synchronous calibration, fusion demand reverse-driven acquisition, and hierarchical preprocessing functions.

[0085] Wind turbine units include blades, transformer substations, and drive trains; substations include main transformers, GIS modules, and distribution rooms; transmission lines include conductors, insulators, and towers.

[0086] The spatiotemporal synchronization calibration mechanism adopts a dual synchronization strategy combining the PTP precision time protocol and hardware pulse triggering to establish a unified time reference for each acquisition terminal. For devices that do not support hardware synchronization, software-level timestamp interpolation and resampling compensation are used to ensure that the alignment error of image, soundprint, and sensor data in the time dimension does not exceed ±5ms. At the same time, the spatial coordinates of the acquired data are accurately bound by the site's GIS geographic information data to solve the problem of spatiotemporal fragmentation of multi-source data. The reverse-driven acquisition based on fusion requirements is achieved through a dynamic optimization unit for acquisition parameters. This unit receives feature difference data fed back by the fusion module. When the contribution of specific types of data (such as soundprint data from a specific area) to defect identification increases, the acquisition accuracy and frequency of the corresponding data are automatically increased.

[0087] The PTP precision time protocol adopts a master-slave clock architecture. The master clock uses a GPS synchronized clock source with a synchronization period of 1 second. Clock deviation compensation uses an adaptive filtering algorithm with a filtering window size of 5 synchronization periods. Hardware pulse triggering uses a 1PPS signal with a pulse width of 100ms and synchronization triggered on the rising edge to ensure that the master-slave clock deviation is ≤1ms.

[0088] Specific upward adjustment criteria for dynamic optimization of collected parameters:

[0089] ① When the feature difference value of a certain type of data is ≥30%, the feature difference value is calculated by the fusion module to determine the degree of deviation between the data and historical normal data. The image data resolution is increased from 1920×1080 to 2560×1440, and the sensor data sampling accuracy is increased from 0.1 level to 0.05 level.

[0090] ② When the feature difference value is ≥20%, the sampling frequency under normal working conditions is increased from 15 minutes / time to 5 minutes / time; when the feature difference value is ≥40%, it is directly increased to 1 second / time, consistent with the high load working conditions.

[0091] ③ When the feature difference value of three consecutive acquisitions is ≤10%, the acquisition accuracy and frequency will be automatically restored to the original to avoid wasting resources.

[0092] The hierarchical preprocessing mechanism includes a data quality grading unit and a lightweight noise reduction unit. The data quality grading unit classifies the standardized data into three levels (A, B, and C) through noise detection and outlier removal algorithms. Level A data is directly connected to the fusion module, while Level B and C data are connected after lightweight noise reduction processing at the edge side. This ensures the quality of the input data while reducing the redundancy of subsequent fusion calculations. The acquisition module maintains a working condition-driven and protocol-unified mode. Under normal working conditions, it acquires data every 15 minutes. Under high load or when approaching the fault threshold, it increases to full acquisition at the 1-second level, uniformly compatible with the IEC61850 standard, and outputs standardized data to the fusion module.

[0093] The fusion module receives standardized multi-source data output by the acquisition module, establishes a collaborative architecture of lightweight edge fusion and deep cloud fusion, and realizes efficient data fusion and real-time feedback. It includes three major mechanisms: dynamic fusion strategy switching, cross-modal alignment enhancement, and bidirectional feedback of fusion results.

[0094] The dynamic fusion strategy switching selects the fusion level based on the quality level of the collected data and the requirements for defect identification. In high-fault scenarios, an edge-side feature-level fusion priority strategy is adopted, which strengthens the weight of key data features such as voiceprint and infrared through an attention mechanism, with the weight dynamically ranging from 0.6 to 0.8. During the stable operation phase of the equipment, an edge-decision-level preliminary fusion combined with a cloud-based deep fusion mode is adopted, balancing fusion efficiency and comprehensiveness. Cross-modal alignment enhancement is achieved through a cross-modal alignment algorithm. For the semantic differences between image, voiceprint, and sensor data, semantic associations of different modal data are established through feature embedding association. For example, the infrared thermal anomaly area of ​​the equipment is bound to the corresponding voiceprint spectrum anomaly, improving the effectiveness of fusion features.

[0095] The two-way feedback mechanism of the fusion results enables collaboration with the acquisition module, which in turn optimizes the acquisition strategy by using the deep fusion results, outputs multimodal fusion feature data to the recognition module, and forms a collaborative closed loop between the edge, cloud and acquisition, reducing cloud computing pressure and improving the real-time performance and accuracy of fusion.

[0096] The identification module receives multimodal fusion feature data from the fusion module, constructs a dedicated library containing multiple few-shot learning (FSL) algorithms, integrates multimodal collaborative sample enhancement, meta-learning-driven model adaptation, and cross-module bidirectional optimization mechanisms, adapts to scenarios where field defect data is scarce, and achieves accurate identification of equipment defects and operational anomalies.

[0097] Multimodal collaborative sample enhancement utilizes cross-modal data correlation to generate complementary samples. Based on the correlation between infrared thermal imaging defect regions, visible light texture features, and abnormal speaker spectrum, an improved GAN network generates multimodal collaborative small samples to solve the recognition bias problem caused by the scarcity of single-modal samples and improve sample expansion efficiency. Meta-learning-driven model adaptation is achieved through the model-independent meta-learning algorithm in the small sample learning algorithm library. The model-independent meta-learning (MAML) algorithm is pre-trained on multiple similar defect recognition tasks and can quickly adjust model parameters with a small number of newly added defect samples to adapt to the needs of new types of defect recognition and improve the model's adaptability to rare defects.

[0098] Cross-module bidirectional optimization feeds back the defect identification results to the fusion module, optimizes the weight allocation of data for each mode, and outputs the equipment defect identification results and abnormal operation information; it can accurately identify defects such as oil leakage, metal corrosion, blade cracking, insulator damage, cable damage, and surge arrester aging, as well as abnormal operation information such as voltage fluctuation.

[0099] The decision-making module receives the equipment defect identification results and operation anomaly information output by the identification module, and combines the original data from the acquisition module and the fusion features of the fusion module to construct a two-way interactive architecture of physical entity, virtual model, and operation and maintenance decision-making, integrating multi-physics joint simulation, virtual-real linkage decision-making pre-drill, and three-dimensional enhanced verification and knowledge linkage mechanism.

[0100] Multiphysics co-simulation integrates multiphysics co-simulation engines such as fluid mechanics, structural mechanics, and thermodynamics on the basis of a full-element digital twin. It can simulate the operating status and fault evolution process of equipment under complex working conditions such as extreme wind conditions and high temperature environments, and improve the foresight and accuracy of fault diagnosis. Virtual-real linkage decision-making pre-drills execute virtual pre-drills of candidate operation and maintenance schemes based on multi-source fusion data and digital twin models, evaluate the implementation effect, resource consumption and safety risks of the schemes, and automatically select the optimal scheme. The implementation process and results of the operation and maintenance schemes are synchronized to the digital twin model in real time to correct the model parameters, thereby improving the simulation accuracy.

[0101] Construction and calibration process of a full-element digital twin:

[0102] ① Integrate the three-dimensional dimensions, material parameters, installation location, and historical operating data (such as voltage, current, and temperature over the past year) of the integrated acquisition module, as well as the equipment design drawings and performance parameter tables provided by the manufacturer;

[0103] ② Using geometric modeling and physical property mapping, the appearance and internal structure of the equipment are first restored using 3D modeling tools, and then physical parameters such as material thermal conductivity and structural strength are assigned to the corresponding model parts;

[0104] ③ Every 24 hours, the real-time operating data (such as the actual temperature and vibration amplitude of the equipment) of the acquisition module is compared with the simulation data of the digital twin model. If the deviation exceeds 5%, the physical parameters of the model are automatically adjusted, such as correcting the thermal conductivity and damping coefficient, to ensure the consistency between the virtual model and the physical entity.

[0105] ④ When equipment undergoes maintenance modifications (such as component replacement) or new equipment is added, the modification parameters or new equipment data are imported synchronously to update the structure and parameters of the digital twin.

[0106] The three-dimensional enhanced verification mode adopts a three-dimensional architecture of multi-physics simulation deviation, measured data difference, and historical case matching degree. It obtains historical fault case support through linkage with the knowledge graph module, incorporates the simulation cases of the digital twin model into the knowledge graph to enrich the decision-making basis, classifies the fault diagnosis results into three levels: general, serious, and critical, and finally outputs the graded fault diagnosis results and the optimal operation and maintenance plan.

[0107] The scheduling module receives the graded fault results and optimal operation and maintenance plan output by the decision module. Driven by digital twin simulation, it integrates virtual pre-simulation optimization of scheduling plan, four-dimensional priority evaluation and full-process closed-loop feedback mechanism.

[0108] The virtual pre-simulation optimization of the scheduling scheme is initiated before the improved ant colony algorithm route planning. The implementation path, time consumption, resource consumption, and potential safety risks of different scheduling schemes are simulated through a digital twin model. The route planning and resource allocation are adjusted based on the simulation results to ensure the feasibility and optimality of the scheme. The four-dimensional priority evaluation model includes four dimensions: fault level, equipment impact range, operation and maintenance cost, and safety risk level. Critical faults and high safety risk tasks are both reset and prioritized, and resources are allocated to ensure that core equipment can be safely and quickly restored to operation.

[0109] Details of the adaptation and optimization of the improved ant colony algorithm:

[0110] ① Heuristic function optimization: Combining the distribution characteristics of new energy power station equipment, the heuristic function introduces equipment spacing weight and fault urgency weight. The higher the fault urgency, the more the heuristic function guides the ant colony to prioritize the path where the equipment is located.

[0111] ②Pheromone update strategy: A combination of local and global updates is adopted. During local updates, only a small amount of pheromone is updated between adjacent nodes of the planned path to avoid premature convergence. During global updates, a large amount of pheromone is deposited only on the optimal path (shortest time and least resource consumption).

[0112] ③ Set up obstacles such as ditches and high-voltage dangerous areas as pheromone-restricted areas, and automatically avoid them when the ant colony searches for a path;

[0113] ④ When the algorithm has not found the optimal path after 50 iterations, the pheromone evaporation coefficient is automatically adjusted from 0.1 to 0.05 to expand the search range.

[0114] Algorithm adaptability to site size: Supports path planning for ≤500 devices per site, with planning time ≤10 seconds.

[0115] Initial parameters for the improved ant colony algorithm:

[0116] ① Initial pheromone concentration τ0 = 0.5;

[0117] ② The initial value of the pheromone evaporation coefficient ρ = 0.1 (the dynamic adjustment range is 0.05~0.2);

[0118] ③ Heuristic function weights α=1.2 (pheromone weight), β=1.8 (fault urgency weight);

[0119] ④ Number of ant colonies = 1 / 10 of the number of device nodes (minimum of 5 and maximum of 50).

[0120] ⑤ Threshold for determining the optimal path: When the path time fluctuation is ≤3% for 10 consecutive iterations, it is determined to be the optimal path.

[0121] The end-to-end closed-loop feedback enables the full-link flow of operation and maintenance data. Operation and maintenance personnel upload the progress and results of handling through smart terminals, generate closed-loop reports, and feed them back to the decision-making module and knowledge graph module in real time. These reports are used to correct the parameters of the digital twin model, enrich the operation and maintenance case library, and simultaneously optimize the scheduling algorithm parameter settings. The scheduling task information is also synchronized to the risk linkage module.

[0122] The risk linkage module links the real-time data of the acquisition module, the defect results of the identification module, and the task information of the scheduling module to build a dynamic risk adaptation linkage mechanism, integrating dynamic risk level adaptation, full-process security linkage of operation and maintenance, and digital twin backtracking of security events.

[0123] Dynamic risk level adaptation is based on real-time data from the acquisition module and defect identification results to dynamically assess the site's safety risk level. Risk levels are divided into low, medium, high, and extremely high, with different risk levels corresponding to differentiated safety monitoring frequencies and linkage rule thresholds. Under extremely high risk levels, the monitoring frequency is automatically increased to the 0.5-second level, expanding the warning range for dangerous areas. Safety linkage throughout the entire operation and maintenance process integrates safety prevention and control into all aspects of operation and maintenance. When receiving operation and maintenance tasks pushed by the scheduling module, it simultaneously sends safety risk warnings, protection requirements, and emergency response plans for the task area. During the operation, it monitors personnel location and environmental parameters in real time. When safety hazards occur, it immediately triggers linkage measures such as audible and visual alerts and access control locking, and dynamically adjusts the priority of operation and maintenance tasks and feeds them back to the scheduling module.

[0124] Specific threshold indicators for risk level determination:

[0125] ①Low risk: Equipment operating parameters deviate from rated values ​​by ≤10%, with no obvious defects, and environmental parameters such as temperature, humidity, and gas concentration are within safe ranges. The safe range is: temperature -20℃ to 60℃, humidity ≤85%, and combustible gas concentration ≤10% of the lower explosive limit (LEL).

[0126] ② Medium risk: Equipment operating parameters deviate from rated values ​​by 10% to 20%, there are minor defects, such as slight corrosion or minor leakage, environmental parameters are close to the safety threshold, and the concentration of combustible gas is 10% to 30% of the lower explosive limit;

[0127] ③ High risk: Equipment operating parameters deviate from rated values ​​by 20%~30%, serious defects exist, such as blade crack length ≤5cm, insulator damage area ≤10cm², environmental parameters exceed safety thresholds, combustible gas concentration 30%~50% of the lower explosive limit, temperature >80℃;

[0128] ④ Extremely high risk: Equipment operating parameters deviate from rated values ​​by ≥30%, there are critical defects, such as blade crack length >5cm, cable damage exposing the core, environmental parameters seriously exceeding standards, combustible gas concentration ≥50% of the lower explosive limit, toxic gas concentration ≥ occupational exposure limit.

[0129] The above thresholds can be calibrated periodically using historical fault data from the knowledge graph module, with a calibration cycle of 3 months.

[0130] Table 1. Correspondence between risk level, monitoring frequency, and linkage rule threshold:

[0131] ,

[0132] The security incident twin retrospective records security alarm events and handling processes into a digital twin model, forming a security incident retrospective archive. Through virtual simulation, it restores the event evolution process, analyzes the causes of risks and optimizes security linkage rules. At the same time, it synchronizes security incident cases to the knowledge graph module, improving the risk prediction and prevention capabilities of the entire system.

[0133] The knowledge graph module integrates historical data from the acquisition module, fusion features from the fusion module, defect cases from the identification module, simulation cases from the decision-making module, operation and maintenance records from the scheduling module, and security event archives from the risk linkage module to construct a dynamic iterative architecture driven by multi-source data, integrating multi-source data fusion and updates, cross-module knowledge collaboration, and dynamic optimization matching algorithms.

[0134] Multi-source data fusion and updating automatically mines the correlation between data in each module through reinforcement learning algorithms, updating the five-dimensional architecture of the knowledge graph, including equipment, defects, causes, solutions, and risks. Cross-module knowledge collaboration provides the decision-making module with historical fault correlation data and simulation boundary conditions, the identification module with similar defect cases and feature matching basis, the scheduling module with historical statistical data on operation and maintenance costs and security risks, and the risk linkage module with security event correlation and prevention rules, realizing the efficient flow and reuse of knowledge between modules.

[0135] Specific data interaction methods for cross-module knowledge collaboration:

[0136] ① Provide data to the decision-making module: Push historical fault cases with a matching degree of ≥80% with the current defect, including fault phenomena, handling process, effect feedback, as well as equipment material parameters and historical operating limit data required for simulation;

[0137] ② Provide data to the identification module: output multimodal feature templates of similar defects, such as the infrared feature range and acoustic spectrum range of different types of corrosion, as a reference benchmark for defect identification;

[0138] ③ Provide data to the scheduling module: Feedback on resource consumption statistics of historical maintenance tasks (such as average working hours for replacing insulators and required spare parts models), and probability of safety risks, such as the accident rate of high-altitude operations;

[0139] ④ Provide data to the risk linkage module: Push related prevention and control rules for similar safety events. For example, if a gas leak accident has occurred in a certain area, push the gas concentration monitoring threshold and emergency response procedures for that area.

[0140] Key data is pushed in real time, while non-key data is synchronized in batches every hour.

[0141] Association rules of the five-dimensional architecture of knowledge graph:

[0142] ① The association between equipment and defects is based on equipment type and years of operation. A mapping is established between equipment model and operating time → types of defects that are prone to occur. For example, wind turbine blades that have been in operation for more than 5 years are prone to cracking and corrosion defects.

[0143] ② Correlation between defects and causes: By statistically analyzing historical cases, the correspondence between defects and causes can be clarified. For example, oil leakage defects → causes include aging of seals, loose bolts, and damaged oil pipes.

[0144] ③ Defect and Solution Correlation: Match maintenance solutions to defects based on their severity. For example, minor rust → grinding and rust removal + applying anti-corrosion paint; severe rust → replacing parts.

[0145] ④ Defect and Risk Correlation: Determine the risk level based on the scope of the defect's impact and its evolution speed. For example, blade cracking → extremely high risk; minor leakage → medium risk.

[0146] ⑤ Record the potential risks in the implementation of the operation and maintenance plan, such as replacing parts at height → high safety risk level, which requires the use of a fall protection plan.

[0147] For every 100 new operation and maintenance cases added, new association rules are automatically discovered, such as the association between new defects and specific environmental parameters, and added to the knowledge graph.

[0148] The dynamic optimization matching algorithm continuously optimizes through incremental learning, dynamically adjusting the case matching weights based on changes in operation and maintenance scenarios such as new equipment and environmental changes, thereby improving the accuracy of matching the optimal solution and forming a virtuous cycle of knowledge accumulation, collaborative application, and iterative optimization, which in turn empowers the performance improvement of all modules in the entire system.

[0149] Core parameters for dynamic iteration:

[0150] ① The Proximal Policy Optimization (PPO) algorithm is adopted. The state space is the feature vector of data association in each module, the action space is the update operation of the association rule of the knowledge graph, the reward function is the improvement value of case matching accuracy, and the target reward value is the improvement of case matching accuracy ≥ 5%;

[0151] ② Incremental learning trigger conditions: The cumulative number of newly added operation and maintenance cases is ≥100, or the number of newly added equipment types is ≥3, or the change in environmental parameters is ≥20%, such as a sudden change in regional temperature and humidity;

[0152] ③ Knowledge graph update frequency: Real-time updates after incremental learning is triggered; if not triggered, batch optimization is performed at 2:00 AM daily. The update process does not affect the real-time data access of each module.

[0153] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0154] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A new energy power station unmanned intelligent operation and maintenance system based on multi-source data fusion, characterized in that, It includes a data acquisition module, a data fusion module, a recognition module, a decision-making module, a scheduling module, a risk linkage module, and a knowledge graph module. The data acquisition module integrates full-scene sensing terminals, covering all equipment and transmission lines of new energy power plants. It has functions such as spatiotemporal synchronous calibration, dynamic optimization of acquisition parameters based on demand, and data quality classification and preprocessing. It adopts an adaptive acquisition mode based on operating conditions and a standardized protocol to output standardized data. Fusion Module: Receives the standardized data, establishes a collaborative data fusion architecture between the edge and cloud, integrates dynamic adaptation fusion strategies and a two-way feedback adjustment mechanism, realizes multi-source data fusion and feature extraction, and outputs multimodal fusion feature data; Identification module: Based on the multimodal fusion feature data, a small sample learning algorithm system is established to adapt to the scenario of scarce site defect samples, realize the identification of equipment defects and operational anomalies, and output fault information; Decision module: Based on the defect identification results, it establishes a virtual-real interaction architecture by combining multi-source raw data and fused data, integrates multi-physics joint simulation and virtual pre-simulation mechanism of operation and maintenance scheme, and realizes fault diagnosis and optimal operation and maintenance scheme matching output; The scheduling module receives the operation and maintenance decision results, is driven by digital twin simulation, integrates scheduling scheme pre-simulation and optimization and a full-link closed-loop feedback mechanism to ensure the scheduling of operation and maintenance resources; Risk linkage module: Establishes a dynamic risk adaptation mechanism, integrating risk level assessment and full-process security control functions; Knowledge Graph Module: Establishes a dynamic and iterative knowledge system, integrates multi-source data fusion and updates, and cross-module knowledge collaboration mechanisms; The data acquisition module integrates a full-scene perception terminal including a visible light camera, an infrared thermal imaging device, a voiceprint sensor, a temperature and humidity sensor, a gas sensor, a drone, a rail-mounted robot, and an intelligent inspection terminal, covering the entire scene of wind turbines, substations, and transmission lines. The spatiotemporal synchronization calibration mechanism adopts a dual synchronization strategy to establish a unified time reference and combines the station's GIS geographic information data to achieve spatial coordinate binding of the collected data. The acquisition parameters are dynamically optimized to receive feature difference data fed back by the fusion module, and the acquisition accuracy and frequency of the corresponding data are adjusted as needed. The data quality classification preprocessing is achieved by noise detection and outlier removal algorithms. After corresponding processing, the classified data is connected to the fusion module. The working condition adaptive acquisition mode uses different acquisition frequencies under normal working conditions, high load working conditions or near fault threshold working conditions, and is uniformly compatible with the IEC61850 standard. The edge-cloud collaborative architecture of the fusion module specifically consists of lightweight edge fusion and deep cloud fusion, and also integrates a cross-modal alignment enhancement mechanism; The dynamic adaptation and fusion strategy switches the fusion level according to the quality level of the collected data and the requirements for defect identification. In high-fault scenarios, the edge-side feature-level fusion strategy is preferred, while in the stable operation phase of the equipment, the edge decision-level preliminary fusion combined with the cloud-based deep fusion mode is adopted. The cross-modal alignment enhancement mechanism establishes semantic associations between different modal data through feature embedding and association. The bidirectional feedback adjustment mechanism optimizes the acquisition strategy of the acquisition module by using the deep fusion results, forming a collaborative closed loop between the edge, cloud, and acquisition.

2. The unmanned intelligent operation and maintenance system for new energy power stations based on multi-source data fusion as described in claim 1, characterized in that, The few-shot learning algorithm system of the recognition module includes a dedicated library of few-shot learning algorithms, and also integrates multimodal collaborative sample enhancement, meta-learning-driven model adaptation and cross-module bidirectional optimization mechanism; The multimodal collaborative sample enhancement utilizes cross-modal data correlation to generate complementary samples, thereby improving sample expansion efficiency; the meta-learning-driven model adaptation is achieved through model-independent meta-learning algorithms in the algorithm library. After pre-training on multiple similar defect type identification tasks, it can quickly adapt to the requirements of new type defect identification with a small number of newly added samples. The cross-module bidirectional optimization feeds back the defect identification results to the fusion module to optimize the weight allocation of each modality's data.

3. The unmanned intelligent operation and maintenance system for new energy power stations based on multi-source data fusion as described in claim 2, characterized in that, The virtual-real interaction architecture of the decision module is a two-way interaction architecture of physical entity-virtual model-operation and maintenance decision, and also integrates three-dimensional enhanced verification and knowledge linkage mechanism. The multiphysics co-simulation integrates a multiphysics co-simulation engine on the basis of a full-element digital twin to simulate the equipment's operating status and fault evolution process; the virtual pre-simulation of the operation and maintenance plan is based on multi-source fused data and digital twin model to perform candidate plan pre-simulation, and automatically selects the optimal plan after evaluation, and the plan implementation process and results are synchronized to the digital twin model in real time; The three-dimensional enhanced verification is verified by multi-physics simulation deviation, measured data difference and historical case matching degree, combined with historical fault cases supported by the knowledge graph module, and finally outputs graded fault diagnosis results and optimal operation and maintenance solutions.

4. The unmanned intelligent operation and maintenance system for new energy power stations based on multi-source data fusion as described in claim 3, characterized in that, The scheduling module also integrates a four-dimensional priority evaluation model; The scheduling scheme pre-simulation optimization is initiated before route planning. The implementation path, time consumption, resource consumption and potential safety risks of different schemes are simulated through a digital twin model. The planning and allocation are adjusted based on the simulation results. The four-dimensional priority evaluation model includes four dimensions: fault level, equipment impact range, operation and maintenance cost and safety risk level. It performs dual reset of critical faults and high safety risk tasks. The full-link closed-loop feedback enables the full-link flow of operation and maintenance data. The generated closed-loop report is fed back to the decision-making module and knowledge graph module in real time, and the scheduling task information is pushed to the risk linkage module simultaneously.

5. The unmanned intelligent operation and maintenance system for new energy power stations based on multi-source data fusion according to claim 4, characterized in that, The dynamic risk adaptation mechanism of the risk linkage module links the real-time data of the acquisition module, the defect results of the identification module, and the task information of the scheduling module, and also integrates the digital twin backtracking function of security events. The risk level assessment is based on real-time data and defect identification results to dynamically determine the site safety risk level, and different levels correspond to different monitoring frequencies and linkage rule thresholds. The entire process of security control is integrated into all aspects of operation and maintenance, pushes security alerts and emergency response plans to task areas, monitors the operation process in real time and triggers emergency linkage measures, and dynamically adjusts task priorities and feeds them back to the scheduling module. The digital twin backtracking of security events forms a security event backtracking archive, restores the event evolution process, optimizes security linkage rules, and synchronizes cases to the knowledge graph module.

6. The unmanned intelligent operation and maintenance system for new energy power stations based on multi-source data fusion as described in claim 5, characterized in that, The dynamic iterative knowledge system of the knowledge graph module integrates historical data, fused feature data, defect cases, simulation cases, operation and maintenance records and security event archives of each module to construct a five-dimensional architecture of equipment, defects, causes, solutions and risks, and also integrates dynamic optimization matching algorithms; The multi-source data fusion update automatically mines the data relationships between various modules and updates the knowledge graph through algorithms; the cross-module knowledge collaboration provides data support and basis for each module; the dynamic optimization matching algorithm continuously and dynamically optimizes the case matching weight through incremental learning, forming a cycle of knowledge accumulation, collaborative application, and iterative optimization.

Images