Power plant employee violation intelligent identification method and system based on AI large model

Abstract

The invention relates to the technical field of electric power operation, and discloses a power plant employee violation intelligent identification method and system based on an AI large model, and the system comprises a visual collection module, an edge calculation node, a personnel positioning and tracking module, a three-dimensional scene reconstruction module, a multi-modal analysis module, a grading early warning module, a self-adaptive adjustment module, and an inspection robot. According to the invention, accurate identification of personnel behaviors in a complex industrial environment is realized, the problems of high false alarm rate, response lag and the like of a traditional monitoring system are effectively solved through a specially designed self-adaptive adjustment mechanism and a grading early warning strategy, and the technical scheme has the advantages that through innovative integration of technologies of machine vision, edge calculation, multi-modal analysis and the like, the safety of the system is improved. An intelligent safety management and control system with an autonomous optimization capability is constructed, and breakthrough progress is made in core indexes such as identification precision, response speed, resource utilization efficiency and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of power operations, and more specifically, to a method and system for intelligent identification of violations by power plant employees based on an AI large model. Background Art

[0002] The power generation industry is a technology, capital, and labor-intensive industry. The operating environment of power plants is complex, with many risk factors, prone to safety accidents, and the consequences of safety accidents are serious and the losses are huge.

[0003] In the field of power production, the power plant safety control system is the core infrastructure for ensuring personnel safety and stable operation of equipment.

[0004] Traditional monitoring methods mainly rely on the combination of a fixed camera network and manual inspections, with multiple technical bottlenecks: First, in a complex industrial environment, interference factors cause serious degradation of video image quality, and the single visible light imaging technology used in existing systems is difficult to accurately capture the details of personnel operations; Second, the positioning technology has a large error and cannot meet the precise control requirements of high-risk areas. Research shows that more than 90% of power plant safety accidents occur during production operations, and more than 85% of them are related to human unsafe behavior factors. There are many outsourcing engineering projects such as infrastructure construction, technological transformation, and maintenance in power plants every year. The number of outsourced personnel is large and the personnel mobility is high. How to strictly control the safety of outsourced personnel access, strengthen risk prevention, hidden danger investigation, and safety control at the operation site, and eliminate potential safety hazards in production from the source is crucial for building an intrinsically safe power plant.

[0005] Therefore, how to provide a method and system for intelligent identification of violations by power plant employees based on an AI large model is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for intelligent identification of violations by power plant employees based on an AI large model to solve the problems raised in the above background art section.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] An intelligent identification system for violations by power plant employees based on an AI large model, comprising: a visual acquisition module, an edge computing node, a personnel positioning and tracking module, a three-dimensional scene reconstruction module, a multimodal analysis module, a hierarchical early warning module, an adaptive adjustment module, and an inspection robot;

[0009] Wherein, the visual acquisition module forms a dynamic perception network covering the fuel transportation channel and the operation area through a rotatable high-definition camera group, and the output end of the visual acquisition module is connected to the edge computing node for background difference processing;

[0010] The edge computing node establishes a data channel with the personnel positioning and tracking module through an optical fiber network. The personnel positioning and tracking module performs spatio-temporal alignment on the UWB positioning data and the video feature matching result, and inputs the generated three-dimensional coordinate data stream into the three-dimensional scene reconstruction module to construct a virtual power plant model;

[0011] The virtual power plant model establishes two-way data interaction with the multimodal analysis module, and the abnormal behavior recognition result of the multimodal analysis module triggers the corresponding response mechanism of the hierarchical warning module;

[0012] The adaptive adjustment module dynamically controls the working parameters of the visual acquisition module, and the adaptive adjustment module receives real-time personnel distribution data from the personnel positioning and tracking module.

[0013] The rotatable high-definition camera groups deployed by the visual acquisition module form a dynamic perception network covering the fuel transportation channel and the core operation area. Combined with the real-time preprocessing of the edge computing node, the processing delay of the original video data is reduced to within 200 ms, solving the problem of monitoring lag caused by data transmission delay in the traditional system; The personnel positioning and tracking module uses the spatio-temporal alignment technology of UWB positioning and visual feature matching to reduce the three-dimensional coordinate positioning error from the traditional 20 cm to ±5 cm, effectively eliminating the spatial offset in personnel trajectory tracking; The three-dimensional scene reconstruction module generates an interactive virtual power plant model through multi-view image fusion, and maps the device coordinates and personnel positions in real time, improving the global perception efficiency of monitoring personnel for complex scenes; The linkage mechanism of the multimodal analysis module and the hierarchical warning module automatically matches the response strategy according to the type of violation, shortening the disposal time limit of high-risk violation events from 5 minutes to 30 seconds; The adaptive adjustment module dynamically adjusts the camera group parameters according to the personnel density, improving the image resolution in high-risk areas and reducing the bandwidth occupancy in low-risk areas, achieving the optimal allocation of monitoring resources.

[0014] Preferably, in the above-mentioned intelligent identification system for power plant employee violations based on the AI large model, the dynamic perception network is provided with dynamic reference points, and the dynamic reference points include a reference locator and an environmental sensor group;

[0015] The reference locator is composed of a laser rangefinder and a radio frequency identification device; The environmental sensor group integrates temperature and humidity, gas concentration, and vibration detection units.

[0016] Preferably, in the above-mentioned intelligent identification system for employees' violations in power plants based on the AI large model, the dynamic reference points are arranged at intervals along the axis of the conveyor belt. The dynamic reference points are provided with reference locators, and the laser ranging signals emitted by the reference locators and the UWB positioning chip group form a composite positioning field. The environmental data collected by the environmental sensor group is uploaded to the edge computing node through the communication interface, where the edge computing node controls the supplementary light intensity of the camera group according to the change of environmental illumination.

[0017] The dynamic reference points construct a composite positioning field through a laser rangefinder and a radio frequency identification device, which triples the anti-interference ability of the UWB positioning signal and still maintains centimeter-level positioning accuracy in a strong electromagnetic environment; the temperature and humidity data collected by the environmental sensor group and the control parameters of the camera group form a closed-loop regulation. For example, when the dust concentration exceeds 50mg / m 3 ³, the image noise reduction algorithm is automatically started to improve the image clarity; the spacing of the reference points arranged along the conveyor belt is optimized to 10 meters, and combined with the data fusion of the edge computing node, the full-process monitoring coverage rate of the fuel transportation path is improved.

[0018] Preferably, in the above-mentioned intelligent identification system for employees' violations in power plants based on the AI large model, the edge computing node includes an image preprocessing unit, a data compression unit, and a cache management unit;

[0019] The image preprocessing unit implements background difference and light compensation algorithms; the data compression unit uses adaptive coding technology to reduce the transmission bandwidth; the cache management unit is configured with a dual-channel storage architecture to process real-time stream data and historical comparison data respectively.

[0020] Preferably, in the above-mentioned intelligent identification system for employees' violations in power plants based on the AI large model, the motion target detection results implemented by the image preprocessing unit generate a low-bitrate video stream through the data compression unit. This video stream is distributed to the dual channels of the local memory and the cloud server through the cache management unit, where the local memory retains the video segments of the most recent 15 minutes for real-time behavior analysis, and the historical data stored in the cloud server is used to train the equipment safety operation boundary of the virtual power plant model.

[0021] The background difference algorithm of the image preprocessing unit combined with adaptive light compensation improves the accuracy of motion target detection and effectively eliminates false detections caused by shadows and reflections; the data compression unit uses H.265 encoding and region of interest enhancement technology, while ensuring the image quality of key regions, reducing the overall video stream bandwidth occupancy by 60%; the dual-channel architecture of the cache management unit shortens the historical data retrieval time required for real-time behavior analysis from 3 seconds to 0.5 seconds, and at the same time meets the integrity requirements of accident traceability through a 15-minute rolling storage mechanism.

[0022] Preferably, in the above intelligent identification system for power plant employees' violations based on an AI large model, the personnel positioning and tracking module includes an inertial navigation unit, a trajectory prediction unit, and a safety area mapping unit. The inertial navigation unit integrates a gyroscope and an acceleration sensor; the trajectory prediction unit establishes a personnel movement path prediction model based on the Markov chain model; the safety area mapping unit divides the virtual power plant model into working partitions with different risk levels.

[0023] Preferably, in the above intelligent identification system for power plant employees' violations based on an AI large model, the personnel positioning and tracking module achieves precise positioning in the following way:

[0024] The acceleration data collected by the inertial navigation unit compensates for the transmission delay of the UWB positioning signal. The trajectory prediction unit generates a personnel movement path probability map based on the corrected positioning coordinates. After the personnel movement path probability map is input into the safety area mapping unit, it is subjected to spatial overlay analysis with the high-risk operation area in the virtual power plant model, and the analysis result is dynamically updated to the risk level assessment unit of the hierarchical early warning module.

[0025] The gyroscope data of the inertial navigation unit compensates for the transmission delay of the UWB positioning signal, reducing the trajectory tracking error when the personnel accelerates suddenly or turns sharply to 2 cm; the trajectory prediction unit generates a movement path probability map for the next 30 seconds based on the Markov chain model, increasing the prediction accuracy of the personnel's illegal entry into high-risk areas from 75% to 90%; the safety area mapping unit divides the virtual model into three-level risk areas. When the personnel stays in the red area for more than the preset duration, a directional sound and light alarm is triggered, increasing the disposal efficiency of illegal stay incidents by 3 times.

[0026] Preferably, in the above intelligent identification system for power plant employees' violations based on an AI large model, the adaptive adjustment module includes a density analysis unit, an optical parameter optimization unit, and a network load balancing unit. The density analysis unit calculates the personnel aggregation degree in the area based on the video stream; the optical parameter optimization unit dynamically adjusts the exposure parameters and fill light intensity of the camera; the network load balancing unit allocates edge computing resources according to the data transmission requirements;

[0027] The density analysis unit controls the optical parameter optimization unit to increase the camera frame rate in the high-density area to 60 fps based on the heat map output by the personnel positioning and tracking module. At the same time, the network load balancing unit adaptively reduces the video stream resolution in the low-risk area to 720p, where the focal length adjustment strategy is dynamically associated with the spatial coordinates of the equipment safety operation boundary in the three-dimensional scene reconstruction module.

[0028] The density analysis unit automatically divides the monitoring priority based on the crowd density heat map of the video stream, improving the capture rate of the camera group for group violations; the optical parameter optimization unit automatically switches to the infrared mode and adjusts the gain parameters in low-light environments, increasing the proportion of effective images in night monitoring; the network load balancing unit dynamically allocates edge computing resources according to the risk level, improving the overall resource utilization rate of the system and reducing the hardware investment cost.

[0029] Preferably, in the above intelligent identification system for employee violations in power plants based on the AI large model, the inspection robot is equipped with a multi-spectral imager and a gas detector; and the inspection robot is provided with a robotic arm capable of remotely operating equipment in dangerous areas; the inspection robot is provided with an autonomous navigation unit that combines the SLAM algorithm and UWB positioning information to plan the inspection path;

[0030] In the path planned by the autonomous navigation unit of the inspection robot, it can collect multi-spectral images of dangerous areas. The collected data is transmitted back to the multi-modal analysis module via a 5G private network. When a gas leak is detected, the historical data playback layer of the virtual power plant model is triggered to retrieve the disposal records of similar events, and the simulation training layer is driven to generate an emergency disposal virtual scenario.

[0031] The combination of the multi-spectral imager and the gas detector carried by the inspection robot improves the methane leak detection sensitivity in the coal conveying corridor to 0.5 ppm, reducing the detection blind area compared with traditional fixed sensors; the six-axis robotic arm of the robotic arm can remotely operate the valve switch, increasing the maintenance efficiency of equipment in dangerous areas by 4 times; the SLAM algorithm of the autonomous navigation unit integrates UWB positioning data, shortening the inspection path planning time from 30 minutes to 5 minutes and improving the obstacle avoidance success rate.

[0032] An intelligent identification method for employee violations in power plants based on the AI large model, the specific method is as follows:

[0033] Deploy a rotatable high-definition camera group and a UWB positioning base station to form a dynamic perception network covering the entire plant area. The visual acquisition module uses the background difference algorithm to extract moving targets and realizes centimeter-level accurate three-dimensional trajectory tracking with UWB positioning data through the spatio-temporal alignment algorithm; the environmental sensor group collects parameters such as temperature, humidity, and vibration in real time to construct a baseline database for the operating state of equipment;

[0034] The edge computing node performs preprocessing on the video stream. After compressing the data volume by 60% through adaptive coding technology, it is transmitted to the local real-time analysis unit and the cloud training platform in channels; the personnel positioning data is input into the trajectory prediction model to generate a moving path probability map after inertial navigation compensation, and collision detection is performed with the safety boundary in the three-dimensional virtual model to achieve early warning of violation behaviors;

[0035] The multi-modal analysis module adopts a feature-level fusion strategy to spatially register visual behavior features, infrared temperature distribution, and acoustic fingerprint spectrum features, and constructs a violation determination matrix with spatio-temporal correlation. The hierarchical early warning module dynamically selects response strategies based on the risk index and triggers a progressive control from acoustic-optical warning to equipment interlock.

[0036] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an intelligent method and system for identifying violations of power plant employees based on an AI large model, which has the following beneficial effects:

[0037] Through the collaborative operation of the visual acquisition module, UWB positioning system, and multi-modal analysis module, the problem of inconsistent spatio-temporal benchmarks existing in traditional monitoring systems is effectively solved. The composite positioning field established by the dynamic reference point improves the positioning accuracy of personnel. Combining with the real-time data preprocessing technology of edge computing nodes, the spatio-temporal alignment error between the video stream and positioning data is reduced. The multi-modal analysis module adopts a feature-level fusion strategy, and the comprehensive recognition accuracy is improved.

[0038] The closed-loop control system constructed by the adaptive adjustment module realizes the intelligent allocation of monitoring resources. The virtual power plant model constructed by the three-dimensional scene reconstruction module and the three-level response mechanism established by the hierarchical early warning module achieve precise control through the risk index dynamic evaluation model.

[0039] The multi-spectral imager carried by the inspection robot reduces the detection blind area of dangerous areas by 92%. The fusion of the SLAM navigation algorithm and UWB positioning effectively improves the inspection path planning efficiency. The system as a whole realizes full-time seamless monitoring of personnel behavior, equipment status, and fuel safety, and reduces the incidence of major safety accidents.

[0040] Through the innovative integration of technologies such as machine vision, edge computing, and multi-modal analysis, this technical solution constructs an intelligent safety control system with autonomous optimization capabilities, and has made breakthrough progress in core indicators such as recognition accuracy, response speed, and resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to the provided drawings.

[0042] Figure 1 The drawings are the structural framework diagrams of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] An intelligent system for identifying violations of power plant employees based on an AI large model is disclosed in an embodiment of the present invention, including: a visual acquisition module, an edge computing node, a personnel positioning and tracking module, a three-dimensional scene reconstruction module, a multimodal analysis module, a hierarchical warning module, an adaptive adjustment module, and an inspection robot;

[0045] Among them, the visual acquisition module of the inspection robot forms a dynamic perception network covering the fuel transportation channel and the operation area through a rotatable high-definition camera group, and the output end of the visual acquisition module of the inspection robot is connected to the edge computing node of the inspection robot for background difference processing;

[0046] The edge computing node of the inspection robot establishes a data channel with the personnel positioning and tracking module of the inspection robot through an optical fiber network. The personnel positioning and tracking module of the inspection robot aligns the UWB positioning data and the video feature matching result in space and time, and the generated three-dimensional coordinate data stream is input into the three-dimensional scene reconstruction module of the inspection robot to construct a virtual power plant model;

[0047] A two-way data interaction is established between the virtual power plant model of the inspection robot and the multimodal analysis module of the inspection robot. Among them, the abnormal behavior recognition result of the multimodal analysis module of the inspection robot triggers the corresponding response mechanism of the hierarchical warning module of the inspection robot;

[0048] The adaptive adjustment module of the inspection robot dynamically controls the working parameters of the visual acquisition module of the inspection robot, and the adaptive adjustment module of the inspection robot receives the real-time personnel distribution data from the personnel positioning and tracking module of the inspection robot.

[0049] The rotatable high-definition camera group deployed by the visual acquisition module forms a dynamic perception network covering the fuel transportation channel and the core operation area. Combining the real-time preprocessing of the edge computing node (such as background difference and light compensation), the processing delay of the original video data is reduced to within 200 ms, solving the problem of monitoring lag caused by data transmission delay in the traditional system; the personnel positioning and tracking module adopts the spatio-temporal alignment technology of UWB positioning and visual feature matching, reducing the three-dimensional coordinate positioning error from the traditional 20 cm to ±5 cm, effectively eliminating the spatial offset in personnel trajectory tracking; the three-dimensional scene reconstruction module generates an interactive virtual power plant model through multi-view image fusion, mapping the equipment coordinates and personnel positions in real time, improving the global perception efficiency of the monitoring personnel for complex scenes; the linkage mechanism of the multi-modal analysis module and the hierarchical warning module automatically matches the response strategy (acoustic and optical alarm / equipment interlock) according to the type of violation (such as not wearing a safety helmet, approaching high-temperature equipment illegally), shortening the disposal time of high-risk violation events from 5 minutes to 30 seconds; the adaptive adjustment module dynamically adjusts the camera group parameters according to the personnel density, improving the image resolution in high-risk areas while reducing the bandwidth occupancy in low-risk areas, realizing the optimal allocation of monitoring resources.

[0050] In order to further optimize the above technical solution, the dynamic perception network of the inspection robot is provided with a dynamic reference point, and the dynamic reference point of the inspection robot includes a reference locator and an environmental sensor group;

[0051] The reference locator of the inspection robot is composed of a laser rangefinder and a radio frequency identification device; the environmental sensor group of the inspection robot integrates temperature and humidity, gas concentration and vibration detection units.

[0052] In order to further optimize the above technical solution, the dynamic reference points of the inspection robot are arranged at intervals along the axis of the conveyor belt. The reference locator of the inspection robot emits a laser ranging signal to form a composite positioning field with the UWB positioning chip group. The environmental data collected by the environmental sensor group of the inspection robot is uploaded to the edge computing node of the inspection robot through the communication interface, where the edge computing node of the inspection robot controls the fill light intensity of the camera group according to the change of environmental illumination.

[0053] The dynamic reference point constructs a composite positioning field through a laser rangefinder and a radio frequency identification device, increasing the anti-interference ability of the UWB positioning signal by 3 times and still maintaining centimeter-level positioning accuracy in a strong electromagnetic environment; the temperature and humidity data collected by the environmental sensor group and the control parameters of the camera group form a closed-loop regulation. For example, when the dust concentration exceeds 50 mg / m 3 ³, the image noise reduction algorithm is automatically started, improving the image clarity by 40%; the interval of the reference points arranged along the conveyor belt is optimized to 10 meters, combined with the data fusion of the edge computing node, improving the full-process monitoring coverage rate of the fuel transportation path.

[0054] To further optimize the above technical solution, the edge computing node of the patrol robot includes an image preprocessing unit, a data compression unit, and a cache management unit;

[0055] The patrol robot image preprocessing unit implements background difference and light compensation algorithms; the patrol robot data compression unit uses adaptive coding technology to reduce the transmission bandwidth; the patrol robot cache management unit is configured with a dual-channel storage architecture to process real-time stream data and historical comparison data respectively.

[0056] To further optimize the above technical solution, the motion target detection results implemented by the patrol robot image preprocessing unit generate a low-bitrate video stream through the data compression unit. This video stream is distributed to the dual channels of the local memory and the cloud server by the cache management unit. Among them, the local memory retains the video segments of the most recent 15 minutes for real-time behavior analysis, and the historical data stored in the cloud server is used to train the device safety operation boundary of the virtual power plant model.

[0057] The background difference algorithm of the image preprocessing unit combined with adaptive light compensation improves the accuracy of motion target detection and effectively eliminates false detections caused by shadows and reflections; the data compression unit uses H.265 encoding and region of interest (ROI) enhancement technology to reduce the overall video stream bandwidth occupancy by 60% while ensuring the image quality of the key area (bitrate > 8Mbps); the dual-channel architecture (local SSD + cloud storage) of the cache management unit shortens the historical data retrieval time required for real-time behavior analysis from 3 seconds to 0.5 seconds, and at the same time meets the integrity requirements of accident traceability through the 15-minute rolling storage mechanism.

[0058] To further optimize the above technical solution, the personnel positioning and tracking module of the patrol robot includes an inertial navigation unit, a trajectory prediction unit, and a safety area mapping unit. The inertial navigation unit integrates a gyroscope and an acceleration sensor; the trajectory prediction unit establishes a personnel movement path prediction model based on the Markov chain model; the safety area mapping unit divides the virtual power plant model into working partitions with different risk levels.

[0059] To further optimize the above technical solution, the personnel positioning and tracking module of the patrol robot achieves precise positioning through the following methods:

[0060] The acceleration data collected by the inertial navigation unit compensates for the transmission delay of the UWB positioning signal. The trajectory prediction unit generates a personnel movement path probability map based on the corrected positioning coordinates. After the personnel movement path probability map is input into the safety area mapping unit, it performs spatial overlay analysis with the high-risk operation area in the virtual power plant model, and the analysis results are dynamically updated to the risk level assessment unit of the hierarchical warning module.

[0061] The gyroscope data of the inertial navigation unit compensates for the UWB positioning signal transmission delay (with a compensation accuracy of 0.1 ms), reducing the trajectory tracking error during rapid acceleration (>2 m / s 2 ) or sharp turning of personnel to 2 cm; the trajectory prediction unit generates a moving path probability map for the next 30 seconds based on the Markov chain model, increasing the prediction accuracy of personnel illegally entering high-risk areas (such as high-voltage switchrooms) from 75% to 90%; the safe area mapping unit divides the virtual model into three levels of risk areas (red / yellow / green), and when a person stays in the red area for more than a preset duration (such as 30 seconds), a directional audible and visual alarm is triggered, increasing the disposal efficiency of illegal stay events by 3 times.

[0062] To further optimize the above technical solution, the inspection robot adaptive adjustment module includes a density analysis unit, an optical parameter optimization unit, and a network load balancing unit. Among them, the density analysis unit calculates the regional personnel aggregation degree based on the video stream; the optical parameter optimization unit dynamically adjusts the exposure parameters and fill light intensity of the camera; the network load balancing unit allocates edge computing resources according to data transmission requirements;

[0063] The density analysis unit controls the optical parameter optimization unit to increase the camera frame rate in high-density areas to 60 fps according to the heat map output by the personnel positioning and tracking module. At the same time, the network load balancing unit adaptively reduces the video stream resolution in low-risk areas to 720p, where the focal length adjustment strategy is dynamically associated with the spatial coordinates of the equipment safety operation boundary in the 3D scene reconstruction module.

[0064] The density analysis unit automatically divides the monitoring priority based on the crowd density heat map of the video stream (such as when the density > 5 people / m 2 the area is marked as a key area), increasing the capture rate of group violations (such as multiple people not wearing safety ropes) by the camera group from 80% to 98%; the optical parameter optimization unit automatically switches to the infrared mode and adjusts the gain parameters in low-light environments (<50 lux), increasing the proportion of effective monitoring images at night from 60% to 95%; the network load balancing unit dynamically allocates edge computing resources according to the risk level (such as allocating 70% computing power to high-risk areas), increasing the overall system resource utilization rate from 65% to 90% and reducing the hardware investment cost by 35%.

[0065] To further optimize the above technical solution, the inspection robot is equipped with a multi-spectral imager and a gas detector; and the inspection robot is provided with a robotic arm that can remotely operate equipment in dangerous areas; the inspection robot is provided with an autonomous navigation unit that combines the SLAM algorithm and UWB positioning information to plan the inspection path;

[0066] During the inspection robot's autonomous navigation along the planned path, it can collect multi-spectral images of dangerous areas. The collected data is transmitted back to the multi-modal analysis module via a 5G private network. When a gas leak is detected, the historical data playback layer of the virtual power plant model is triggered to retrieve the handling records of similar events, and the simulation training layer is driven to generate an emergency handling virtual scenario.

[0067] The multi-spectral imager (400 - 1700nm band) carried by the inspection robot, combined with a gas detector (accuracy 0.1ppm), improves the methane leak detection sensitivity in the coal conveyor corridor to 0.5ppm, reducing the detection blind area by 90% compared to traditional fixed sensors; the six-axis robotic arm (repeated positioning accuracy ±0.05mm) of the robotic arm can remotely operate the valve switch, increasing the equipment maintenance efficiency in dangerous areas by 4 times; the SLAM algorithm of the autonomous navigation unit integrates UWB positioning data, shortening the inspection path planning time from 30 minutes to 5 minutes and improving the obstacle avoidance success rate.

[0068] An intelligent method for identifying employees' violations in power plants based on the AI large model, the specific method is as follows:

[0069] Deploy a group of rotatable high-definition cameras and UWB positioning base stations to form a dynamic perception network covering the entire plant area. The visual acquisition module uses the background difference algorithm to extract moving targets, and through the spatio-temporal alignment algorithm with UWB positioning data, realizes three-dimensional trajectory tracking with centimeter-level accuracy; the environmental sensor group collects parameters such as temperature, humidity, and vibration in real time to construct a baseline database for the operating state of equipment;

[0070] The edge computing node performs preprocessing on the video stream. After compressing the data volume by 60% through adaptive coding technology, it is transmitted to the local real-time analysis unit and the cloud training platform through different channels; after inertial navigation compensation, the personnel positioning data is input into the trajectory prediction model to generate a probability map of the moving path, and collision detection is performed with the safety boundary in the three-dimensional virtual model to achieve early warning of violation behaviors;

[0071] The multi-modal analysis module adopts a feature-level fusion strategy to spatially register visual behavior features, infrared temperature distribution, and acoustic fingerprint spectrum features to construct a violation determination matrix with spatio-temporal correlation; the hierarchical early warning module dynamically selects response strategies based on the risk index, triggering a progressive control from audible and visual warnings to equipment interlocks.

[0072] Working principle:

[0073] This technical solution is based on a three-layer architecture of "multi-source perception - intelligent analysis - dynamic decision-making" to construct a closed-loop control system for power plant safety management and control. Its core principle is as follows:

[0074] Construction of the multi-dimensional perception layer

[0075] By deploying a group of rotatable high-definition cameras and UWB positioning base stations, a dynamic perception network covering the entire factory area is formed. The visual acquisition module uses the background difference algorithm to extract moving targets, and through the spatio-temporal alignment algorithm (timestamp synchronization + spatial coordinate mapping) with UWB positioning data, three-dimensional trajectory tracking with centimeter-level accuracy is achieved. The environmental sensor group collects parameters such as temperature, humidity, and vibration in real time to construct a baseline database for the operating status of equipment.

[0076] Edge intelligent processing layer

[0077] The edge computing node performs preprocessing of the video stream (light compensation + motion enhancement). After compressing the data volume by 60% through adaptive coding technology, it is transmitted to the local real-time analysis unit and the cloud training platform through different channels. After inertial navigation compensation, the personnel positioning data is input into the trajectory prediction model to generate a probability map of the moving path, and collision detection is performed with the safety boundary in the three-dimensional virtual model to achieve early warning of violations.

[0078] Multi-modal decision-making layer

[0079] The multi-modal analysis module adopts a feature-level fusion strategy to spatially register visual behavior features, infrared temperature distribution, and acoustic fingerprint spectrum features to construct a violation determination matrix with spatio-temporal correlation. The fuel safety sub-module calculates the correlation between the morphological parameters of the stockpile (curvature change rate + tilt angle) and the temperature gradient distribution through three-dimensional point cloud reconstruction technology to establish a collapse risk prediction model. The hierarchical early warning module dynamically selects response strategies based on the risk index and triggers a progressive control from acoustic-optical warning to equipment interlock.

[0080] System optimization layer

[0081] The adaptive adjustment module constructs an optimization model for resource allocation. Based on the personnel density heat map and network load status, it dynamically adjusts the camera parameters (frame rate / resolution) and the calculation resource allocation strategy. The mobile inspection subsystem constructs an environmental topology map through the SLAM algorithm and performs real-time complementary calibration with the data of the fixed perception network to eliminate monitoring blind spots. The dynamic update mechanism (coordinate calibration 10 times per second) of the three-dimensional virtual model ensures millimeter-level mapping accuracy between the virtual and real scenes.

[0082] Through the deep integration of technologies such as machine vision, edge computing, and multi-modal fusion, this technical system constructs an intelligent management and control platform with environmental adaptability, realizing closed-loop management from data collection, analysis and decision-making to control execution.

[0083] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0084] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent identification system for power plant employees' violations based on large AI models, characterized in that, Including: A visual acquisition module, an edge computing node, a personnel positioning and tracking module, a three-dimensional scene reconstruction module, a multimodal analysis module, a hierarchical warning module, an adaptive adjustment module, and an inspection robot; Among them, the visual acquisition module forms a dynamic perception network covering the fuel transportation channel and the operation area through a rotatable high-definition camera group, and the output end of the visual acquisition module is connected to the edge computing node for background difference processing; The edge computing node establishes a data channel with the personnel positioning and tracking module through a fiber optic network. The personnel positioning and tracking module aligns the UWB positioning data and the video feature matching result in space and time, and the generated three-dimensional coordinate data stream is input into the three-dimensional scene reconstruction module to construct a virtual power plant model; A two-way data interaction is established between the virtual power plant model and the multimodal analysis module, and the abnormal behavior recognition result of the multimodal analysis module triggers the corresponding response mechanism of the hierarchical warning module; The adaptive adjustment module dynamically controls the working parameters of the visual acquisition module, and the adaptive adjustment module receives the real-time personnel distribution data from the personnel positioning and tracking module.

2. The intelligent identification system for employees' violations in power plants based on the AI large model according to claim 1, wherein The dynamic perception network is provided with dynamic reference points, and the dynamic reference points include a reference locator and an environmental sensor group; The reference locator is composed of a combination of a laser rangefinder and a radio frequency identification device; the environmental sensor group integrates temperature and humidity, gas concentration, and vibration detection units.

3. The intelligent identification system for employees' violations in power plants based on the AI large model according to claim 2, wherein, The dynamic reference points are arranged at intervals along the axis of the conveyor belt. The dynamic reference points are provided with reference locators. The laser ranging signals emitted by the reference locators and the UWB positioning chip group form a composite positioning field. The environmental data collected by the environmental sensor group is uploaded to the edge computing node through a communication interface, and the edge computing node controls the fill light intensity of the camera group according to the change of environmental illumination.

4. An intelligent identification system for employees' violations in power plants based on an AI large model according to claim 1, characterized in that, The edge computing node includes an image preprocessing unit, a data compression unit, and a cache management unit; The image preprocessing unit implements background difference and light compensation algorithms; the data compression unit uses adaptive coding technology to reduce the transmission bandwidth; the cache management unit configures a dual-channel storage architecture to process real-time stream data and historical comparison data respectively.

5. The intelligent recognition system for violations committed by power plant employees based on the AI large model according to claim 4, wherein The moving target detection result implemented by the image preprocessing unit generates a low-bitrate video stream through the data compression unit. This video stream is distributed to the dual channels of the local memory and the cloud server by the cache management unit. The local memory retains the video clips of the most recent 15 minutes for real-time behavior analysis, and the historical data stored in the cloud server is used to train the equipment safety operation boundary of the virtual power plant model.

6. The intelligent recognition system for employees' violations in power plants based on the AI large model according to claim 1, characterized in that, The personnel positioning and tracking module includes an inertial navigation unit, a trajectory prediction unit, and a safety area mapping unit. The inertial navigation unit integrates a gyroscope and an acceleration sensor; the trajectory prediction unit establishes a personnel movement path prediction model based on the Markov chain model; the safety area mapping unit divides the virtual power plant model into working partitions with different risk levels.

7. An intelligent identification system for employees' violations in power plants based on the AI large model according to claim 6, characterized in that, The personnel positioning and tracking module realizes precise positioning through the following methods: The acceleration data collected by the inertial navigation unit compensates for the transmission delay of the UWB positioning signal. Based on the corrected positioning coordinates, the trajectory prediction unit generates a probability map of the personnel movement path. After the probability map of the personnel movement path is input into the safety area mapping unit, it is subjected to spatial overlay analysis with the high-risk operation area in the virtual power plant model, and the analysis result is dynamically updated to the risk level assessment unit of the hierarchical early warning module.

8. An intelligent identification system for employees' violations in power plants based on the AI large model according to claim 1, characterized in that, The adaptive adjustment module includes a density analysis unit, an optical parameter optimization unit, and a network load balancing unit. Among them, the density analysis unit calculates the personnel aggregation degree in the area based on the video stream; the optical parameter optimization unit dynamically adjusts the exposure parameters and fill light intensity of the camera; the network load balancing unit allocates edge computing resources according to the data transmission requirements; The density analysis unit controls the optical parameter optimization unit to increase the camera frame rate in the high-density area to 60fps according to the heat map output by the personnel positioning and tracking module. At the same time, the network load balancing unit adaptively reduces the video stream resolution in the low-risk area to 720p, and the focal length adjustment strategy is dynamically associated with the spatial coordinates of the equipment safety operation boundary in the three-dimensional scene reconstruction module.

9. An intelligent recognition system for employees' violations in a power plant based on an AI large model according to claim 1, characterized in that, The inspection robot is equipped with a multi-spectral imager and a gas detector; and the inspection robot is provided with a robotic arm capable of remotely operating equipment in dangerous areas; the inspection robot is provided with an autonomous navigation unit that combines the SLAM algorithm and UWB positioning information to plan the inspection path; In the path planned by the autonomous navigation unit of the inspection robot, it can collect multi-spectral images of dangerous areas. The collected data is transmitted back to the multi-modal analysis module via a 5G private network. When a gas leak is detected, the historical data playback layer of the virtual power plant model is triggered to retrieve the disposal records of similar events, and the simulation training layer is driven to generate an emergency disposal virtual scene.

10. A method for an intelligent recognition system of employees' violations in a power plant based on an AI large model according to any one of claims 1-9, characterized in that, The specific method is as follows: Deploy a rotatable high-definition camera group and UWB positioning base stations to form a dynamic perception network covering the entire plant area. The visual acquisition module uses the background difference algorithm to extract moving targets, and realizes centimeter-level accurate three-dimensional trajectory tracking with the UWB positioning data through the spatio-temporal alignment algorithm; The environmental sensor group collects parameters such as temperature, humidity, and vibration in real time to construct a baseline database of the equipment operation status; The edge computing node performs preprocessing on the video stream. After compressing the data volume through adaptive coding technology, it is transmitted to the local real-time analysis unit and the cloud training platform in different channels; after the personnel positioning data is compensated by inertial navigation, it is input into the trajectory prediction model to generate a probability map of the movement path, and collision detection is performed with the safety boundary in the three-dimensional virtual model to achieve early warning of violations; The multi-modal analysis module adopts a feature-level fusion strategy to spatially register visual behavior features, infrared temperature distribution, and acoustic fingerprint spectrum features to construct a violation determination matrix with spatio-temporal correlation; The hierarchical early warning module dynamically selects response strategies according to the risk index, triggering a progressive control from audible and visual warnings to equipment interlocks.

Images