Power utilization inspection hidden danger detection method based on intelligent image recognition and three-dimensional simulation

Abstract

The invention relates to the technical field of electricity utilization inspection, in particular to an electricity utilization inspection hidden danger detection method based on intelligent image recognition and three-dimensional simulation. According to the technical scheme, the power utilization inspection hidden danger detection method based on intelligent image recognition and three-dimensional simulation comprises the following steps that S1, image data and environment parameter data of power field equipment are collected through a mobile terminal provided with a multispectral imaging module, and a multi-modal input set is formed; wherein the multispectral imaging module synchronously collects visible light, infrared thermal imaging and ultraviolet corona discharge data, and obtains temperature, humidity and noise level parameters in combination with an environment sensor; s2, carrying out adaptive preprocessing on the collected data, wherein the adaptive preprocessing comprises data cleaning, feature extraction and quality evaluation; the data quality is effectively improved through multispectral synchronous acquisition and self-adaptive preprocessing, so that the hidden danger detection accuracy is remarkably improved; and secondly, the intelligent identification model enhances the capability of capturing tiny defects, and the risk of missing detection is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of electrical inspection technology, and in particular to a method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation. Background Technology

[0002] Traditional power industry inspections primarily rely on manual on-site operations, with inspectors identifying equipment hazards through visual observation and simple instruments. This method has significant limitations: manual inspections are susceptible to environmental factors and subjective experience, easily leading to missed or false detections; inspection efficiency is low, making it difficult to meet the demands of rapid screening of large-scale equipment; personnel training is lengthy and costly, and standardized procedures are lacking. While existing technologies have introduced some testing equipment, the asynchronous acquisition of multi-source data and reliance on manual experience in analysis processes prevent the formation of intelligent closed-loop management. With the increasing complexity of power equipment, traditional methods can no longer meet the demands of modern power grids for efficient and accurate inspections.

[0003] To address the aforementioned shortcomings, the following technical issues urgently need to be resolved: insufficient synchronization in multimodal data acquisition leads to analysis errors; preprocessing lacks adaptability and struggles to cope with different equipment and environmental conditions; the hazard identification model has weak generalization ability and poor interpretability; the training system is disconnected from actual testing; and report generation is inefficient and carries the risk of tampering. These problems severely restrict the development of intelligent electricity inspection and necessitate a complete solution. Summary of the Invention

[0004] This invention proposes a method for detecting potential hazards in electricity inspection based on intelligent image recognition and 3D simulation. This method solves the problem that traditional manual inspection methods in the prior art are significantly lacking in efficiency, accuracy, and intelligence, and cannot meet the needs of modern power grids for efficient and accurate hazard investigation.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting potential electrical hazards based on intelligent image recognition and 3D simulation includes the following steps: S1: Image data and environmental parameter data of power field equipment are collected by a mobile terminal equipped with a multispectral imaging module to form a multimodal input set. The multispectral imaging module simultaneously collects visible light, infrared thermal imaging and ultraviolet corona discharge data, and combines environmental sensors to obtain temperature, humidity and noise level parameters. S2: Adaptive preprocessing of the collected data, including data cleaning, feature extraction and quality assessment, wherein the preprocessing adopts an optical correction algorithm based on device type and an environmental parameter filtering chain to optimize data quality; S3: The preprocessed data is input into a deep convolutional neural network model based on the attention mechanism to achieve intelligent identification and classification of hidden danger features. The model adopts a multi-scale feature pyramid structure and embeds a spatial-channel dual attention module to focus on key areas.

[0006] Furthermore, it also includes: S4: The model output results drive the 3D simulation system to generate interactive training scenarios in real time, including equipment disassembly animation, hazard evolution simulation and standard handling process demonstration. The 3D simulation system is based on a physics engine to achieve multi-field coupled rendering. S5: Based on the identification results and simulation feedback, a structured hazard report is automatically generated and digital evidence is stored. The report generation uses knowledge graph technology to link historical cases, and the evidence storage process is based on blockchain to ensure that the data is tamper-proof.

[0007] Furthermore, the acquisition method of the multispectral imaging module in step S1 includes: Develop an adaptive acquisition strategy based on the equipment's operating status, and dynamically adjust imaging parameters, including resolution, frame rate, and exposure time, according to the equipment type (such as transformers and high-voltage switchgear) to optimize data quality in different scenarios; A spatiotemporal registration algorithm for multi-source data is designed to solve the synchronization problem between visible light, infrared and ultraviolet data through feature point matching and coordinate transformation, thereby ensuring data alignment accuracy. A real-time monitoring mechanism for acquisition quality is introduced, which uses a deep learning model to evaluate image clarity and integrity, and automatically prompts for reacquisition or parameter adjustment.

[0008] Furthermore, the adaptive preprocessing method described in step S2 includes: To address surface reflections on metal equipment, polarization imaging technology is used to suppress specular reflections, and histogram equalization is combined to enhance texture features. To address environmental noise, an adaptive filtering combination strategy is developed, which automatically selects either Kalman filtering or wavelet threshold denoising algorithm based on the noise spectrum characteristics. Establish a preprocessing effect evaluation system, quantify data quality through peak signal-to-noise ratio and structural similarity indicators, and provide feedback to optimize preprocessing parameters.

[0009] Furthermore, the intelligent recognition method for the deep convolutional neural network model described in step S3 includes: A multi-level feature extraction network is constructed, and local details and global semantic information are fused through a feature pyramid to improve the robustness of hazard detection; Embedding deformable convolutional layers adapts to changes in the viewing angle and occlusion of device images, enhancing the model's generalization ability; Adversarial training techniques are employed to generate indistinguishable samples to optimize the model's decision boundary and improve the recognition rate of rare defects.

[0010] Furthermore, the scene generation method of the three-dimensional simulation system in step S4 includes: A multi-field coupling model of power equipment is established based on physical laws to simulate electromagnetic, thermal and mechanical interactions and realistically reproduce the fault evolution process. Develop a real-time data-driven mechanism to map recognition results to virtual scenes and dynamically adjust animation content and interaction logic; It integrates a natural interaction interface, supports gesture recognition and voice control, and allows users to operate virtual devices from multiple angles, enhancing the immersive experience of training.

[0011] Furthermore, the structured hazard report generation method described in step S5 includes: Construct a knowledge graph for power equipment, integrate design specifications, historical hazards, and rectification plans, and use graph neural networks to infer the root causes of hazards. The design report automatic generation engine populates key fields based on templates and real-time data, and recommends optimization and rectification measures; Implement a blockchain-based evidence storage process, using a hash chain structure to store report fingerprints, supporting distributed verification and traceability.

[0012] Furthermore, the model described in step S3 also includes an interpretability analysis module: Heatmaps are generated by gradient-weighted activation mapping to visualize the regions of interest of the model and provide a basis for decision-making. Develop a feature importance ranking algorithm to identify key features that affect classification results and enhance model transparency; By combining expert knowledge bases, semantic interpretation of model outputs is performed to reduce the risk of false alarms.

[0013] Furthermore, the 3D simulation system described in step S4 supports personalized training and optimization: Establish a trainee competency assessment model, quantify skill levels through operational data, and dynamically adjust training difficulty and content; Design a path planning algorithm based on reinforcement learning to recommend the best learning sequence based on the learner's progress; Enables multi-student collaborative training mode, supports role assignment and team tasks, and improves collaboration efficiency; In step S5, the digital evidence storage method includes: The model is updated using a federated learning framework, and the recognition algorithm is optimized using distributed data, while protecting user privacy. Develop an automatic verification mechanism for smart contracts to check data compliance during the evidence storage process and reduce manual intervention; Establish a data lifecycle management system for evidence storage, regularly archive and clean up expired information, and ensure the efficient operation of the system.

[0014] Furthermore, the method also includes a cloud-edge collaborative processing architecture: Deploy lightweight models at the edge to achieve low-latency data collection and preliminary identification; Perform complex model training and big data analysis in the cloud, supporting global system optimization; Design a dynamic task scheduling algorithm to allocate computing resources based on network conditions and ensure system reliability.

[0015] This invention effectively improves data quality through multispectral synchronous acquisition and adaptive preprocessing, significantly enhancing the accuracy of hazard detection. Secondly, the intelligent identification model strengthens the ability to capture subtle defects, greatly reducing the risk of missed detections. Immersive training enabled by the 3D simulation system not only shortens the skills development cycle but also improves training effectiveness. The combination of knowledge graph and blockchain technology ensures the credibility and traceability of reports while improving management efficiency. By reducing reliance on manual labor and operational costs, a complete intelligent detection closed loop is formed, providing reliable technical support for the digital transformation of the power industry. These improvements work together to promote the standardization and intelligentization of electricity inspection work, providing strong guarantees for the safe and stable operation of the power grid.

[0016] By organically integrating advanced sensing technology, artificial intelligence algorithms, and simulation training systems, a highly efficient and reliable system for detecting potential hazards in electrical equipment has been constructed. This system not only solves the pain points of traditional methods but also pioneers a new model for intelligent detection of power equipment, possessing significant value for widespread application. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 A method for detecting potential electrical hazards based on intelligent image recognition and 3D simulation includes the following steps: S1: Image data and environmental parameter data of power field equipment are collected by a mobile terminal equipped with a multispectral imaging module to form a multimodal input set. The multispectral imaging module simultaneously collects visible light, infrared thermal imaging and ultraviolet corona discharge data, and combines environmental sensors to obtain temperature, humidity and noise level parameters. S2: Adaptive preprocessing of the collected data, including data cleaning, feature extraction and quality assessment, wherein the preprocessing adopts an optical correction algorithm based on device type and an environmental parameter filtering chain to optimize data quality; S3: The preprocessed data is input into a deep convolutional neural network model based on the attention mechanism to achieve intelligent identification and classification of hidden danger features. The model adopts a multi-scale feature pyramid structure and embeds a spatial-channel dual attention module to focus on key areas.

[0019] It also includes: S4: The model output results drive the 3D simulation system to generate interactive training scenarios in real time, including equipment disassembly animation, hazard evolution simulation and standard handling process demonstration, in which the 3D simulation system realizes multi-field coupled rendering based on the physics engine; S5: Based on the identification results and simulation feedback, a structured hazard report is automatically generated and digital evidence is stored. The report generation uses knowledge graph technology to link historical cases, and the evidence storage process is based on blockchain to ensure that the data is tamper-proof.

[0020] Specifically, the process is as follows: First, data is collected via a mobile terminal equipped with a multispectral imaging module. This terminal integrates a visible light camera, an infrared thermal imager, and an ultraviolet sensor. The visible light camera employs high dynamic range imaging technology, automatically adjusting exposure and focus parameters to adapt to different lighting conditions, ensuring the capture of details on the device surface. The infrared thermal imager, based on the principle of microbolometers, detects the temperature distribution of the device and identifies overheating anomalies. The ultraviolet sensor is specifically designed to capture corona discharge phenomena, employing solar blind zone ultraviolet detection technology to eliminate sunlight interference. Environmental parameter acquisition is achieved through digital temperature and humidity sensors and a wideband noise detector. All sensor data is synchronized using a hardware mechanism (such as GPS timestamps) to ensure time consistency, forming a multimodal input set. Next, the collected data undergoes adaptive preprocessing: data cleaning uses wavelet transform denoising algorithms and histogram equalization to enhance contrast; feature extraction is tailored to the material characteristics of the equipment, with polarization imaging technology applied to metal surfaces (suppressing specular reflection through rotating polarization filters and Stokes vector calculations), and multi-scale texture analysis algorithms (such as Gabor filter banks to extract surface defects) used for insulating materials; quality assessment is based on a lightweight convolutional neural network model that monitors image sharpness and integrity in real time and provides feedback to optimize acquisition parameters. After preprocessing, the data is input into a deep convolutional neural network model based on an attention mechanism. This model adopts an improved ResNet architecture, embedding spatial attention and channel attention modules. Spatial attention focuses on key parts of the equipment, while channel attention optimizes feature weight allocation. Multi-scale features are fused through a feature pyramid structure to achieve intelligent identification and classification of potential hazards; model training employs a transfer learning strategy, fine-tuning on an electricity inspection dataset, and the output results support real-time inference.

[0021] Subsequently, the model output drives the 3D simulation system to generate interactive training scenarios. This system, developed based on the Unity3D engine, integrates a physics engine (such as NVIDIA PhysX) to simulate multi-field coupling effects (including heat conduction, electromagnetic fields, and mechanical interactions), generating equipment disassembly animations, hazard evolution simulations, and standard handling process demonstrations. The scenarios receive recognition results through a real-time data interface, dynamically adjusting the virtual content. Interactive training supports gesture recognition (using Leap Motion sensors to capture movements) and voice control (integrating a speech recognition API), enhancing immersion. Finally, based on the recognition results and simulation feedback, a structured hazard report is automatically generated. Report generation employs knowledge graph technology, using the Neo4j graph database to integrate historical equipment data, hazard cases, and rectification plans, and inferring the root causes of hazards through graph neural networks (such as GraphSAGE). The report engine automatically fills in fields based on a template design. Digital evidence storage is based on a blockchain framework (such as Hyperledger Fabric), using a hash chain structure to store report fingerprints, and ensuring data immutability and traceability through smart contracts. The entire solution, through the integration of multiple technologies, achieves a closed loop of detection, training, and management, improving the efficiency and reliability of power inspections.

[0022] Through multimodal data acquisition and intelligent identification processes, comprehensive automation of hazard detection in power inspections has been achieved. Beneficial effects include: improved data completeness and accuracy by reducing human error through multispectral imaging and simultaneous acquisition of environmental parameters; enhanced preprocessing adaptability and optimized data quality, providing reliable input for subsequent identification; and improved hazard identification efficiency, as the attention-based deep learning model can quickly classify defects, reducing the false negative rate. Overall, this solution provides an efficient and standardized detection method for power inspections, significantly improving operational safety and reliability.

[0023] Expanding 3D simulation and report generation capabilities, the system provides an immersive training experience through 3D simulation, helping trainees quickly master the operation procedures of complex equipment and reducing on-site training costs. Digital report generation combines knowledge graphs and blockchain notarization to ensure the traceability and tamper-proof nature of reports, enhancing data credibility. The overall system achieves a closed loop of testing, training, and management, optimizing resource allocation and promoting the digital transformation of the power industry.

[0024] The acquisition method of the multispectral imaging module in step S1 includes: Develop an adaptive acquisition strategy based on device operating status, and dynamically adjust imaging parameters, including resolution, frame rate and exposure time, according to device type to optimize data quality in different scenarios; A spatiotemporal registration algorithm for multi-source data is designed to solve the synchronization problem between visible light, infrared and ultraviolet data through feature point matching and coordinate transformation, thereby ensuring data alignment accuracy. A real-time monitoring mechanism for acquisition quality is introduced, which uses a deep learning model to evaluate image clarity and integrity, and automatically prompts for reacquisition or parameter adjustment.

[0025] Specifically, the process is as follows: First, a device feature database is established, containing optimal acquisition parameter configurations for various types of power equipment. For example, high infrared resolution mode is recommended for transformer equipment, while multi-angle ultraviolet acquisition strategy is suggested for switchgear equipment. The acquisition and control system dynamically adjusts imaging parameters based on the equipment's operating status, analyzes scene complexity through intelligent algorithms, and automatically optimizes exposure time (adjustable from 0.1 to 10 seconds), frame rate (1 to 30 fps), and gain parameters (0 to 40 dB).

[0026] The spatiotemporal registration system employs an improved feature point matching algorithm. Specifically, it utilizes an enhanced SIFT feature detector to extract stable feature points, removes mismatched points using the RANSAC algorithm, and calculates an accurate affine transformation matrix. Registration accuracy reaches the sub-pixel level, ensuring precise spatial alignment of multi-source data. The system also establishes a registration quality evaluation mechanism to verify the registration results in real time, automatically triggering a re-registration process when the registration error exceeds a threshold.

[0027] The quality monitoring mechanism integrates a lightweight convolutional neural network model (based on the MobileNetV2 architecture) to evaluate image quality metrics in real time, including sharpness, contrast, and noise levels. When a deterioration in image quality is detected, the system automatically prompts the operator to adjust the acquisition angle or parameter settings, and records the environmental conditions and equipment status during acquisition, providing data support for quality traceability.

[0028] The multispectral imaging acquisition method has been refined, with the following benefits: the adaptive acquisition strategy dynamically adjusts parameters according to the device type, improving the targeting and efficiency of data acquisition; the spatiotemporal registration algorithm ensures accurate alignment of multi-source data, enhancing data consistency; and the quality monitoring mechanism reduces acquisition errors and lowers the cost of repetitive work through real-time evaluation. These improvements make the data acquisition process more intelligent and reliable, laying a solid foundation for subsequent processing.

[0029] The adaptive preprocessing method described in step S2 includes: To address surface reflections on metal equipment, polarization imaging technology is used to suppress specular reflections, and histogram equalization is combined to enhance texture features. To address environmental noise, an adaptive filtering combination strategy is developed, which automatically selects either Kalman filtering or wavelet threshold denoising algorithm based on the noise spectrum characteristics. Establish a preprocessing effect evaluation system, quantify data quality through peak signal-to-noise ratio and structural similarity indicators, and provide feedback to optimize preprocessing parameters.

[0030] The above data preprocessing method establishes a three-stage processing pipeline. The first-stage quality assessment uses a deep learning-based quality detection network, which contains 5 convolutional layers and 3 fully connected layers, and can simultaneously output image sharpness score, contrast score, and noise level assessment results.

[0031] The second level of feature enhancement implements differentiated processing for different equipment materials: polarization imaging technology is used for metal surfaces, and the angle of the polarizer is adjusted by an electric rotating bracket (adjustable from 0-180°) to acquire multiple sets of images with different polarization directions, and Stokes vector calculation is used to suppress specular reflection; multi-scale texture analysis is applied to insulating materials, and Gabor filter banks (4 scales, 6 directions) are used to extract surface defect features.

[0032] The third-level data fusion employs the Laplacian pyramid algorithm, decomposing multi-source images into five scale spaces and performing adaptive weighted fusion based on feature importance. For environmental noise processing, an adaptive filtering combination strategy was developed: high-frequency noise is denoised using wavelet thresholding (emphasizing a soft thresholding function), while low-frequency noise is denoised using Kalman filtering (the state vector contains six dimensions).

[0033] The preprocessing performance evaluation system employs a multi-indicator approach, including traditional PSNR and SSIM metrics, as well as perception-based image quality evaluation metrics (using VGG networks for feature extraction). Evaluation results are displayed in real-time on a visual dashboard, providing operators with intuitive quality feedback.

[0034] Optimization of the preprocessing stage brings significant technical advantages: polarization imaging technology effectively suppresses reflections from metal surfaces, improving feature extraction accuracy; adaptive filtering strategies address different noise environments, enhancing data cleanliness; and the evaluation system uses quantitative indicators to optimize preprocessing parameters, ensuring process controllability. This helps reduce the risk of misjudgment and improves the robustness of hazard identification, making it particularly suitable for complex field environments.

[0035] The intelligent recognition method for the deep convolutional neural network model described in step S3 includes: A multi-level feature extraction network is constructed, and local details and global semantic information are fused through a feature pyramid to improve the robustness of hazard detection; Embedding deformable convolutional layers adapts to changes in the viewing angle and occlusion of device images, enhancing the model's generalization ability; Adversarial training techniques are employed to generate indistinguishable samples to optimize the model's decision boundary and improve the recognition rate of rare defects.

[0036] The backbone network is based on an improved ResNet-50 architecture, in which a dual attention mechanism is embedded: the spatial attention module uses convolutional layers to generate attention maps and focus on key parts of the device; the channel attention module uses the SE module to dynamically adjust feature weights.

[0037] The feature extraction network constructs a multi-scale feature pyramid (FPN), fusing features from different levels through top-down and bottom-up paths. Deformable convolutional layers are added to the network to learn offsets and adapt to changes in the device image's viewpoint and occlusion, significantly improving the model's adaptability to complex scenes.

[0038] The training process employs a progressive learning strategy, first pre-training on the ImageNet dataset and then fine-tuning using a power equipment-specific dataset. Adversarial training techniques are introduced during training, using Wasserstein GAN to generate indistinguishable samples to optimize the model's decision boundaries. Online hard sample mining techniques focus on difficult-to-classify samples, improving the model's ability to identify rare defects.

[0039] Several optimization techniques are employed during model deployment: knowledge distillation transfers knowledge from a large teacher model to a lightweight student model; model pruning removes redundant parameters, reducing computational complexity; and quantization converts floating-point weights into low-precision representations, improving inference speed. The final model is deployed on edge computing devices, supporting real-time inference.

[0040] By improving the intelligent recognition model, a multi-level feature extraction network enhances the model's ability to capture subtle defects, thereby increasing detection sensitivity; deformable convolutional layers improve the model's generalization ability, adapting to equipment diversity and changing perspectives; and adversarial training techniques optimize the decision boundary, reducing the false negative rate for rare defects. These effects collectively improve the system's accuracy and adaptability, meeting the high standards required for power equipment inspection.

[0041] The scene generation method of the 3D simulation system in step S4 includes: A multi-field coupling model of power equipment is established based on physical laws to simulate electromagnetic, thermal and mechanical interactions and realistically reproduce the fault evolution process. Develop a real-time data-driven mechanism to map recognition results to virtual scenes and dynamically adjust animation content and interaction logic; It integrates a natural interaction interface, supports gesture recognition and voice control, and allows users to operate virtual devices from multiple angles, enhancing the immersive experience of training.

[0042] The 3D simulation system is developed based on the Unity3D engine and uses the High Definition Render Pipeline (HDRP) to achieve realistic visual effects. The physics simulation integrates the NVIDIA PhysX engine to establish a multiphysics coupling model: the heat conduction simulation uses the finite element method with mesh generation accuracy down to the millimeter level; the electromagnetic field simulation is based on the finite-difference time-domain method with a calculation step size set to the nanosecond level; and the structural mechanics simulation uses finite element analysis to calculate the stress and strain distribution of the equipment under various working conditions.

[0043] The scene generation system includes a device disassembly animation module, which uses keyframe animation technology to showcase the internal structure and working principles of the equipment. The hazard evolution simulation is based on physical laws, realistically reproducing the fault development process, such as equipment deformation due to overheating or insulation damage caused by discharge. The standard handling procedure demonstration provides interactive guidance, showing the correct operating steps step by step.

[0044] The real-time data-driven mechanism establishes a connection with the recognition system via the WebSocket protocol, mapping the recognition results to the virtual scene in real time. The system supports dynamically adjusting animation content and interactive logic, displaying corresponding handling solutions based on different hazard types.

[0045] The natural interaction interface integrates a variety of advanced technologies: gesture recognition uses the Leap Motion sensor and identifies operation intentions through 3D skeleton tracking algorithms; voice control integrates Azure Speech Services and supports the recognition of mixed Chinese and English commands; force feedback devices provide 6 degrees of freedom tactile feedback to enhance the realism of operation.

[0046] Improving the generation of 3D simulation scenarios offers several benefits: multi-field coupled models realistically reproduce equipment failure evolution, helping users intuitively understand the causes of potential hazards; real-time data-driven mechanisms ensure synchronization between simulation and testing results, enhancing the timeliness of training; and natural interactive interfaces improve operational convenience and promote skill transfer. This not only improves training effectiveness but also reduces practical risks and supports rapid emergency response.

[0047] The structured hazard report generation method described in step S5 includes: Construct a knowledge graph for power equipment, integrate design specifications, historical hazards, and rectification plans, and use graph neural networks to infer the root causes of hazards. The design report automatic generation engine populates key fields based on templates and real-time data, and recommends optimization and rectification measures; Implement a blockchain-based evidence storage process, using a hash chain structure to store report fingerprints, supporting distributed verification and traceability.

[0048] The aforementioned report generation system constructs a complete knowledge graph system. The knowledge graph is stored in the Neo4j graph database, containing over 500,000 entity nodes and over 2 million relationship edges. The graph construction adopts a bottom-up approach: first, entities and relationships are extracted from structured data sources (equipment ledgers, inspection records); then, natural language processing technology is used to extract knowledge from unstructured text; and finally, the BERT model is used for named entity recognition.

[0049] The graph neural network inference uses the GraphSAGE algorithm, which achieves intelligent reasoning of the root cause of potential hazards through neighbor sampling (50 samples) and feature aggregation. The inference process considers multi-dimensional factors, including the equipment's historical status, environmental conditions, and operating parameters, to provide accurate diagnostic results.

[0050] The report generation engine employs a template-based design, supporting multiple output formats such as Word and PDF. The template library includes over 20 standard report templates, and user-defined template design is also supported. The intelligent fill module dynamically generates report content based on real-time data and automatically recommends optimization and rectification measures.

[0051] The blockchain-based evidence storage system is developed based on the Hyperledger Fabric framework and employs a Byzantine fault-tolerant consensus mechanism. The evidence storage process includes: generating a report SHA-256 hash value, storing the hash value in a distributed ledger, and automatically verifying data integrity through smart contracts. The system provides complete traceability functionality, supporting the querying and verification of historical records.

[0052] The aforementioned optimized report generation method integrates multi-source data using a knowledge graph, providing intelligent reasoning to identify the root causes of potential problems and assisting in decision optimization; a templated report engine automates the generation of structured content, reducing the burden of manual compilation; and blockchain-based evidence storage ensures data security and compliance, meeting industry regulatory requirements. This enhances the transparency and efficiency of the entire inspection process and promotes standardized management.

[0053] Example 2 Based on Example 1: The model described in step S3 further includes an interpretability analysis module: Heatmaps are generated by gradient-weighted activation mapping to visualize the regions of interest of the model and provide a basis for decision-making. Develop a feature importance ranking algorithm to identify key features that affect classification results and enhance model transparency; By combining expert knowledge bases, semantic interpretation of model outputs is performed to reduce the risk of false alarms.

[0054] The aforementioned interpretability analysis module develops a web-based visualization platform. The platform integrates the Gradient Weighted Class Activation Mapping (Grad-CAM) algorithm, which generates high-resolution visualizations of attention regions by calculating the gradient weights of convolutional layers. The heatmap uses color coding to represent the level of attention, with red indicating highly attentional regions and blue indicating low-attention regions.

[0055] Feature importance analysis integrates multiple analytical methods: SHAP value calculation, based on game theory principles, quantifies the contribution of each feature to the prediction result; the LIME local interpretation method generates a locally interpretable linear model by perturbing the input data. Analysis results are presented through interactive charts, allowing users to explore the impact of features in depth.

[0056] The expert knowledge base integrates professional knowledge in the power industry, containing over 200 detection rules and rules of thumb. The rule engine uses the Drools framework; when a discrepancy arises between the model's predictions and the expert rules, the system automatically triggers an early warning mechanism, prompting professionals to intervene and analyze the issue. The knowledge base supports continuous updates, constantly optimizing the rule set through machine learning.

[0057] The system also provides model performance monitoring capabilities, tracking the model's performance in real-time, including metrics such as accuracy and recall. When a decline in model performance is detected, it automatically prompts for retraining or parameter adjustment.

[0058] An interpretability analysis module has been added, the core effects of which are: heatmap visualization of the model's focus areas, enhancing user trust in AI decision-making; transparent analysis of feature importance ranking, helping to pinpoint key factors; and integration with an expert knowledge base to reduce false alarm rates and improve system usability. This makes the model not only efficient but also easy to understand and verify, meeting the reliability requirements of the power safety field.

[0059] The 3D simulation system described in step S4 supports personalized training and optimization. Establish a trainee competency assessment model, quantify skill levels through operational data, and dynamically adjust training difficulty and content; Design a path planning algorithm based on reinforcement learning to recommend the best learning sequence based on the learner's progress; Enables multi-student collaborative training mode, supports role assignment and team tasks, and improves collaboration efficiency; In step S5, the digital evidence storage method includes: The model is updated using a federated learning framework, and the recognition algorithm is optimized using distributed data, while protecting user privacy. Develop an automatic verification mechanism for smart contracts to check data compliance during the evidence storage process and reduce manual intervention; Establish a data lifecycle management system for evidence storage, regularly archive and clean up expired information, and ensure the efficient operation of the system.

[0060] The training optimization system establishes a multi-dimensional learner competency assessment model. Assessment indicators include operational accuracy (number of errors), efficiency (completion time), and compliance with safety regulations. The system uses a clustering algorithm (K-means) to group learners and dynamically adjusts the training difficulty and content based on their competency levels.

[0061] Personalized learning path planning employs a reinforcement learning algorithm (Q-learning) to recommend the optimal learning sequence based on the learner's real-time performance. The system establishes a reward mechanism, providing positive feedback for excellent performance and additional training content for weak areas.

[0062] The multi-student collaborative training mode supports role assignment and team tasks. The system provides a virtual command center where students take on different roles (such as inspectors, recorders, and safety officers) to collaboratively complete complex tasks. The collaboration process utilizes WebRTC technology to achieve real-time audio and video communication and data synchronization.

[0063] The digital evidence storage system employs a federated learning framework, where each edge node trains its model locally and only uploads the model parameters to the central server. Privacy protection utilizes differential privacy technology, adding noise (with a noise standard deviation set to 0.1) to the model parameters to ensure data security.

[0064] The smart contracts are developed on the Ethereum platform and written in Solidity. Contract functionalities include: automatic data integrity verification, access control management, and execution of notarization rules. Contract gas consumption has been optimized to below 50,000 units, improving execution efficiency. A lifecycle management system establishes a data archiving strategy, periodically migrating historical data to low-cost storage media. A cleanup strategy, based on data value and timeliness, automatically identifies and deletes expired information, ensuring efficient system operation.

[0065] The aforementioned solutions encompass training optimization and digital record keeping, generating comprehensive benefits: personalized training systems improve the efficiency of learner skills development through competency assessment and pathway planning; the federated learning framework protects data privacy and supports cross-institutional collaboration; and automated verification of smart contracts reduces human intervention and optimizes resource utilization. These effects promote the intelligent allocation of training resources and data security, making them suitable for large-scale deployment.

[0066] The method also includes a cloud-edge collaborative processing architecture: Deploy lightweight models at the edge to achieve low-latency data collection and preliminary identification; Perform complex model training and big data analysis in the cloud, supporting global system optimization; Design a dynamic task scheduling algorithm to allocate computing resources based on network conditions and ensure system reliability.

[0067] The collaborative architecture in this embodiment adopts a layered design. Lightweight models are deployed on edge computing nodes, and multiple techniques are used for model optimization: model pruning removes redundant connections while retaining important parameters; weight quantization converts 32-bit floating-point numbers into 8-bit integers to reduce storage requirements; and operator fusion merges multiple computational operations to improve execution efficiency.

[0068] The cloud-based training platform is built as a distributed cluster, using Kubernetes for resource management. The cluster contains 10 compute nodes, each equipped with 8 GPU cards. Training task scheduling uses a fair sharing algorithm to ensure fair resource allocation. The model version management system supports the traceability and reproducibility of the training process.

[0069] The dynamic task scheduling algorithm is developed based on reinforcement learning and monitors the computational load, network bandwidth, and energy consumption of each node in real time. The scheduling strategy considers multiple optimization objectives: minimizing response time, maximizing resource utilization, and balancing load distribution. The algorithm continuously optimizes scheduling decisions through Q-learning.

[0070] The security protection system implements defense-in-depth: the network layer uses IPSec VPN to encrypt data transmission; the application layer implements two-factor authentication and role-based access control; the data layer uses AES-256 encrypted storage and implements data anonymization. The auditing system records all operation logs, supporting the tracing and analysis of security incidents.

[0071] Through a cloud-edge collaborative architecture, system-level optimization is achieved: lightweight edge models ensure real-time processing capabilities and reduce latency; big data analytics in the cloud support continuous model evolution, improving system intelligence; and dynamic scheduling algorithms enhance resource adaptability, ensuring stable operation. This makes the solution scalable and reliable, suitable for power application scenarios of different scales.

[0072] The above-described embodiments are detailed and specific, illustrating preferred embodiments of the present invention. They are only used to illustrate the technical ideas and features of the present invention, with the aim of enabling those skilled in the art to understand the content of the present invention and implement it accordingly. However, they are not limited to the present invention, and the patent scope of the present invention cannot be limited by this embodiment alone. That is, any equivalent changes or modifications made to the spirit disclosed in the present invention, without departing from the structure of the present invention, such as local improvements within the system and modifications or transformations between subsystems, are still within the patent scope of the present invention.

Claims

1. A method for detecting potential electrical hazards based on intelligent image recognition and 3D simulation, characterized in that, Includes the following steps: S1: Image data and environmental parameter data of power field equipment are collected by a mobile terminal equipped with a multispectral imaging module to form a multimodal input set. The multispectral imaging module simultaneously collects visible light, infrared thermal imaging and ultraviolet corona discharge data, and combines environmental sensors to obtain temperature, humidity and noise level parameters. S2: Adaptive preprocessing of the collected data, including data cleaning, feature extraction and quality assessment, wherein the preprocessing adopts an optical correction algorithm based on device type and an environmental parameter filtering chain to optimize data quality; S3: The preprocessed data is input into a deep convolutional neural network model based on the attention mechanism to achieve intelligent identification and classification of hidden danger features. The model adopts a multi-scale feature pyramid structure and embeds a spatial-channel dual attention module to focus on key areas.

2. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 1, characterized in that, Also includes: S4: The model output results drive the 3D simulation system to generate interactive training scenarios in real time, including equipment disassembly animation, hazard evolution simulation and standard handling process demonstration. The 3D simulation system is based on a physics engine to achieve multi-field coupled rendering. S5: Based on the identification results and simulation feedback, a structured hazard report is automatically generated and digital evidence is stored. The report generation uses knowledge graph technology to link historical cases, and the evidence storage process is based on blockchain to ensure that the data is tamper-proof.

3. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The acquisition method of the multispectral imaging module in step S1 includes: Develop an adaptive acquisition strategy based on device operating status, and dynamically adjust imaging parameters, including resolution, frame rate and exposure time, according to device type to optimize data quality in different scenarios; A spatiotemporal registration algorithm for multi-source data is designed to solve the synchronization problem between visible light, infrared and ultraviolet data through feature point matching and coordinate transformation, thereby ensuring data alignment accuracy. A real-time monitoring mechanism for acquisition quality is introduced, which uses a deep learning model to evaluate image clarity and integrity, and automatically prompts for reacquisition or parameter adjustment.

4. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The adaptive preprocessing method described in step S2 includes: To address surface reflections on metal equipment, polarization imaging technology is used to suppress specular reflections, and histogram equalization is combined to enhance texture features. To address environmental noise, an adaptive filtering combination strategy is developed, which automatically selects either Kalman filtering or wavelet threshold denoising algorithm based on the noise spectrum characteristics. Establish a preprocessing effect evaluation system, quantify data quality through peak signal-to-noise ratio and structural similarity indicators, and provide feedback to optimize preprocessing parameters.

5. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The intelligent recognition method for the deep convolutional neural network model described in step S3 includes: A multi-level feature extraction network is constructed, and local details and global semantic information are fused through a feature pyramid to improve the robustness of hazard detection; Embedding deformable convolutional layers adapts to changes in the viewing angle and occlusion of device images, enhancing the model's generalization ability; Adversarial training techniques are employed to generate indistinguishable samples to optimize the model's decision boundary and improve the recognition rate of rare defects.

6. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The scene generation method of the 3D simulation system in step S4 includes: A multi-field coupling model of power equipment is established based on physical laws to simulate electromagnetic, thermal and mechanical interactions and realistically reproduce the fault evolution process. Develop a real-time data-driven mechanism to map recognition results to virtual scenes and dynamically adjust animation content and interaction logic; It integrates a natural interaction interface, supports gesture recognition and voice control, and allows users to operate virtual devices from multiple angles, enhancing the immersive experience of training.

7. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The structured hazard report generation method described in step S5 includes: Construct a knowledge graph for power equipment, integrate design specifications, historical hazards, and rectification plans, and use graph neural networks to infer the root causes of hazards. The design report automatic generation engine populates key fields based on templates and real-time data, and recommends optimization and rectification measures; Implement a blockchain-based evidence storage process, using a hash chain structure to store report fingerprints, supporting distributed verification and traceability.

8. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The model described in step S3 also includes an interpretability analysis module: Heatmaps are generated by gradient-weighted activation mapping to visualize the regions of interest of the model and provide a basis for decision-making. Develop a feature importance ranking algorithm to identify key features that affect classification results and enhance model transparency; By combining expert knowledge bases, semantic interpretation of model outputs is performed to reduce the risk of false alarms.

9. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The 3D simulation system described in step S4 supports personalized training and optimization. Establish a trainee competency assessment model, quantify skill levels through operational data, and dynamically adjust training difficulty and content; Design a path planning algorithm based on reinforcement learning to recommend the best learning sequence based on the learner's progress; Enables multi-student collaborative training mode, supports role assignment and team tasks, and improves collaboration efficiency; In step S5, the digital evidence storage method includes: The model is updated using a federated learning framework, and the recognition algorithm is optimized using distributed data, while protecting user privacy. Develop an automatic verification mechanism for smart contracts to check data compliance during the evidence storage process and reduce manual intervention; Establish a data lifecycle management system for evidence storage, regularly archive and clean up expired information, and ensure the efficient operation of the system.

10. The method for detecting potential electrical hazards based on intelligent image recognition and three-dimensional simulation according to claim 2, characterized in that, The method also includes a cloud-edge collaborative processing architecture: Deploy lightweight models at the edge to achieve low-latency data collection and preliminary identification; Perform complex model training and big data analysis in the cloud, supporting global system optimization; Design a dynamic task scheduling algorithm to allocate computing resources based on network conditions and ensure system reliability.