Power plant specific area personnel behavior identification and early warning system

Abstract

This invention discloses a personnel behavior recognition and early warning system for specific areas of a power plant, relating to the field of intelligent industrial safety monitoring technology. It includes a data acquisition module, an edge computing module, and a central processing platform. The data acquisition module is responsible for collecting and initially processing multi-source heterogeneous data, including personnel identity, behavior, location, and environmental parameters. The edge computing module performs lightweight processing and localized decision-making on the raw data, reducing the load on the central processing platform. The central processing platform achieves full-process safety control of high-risk areas through multimodal data fusion, dynamic access control, and trajectory analysis. In this invention, lightweight AI models are embedded in various cameras of the employee badge authentication terminal and visual perception unit, enabling preliminary inference to be performed directly on the employee badge authentication terminal and camera, thus equipping the employee badge authentication terminal and camera with edge computing capabilities. Independent edge servers or industrial control computers are deployed near the data sources in various areas of the power plant as regional edge computing nodes.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for industrial safety, and in particular to a system for identifying and warning of personnel behavior in specific areas of a power plant. Background Technology

[0002] Industrial safety intelligent monitoring technology is a core area for ensuring the safety of personnel and equipment in high-risk industrial environments. This field focuses on achieving proactive safety management of personnel behavior, environmental risks, and equipment status through multimodal sensing, real-time data analysis, and intelligent decision-making. Power plants, as typical high-risk scenarios, present risks such as high-temperature and high-pressure equipment, radiation areas, and toxic gas leaks, placing even more stringent requirements on monitoring systems.

[0003] Existing technologies generally rely on manual inspections and fixed alarm devices, forming a passive monitoring mode. When personnel violate regulations or accidentally enter restricted areas, the system cannot actively identify real-time risks, resulting in response delays. This lag weakens accident prevention capabilities, and violations are often only discovered in post-incident tracing. The single sensing mode leads to insufficient anti-interference capabilities. Traditional solutions use isolated sensors, and the data from each system is fragmented and cannot be cross-verified. This single-modal system is difficult to adapt to the complex operating conditions of power plants with high temperature, high pressure, and high dust. Access control configuration relies on manual updates of the shift schedule, and the change delay often exceeds 2 hours. It is impossible to track the actual activity paths of personnel in authorized areas, making it impossible to prevent behaviors such as unauthorized stays and accidental entry into radiation areas in real time. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a power plant specific area personnel behavior recognition and early warning system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a power plant specific area personnel behavior recognition and early warning system, including a data acquisition module, an edge computing module, and a central processing platform. The data acquisition module is responsible for the collection and preliminary processing of multi-source heterogeneous data, including personnel identity, behavior, location, and environmental parameters. The edge computing module performs lightweight processing and localized decision-making on the raw data to reduce the load on the central processing platform. The central processing platform achieves full-process safety control of high-risk areas through multimodal data fusion, dynamic permission management, and trajectory analysis.

[0006] As a further description of the above technical solution:

[0007] The data acquisition module includes: a visual perception unit that captures personnel movements and environmental information through thermal imaging cameras, wide-angle high-definition cameras, and behavior analysis cameras; a work badge authentication terminal that uses RFID / NFC to read work badge information and integrates liveness detection to prevent identity theft; a wearable device unit that reports personnel location and abnormal status in real time through a UWB positioning module and an accelerometer; and an environmental perception unit that collects temperature, humidity, gas concentration, and radiation dose, and marks high-risk environmental events.

[0008] As a further description of the above technical solution:

[0009] The edge computing module deploys lightweight AI models to extract behavioral features, filter invalid data, and perform data compression and structured encapsulation to adapt to a unified transmission protocol.

[0010] As a further description of the above technical solution:

[0011] The edge computing module has a pre-built security policy library that triggers local responses, reducing reliance on the cloud, blurring the faces and ID cards of irrelevant personnel in real time, and uploading only the de-identified key frames of risk events.

[0012] As a further description of the above technical solution:

[0013] The central processing platform calibrates the employee ID card check-in time, video timestamp, and UWB positioning coordinates to construct a global spatial consistency view and uses the DS evidence theory to comprehensively determine risks.

[0014] As a further description of the above technical solution:

[0015] The central processing platform updates permissions based on a job-region-time matrix, generates dynamic verification logs, issues temporary permissions via mobile approval, limits regions, time periods and operation types, draws electronic fences, detects trajectory deviations or timeouts, optimizes response strategies in conjunction with environmental parameters, and pushes risk levels to the early warning module.

[0016] As a further description of the above technical solution:

[0017] The system process is as follows:

[0018] S1: Multi-source data acquisition

[0019] Visual perception unit: thermal imaging camera, wide-angle high-definition camera, behavior analysis camera capture personnel outlines, safety equipment wearing status and high-risk behaviors in real time;

[0020] Employee ID card authentication terminal: Forces verification of employee ID card information at the area entrance and integrates liveness detection to prevent impersonation;

[0021] Wearable device unit: The smart safety helmet reports the personnel's location and abnormal status in real time;

[0022] Environmental sensing unit: Temperature, humidity, gas concentration, and radiation sensors monitor environmental risk parameters;

[0023] S2: Edge computing preprocessing

[0024] Lightweight AI inference: Deploy models at the camera end to extract behavioral features in real time and filter out invalid data;

[0025] Localized rule execution: Predefined rules trigger local responses;

[0026] Data anonymization and compression: The faces of irrelevant persons in the video are blurred, only key frames of risk events are retained, and the data is compressed before uploading;

[0027] S3: Data Upload

[0028] The preprocessed standardized data is transmitted to the central processing platform via a dual-channel redundant 5G private network or Ethernet, and is automatically cached and incrementally restored when the network is interrupted.

[0029] S4: Multimodal Data Fusion

[0030] Spatiotemporal alignment: calibrate employee ID card check-in time, camera timestamps, and location data to millisecond-level time sequence;

[0031] Integrated decision-making: Risk is comprehensively assessed using the DS evidence theory, and the risk level is calculated by weighting environmental parameters;

[0032] S5: Dynamic Permission Verification

[0033] Personnel permissions are matched in real time based on a job-region-time time permission matrix. Temporary permissions are issued via mobile approval and automatically reset upon expiration.

[0034] S6: Trial-based early warning trigger

[0035] Level 1 Warning: Not wearing safety equipment or entering an unauthorized area;

[0036] Response actions: On-site audible and visual alarms, and push notifications to the individual's wristband;

[0037] Level 2 Warning: Unauthorized stay exceeding time limit, misuse of tools;

[0038] Response actions: Lock the area access control, trigger power rationing in the DCS system, and push the work order to the management personnel;

[0039] Level 3 warning: People are stranded during a fire / leak, and high risks are compounded;

[0040] Response actions: Initiate a plant-wide emergency shutdown, broadcast evacuation instructions, push AR escape routes, and dispatch rescue personnel;

[0041] S7: Edge Response Execution

[0042] The warning command is sent to the edge node through an incremental protocol, directly triggering local actions and deeply integrating with the power plant's DCS system to automatically execute equipment linkage.

[0043] S8: Full-chain traceability

[0044] Store de-identified data associated with the events;

[0045] Supports multi-dimensional searches by time, people, and region, and allows for one-click tracing of the complete event chain;

[0046] S9: Compliance Report Generation

[0047] Automatically extract safety indicators and generate reports that comply with ISO45001 standards to meet regulatory review requirements.

[0048] The present invention has the following beneficial effects:

[0049] 1. This invention first constructs an active security protection system, solving the problems of passivity, singularity, and fragmented access control in traditional power plant monitoring systems through technological innovation. It achieves closed-loop management of personnel behavior in high-risk areas throughout the entire process. The system uses four-dimensional data fusion—work badge authentication (identity), liveness detection (facial recognition), trajectory tracking (behavior), and environmental perception (risk parameters)—to form a closed-loop verification. The work badge authentication terminal integrates RFID readers and near-infrared cameras, employing the ArcFace algorithm to compare work badge photos with facial features in real time. If the cosine similarity is below a threshold, identity theft is determined. Random blinking / head-turning actions are used to resist forgery attacks. Simultaneously, UWB positioning... The positioning module achieves centimeter-level real-time positioning and generates personnel contour coordinates by combining thermal imaging cameras. It effectively penetrates smoke, dust, and low-light environments. When the deviation between the authorized area on the employee badge and the actual positioning exceeds a threshold, it triggers an unauthorized behavior judgment based on DS evidence theory, improving the accuracy of permission violation identification and reducing the false alarm rate. It innovatively designs a job-area-time period permission matrix, predefines the accessible areas and time periods for different jobs, and automatically synchronizes with the scheduling system to avoid manual configuration errors. Temporary permissions are dynamically issued through mobile terminal approval, limiting the area, time period, and operation type. When the permission expires, the access control is automatically reset. The system compares the employee badge permissions with the spatiotemporal coordinates in real time and generates dynamic verification logs.

[0050] 2. In this invention, a hybrid edge computing architecture combining embedded and regional components is adopted. Lightweight AI models are embedded within various cameras of the employee badge authentication terminal and visual perception unit, allowing preliminary inference to be performed directly on the employee badge authentication terminal and camera. This enables the employee badge authentication terminal and camera to possess edge computing capabilities, which can be considered "minimum units of edge computing nodes." However, due to limited computing power, they only support specific tasks. Independent edge servers or industrial control computers are deployed near data sources in various areas of the power plant as regional edge computing nodes. The OpenPose algorithm extracts 18 skeletal key points, combined with optical flow analysis to resolve high-risk behaviors. A liveness detection model is embedded in the employee badge terminal to reduce redundancy in the original data. Unrelated personnel faces in the video stream are blurred in real time, retaining only key frames of high-risk events. Sensitive data is stored locally, and after desensitization, the data is processed through… Dual-channel upload via 5G private network or Ethernet reduces bandwidth consumption. Based on edge node CPU / memory utilization, it prioritizes processing data from high-risk areas. It provides structured storage of anonymized video clips, work badge records, location trajectories, and environmental parameters. It supports one-click backtracking by time, personnel, and region. Thermal imaging and wide-angle high-definition cameras complement each other, with the wide-angle lens covering blind spots. Dynamic HDR algorithms adapt to strong light / backlight scenes. Gas sensors monitor CO and H2S concentrations in real time, and radiation sensors link with location data to delineate exposure areas. It automatically generates compliance reports and extracts indicators such as the number of violations and response latency, reducing audit time. Through four innovations—multimodal fusion verification, dynamic permission adaptation, edge intelligent computing, and hierarchical device linkage—it fills the technological gaps in real-time performance, accuracy, and environmental adaptability in the field of industrial security. Attached Figure Description

[0051] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Reference Figure 1This invention provides an embodiment of a personnel behavior recognition and early warning system for specific areas of a power plant, comprising a data acquisition module, an edge computing module, and a central processing platform. The data acquisition module is responsible for the collection and preliminary processing of multi-source heterogeneous data, including personnel identity, behavior, location, and environmental parameters. The edge computing module performs lightweight processing and localized decision-making on the raw data, reducing the load on the central processing platform. The central processing platform achieves full-process security control of high-risk areas through multimodal data fusion, dynamic access control, and trajectory analysis. The data acquisition module includes: a visual perception unit that captures personnel actions and environmental information through thermal imaging cameras, wide-angle high-definition cameras, and behavior analysis cameras; a work badge authentication terminal that reads work badge information using RFID / NFC and integrates liveness detection to prevent identity theft; and a wearable device unit that reports personnel location and abnormal status in real time through a UWB positioning module and an accelerometer. The sensing unit collects temperature, humidity, gas concentration, and radiation dose, marking high-risk environmental events. The edge computing module deploys a lightweight AI model to extract behavioral features, filter invalid data, and perform data compression and structured encapsulation, adapting to a unified transmission protocol. The edge computing module has a pre-built security policy library that triggers local responses, reducing reliance on the cloud. It blurs the faces and name tags of irrelevant personnel in real time, uploading only anonymized key frames of risk events. The central processing platform calibrates name tag attendance time, video timestamps, and UWB positioning coordinates to construct a globally consistent spatial view. It uses DS evidence theory to comprehensively determine risks. The central processing platform updates permissions based on a job-region-time period matrix, generates dynamic verification logs, and issues temporary permissions through mobile approval, limiting areas, time periods, and operation types. It draws electronic fences, detects trajectory deviations or timeouts, optimizes response strategies in conjunction with environmental parameters, and pushes risk levels to the early warning module.

[0054] The system process is as follows:

[0055] S1: Multi-source data acquisition

[0056] Visual perception unit: thermal imaging camera, wide-angle high-definition camera, behavior analysis camera capture personnel outlines, safety equipment wearing status and high-risk behaviors in real time;

[0057] Employee ID card authentication terminal: Forces verification of employee ID card information at the area entrance and integrates liveness detection to prevent impersonation;

[0058] Wearable device unit: The smart safety helmet reports the personnel's location and abnormal status in real time;

[0059] Environmental sensing unit: Temperature, humidity, gas concentration, and radiation sensors monitor environmental risk parameters;

[0060] S2: Edge computing preprocessing

[0061] Lightweight AI inference: Deploy models at the camera end to extract behavioral features in real time and filter out invalid data;

[0062] Localized rule execution: Predefined rules trigger local responses;

[0063] Data anonymization and compression: The faces of irrelevant persons in the video are blurred, only key frames of risk events are retained, and the data is compressed before uploading;

[0064] S3: Data Upload

[0065] The preprocessed standardized data is transmitted to the central processing platform via a dual-channel redundant 5G private network or Ethernet, and is automatically cached and incrementally restored when the network is interrupted.

[0066] S4: Multimodal Data Fusion

[0067] Spatiotemporal alignment: calibrate employee ID card check-in time, camera timestamps, and location data to millisecond-level time sequence;

[0068] Integrated decision-making: Risk is comprehensively assessed using the DS evidence theory, and the risk level is calculated by weighting environmental parameters;

[0069] S5: Dynamic Permission Verification

[0070] Personnel permissions are matched in real time based on a job-region-time time permission matrix. Temporary permissions are issued via mobile approval and automatically reset upon expiration.

[0071] S6: Trial-based early warning trigger

[0072] Level 1 Warning: Not wearing safety equipment or entering an unauthorized area;

[0073] Response actions: On-site audible and visual alarms, and push notifications to the individual's wristband;

[0074] Level 2 Warning: Unauthorized stay exceeding time limit, misuse of tools;

[0075] Response actions: Lock the area access control, trigger power rationing in the DCS system, and push the work order to the management personnel;

[0076] Level 3 warning: People are stranded during a fire / leak, and high risks are compounded;

[0077] Response actions: Initiate a plant-wide emergency shutdown, broadcast evacuation instructions, push AR escape routes, and dispatch rescue personnel;

[0078] S7: Edge Response Execution

[0079] The warning command is sent to the edge node through an incremental protocol, directly triggering local actions and deeply integrating with the power plant's DCS system to automatically execute equipment linkage.

[0080] S8: Full-chain traceability

[0081] Store de-identified data associated with the events;

[0082] Supports multi-dimensional searches by time, people, and region, and allows for one-click tracing of the complete event chain;

[0083] S9: Compliance Report Generation

[0084] Automatically extract safety indicators and generate reports that comply with ISO45001 standards to meet regulatory review requirements.

[0085] The data acquisition module, as the system's sensing front end, is responsible for the accurate acquisition and preprocessing of multi-source heterogeneous data. Its core modules include a visual perception unit, a badge authentication terminal, a wearable device unit, and an environmental perception unit. Various cameras in the badge authentication terminal and visual perception unit embed lightweight AI models. The visual perception unit achieves accurate monitoring in complex scenes through a combination of multiple camera types. Thermal imaging cameras generate heat maps based on infrared radiation differences, using uncooled microbolometers to effectively penetrate interference from smoke, dust, and low-light environments. They output real-time thermal imaging video streams showing the X, Y, and Z contour coordinates and movement direction of personnel. Wide-angle high-definition cameras, equipped with 120° ultra-wide-angle lenses, cover blind spots in densely populated areas. Combined with an embedded YOLOv5 model, they perform real-time video stream analysis of monitoring blind spots, identifying the wearing status of safety equipment in real time. The YOLOv5 model divides the image into a grid, predicting the bounding box and class probability for each grid. The formula is:

[0086] Among them, the loss function for the bounding box often uses CIoU loss to optimize the prediction. , , These are the coordinates of the bounding box center. These are the width and height of the bounding box. It's a confidence score, representing the probability that an object exists within the bounding box, ranging from 0 to 1. It is a class probability vector, representing the probability that an object belongs to each category, such as "wearing a helmet" or "not wearing a helmet". It is the center point of the prediction box. Center point of the real frame The Euclidean distance between them The minimum diagonal length between the predicted bounding box and the ground truth bounding box. For the width and height of the actual bounding box, As a measure of aspect ratio consistency, As a weighting factor, it is used to balance geometric errors and adapt to strong light reflection or backlight scenes through a dynamic HDR algorithm. The behavior analysis camera extracts 18 skeletal key points of the human body based on the OpenPose algorithm to construct a spatiotemporal motion map, such as joint angle change rate and motion trajectory. OpenPose detects and associates key points, and the key point detection formula is as follows: ,in, It is the first The locations of the key nodes, and the linear assignment of key point associations using PAFs, are as follows: , For in position At the key point The confidence plot values ​​range from 0 to 1. The Diracdelta function is used to indicate the location of key points. It is the first The first of the individuals Confidence level of each key point The PartAffinityField vector represents the limb orientation. For interpolation points between two points on the limb, It is a normalization parameter, ranging from 0 to 1. To start from the key point arrive The unit direction vector, The association score for keypoint pairs is used to construct human pose. The spatial coordinates of skeletal keypoints are analyzed frame-by-frame from the video stream. Combined with optical flow, motion vectors between consecutive frames are calculated to identify high-risk behaviors such as climbing railings, running, or unauthorized operation of equipment. The optical flow method is based on the assumption of constant brightness, and its core equation is the Horn-Schunck optical flow constraint: The global optimization objective function is: , Image brightness is represented by grayscale values. The spatial gradient of the image in the x and y directions is calculated using operators such as Sobel. The change of an image over time, also known as the temporal gradient. It is an optical flow vector, representing the velocity of a pixel in the x and y directions, with units of pixels per frame. This is the regularization parameter, which balances the data terms and the smoothing term; it is usually set to 1.0. It is the spatial derivative of the optical flow vector, used for smoothing constraints. The employee badge authentication terminal deploys a high-frequency RFID reader / writer at the entrance of a specific area to forcibly verify employee badge permissions. It reads the radio frequency signal from the embedded chip in the badge through electromagnetic induction, parsing the employee ID, job title, and permission information. The authentication process is forcibly triggered at the area entrance, allowing entry only if the badge permissions match the pre-set area-job matrix. If no attendance is taken or the chip is damaged, the terminal triggers a red warning light and an audible and visual alarm, and takes a photo as evidence. To prevent badge misuse, the terminal integrates a near-infrared camera to capture facial images. Through liveness detection, the ArcFace algorithm extracts facial features and compares them with the pre-stored photo on the badge using cosine similarity. If the similarity is below a threshold, it is determined to be identity fraud, triggering an alarm and generating an evidence snapshot to prevent badge misuse. ArcFace uses additive angular margin loss to enhance feature discriminative power; the loss function is: The output after feature extraction is a normalized feature vector. , For feature vectors Compared to the real category The radian angle between the weight vectors This is an additive angular margin, typically 0.5, used to increase the distance between classes. This is a feature scaling factor, typically 64, used to magnify angular differences. The number of training samples, For the eigenvector and the th Cosine similarity of the weight vectors of each category The facial feature vector extracted after L2 normalization is designed with anti-fraud features, randomly requiring blinking or head turning actions to detect liveness and resist photo / video forgery attacks. The wearable device unit includes a smart helmet, which integrates a UWB ultra-wideband positioning module and a three-axis accelerometer. It has a built-in Decawave DW1000 chip to generate real-time point cloud data of the person's location. It measures the distance difference with the positioning base station using the time-of-flight method to achieve centimeter-level real-time positioning and calculates the distance based on the signal propagation time. , The distance is in meters. At the speed of light, The signal round-trip time, measured in seconds, represents the time difference between transmission and reception. Coordinates are reported to the central platform once per second. An integrated MEMS three-axis accelerometer continuously monitors personnel movement, using thresholds to distinguish between normal movement and falls or impacts, triggering local buzzer alarms and uploading event codes. A multi-base station network is deployed with four base stations, employing a time-difference-of-arrival (TDOA) algorithm to achieve 3D spatial coverage. The TDOA positioning equation... Is the tag to the first The distance difference between each base station and the reference base station, in meters. It's the speed of light. The signal arrived at the The latency of each base station, measured in seconds. It is the time delay for the signal to reach the reference base station, measured in seconds. These are the coordinates of the tagged personnel. It is the first The coordinates of each base station It refers to the coordinates of the base station, the formula is: The environmental perception unit monitors the physical state of high-risk areas in real time through temperature and humidity sensors, gas concentration sensors, and radiation sensors to assist in behavioral risk assessment. The DHT22 digital temperature and humidity sensor collects data every 5 seconds and transmits it to the edge node via the LoRaWAN protocol. The gas concentration sensor uses the MQ-135 electrochemical type to detect hazardous gases CO and H2S, outputting a 0-5V analog signal which is converted by an ADC to generate a PPM value. When the CO concentration is ≥50ppm, it is marked as high risk. The radiation sensor uses an LND-712 GM counter tube to detect gamma ray intensity, converting it into radiation dose rate through pulse count rate, and linking it with personnel positioning data to delineate exposure areas. All raw data, including visual features, work badge information, positioning coordinates, and environmental parameters, are pre-processed by the edge computing module and then uploaded to the central processing platform to ensure the efficiency of spatiotemporal alignment and fusion decision-making of multimodal data.

[0087] Edge computing nodes adopt a hybrid deployment mode of embedded and regional approaches. Lightweight AI models are embedded in various cameras of the employee badge authentication terminal and visual perception unit, and preliminary inference is performed directly on the employee badge authentication terminal and camera, enabling the employee badge authentication terminal and camera to have edge computing capabilities. They can be regarded as the "minimum unit of edge computing nodes," but the computing power is limited and only supports specific tasks. Independent edge servers or industrial control computers are deployed near the data sources in various areas of the power plant as regional edge computing nodes, undertaking localized data processing, real-time decision-making, and data preprocessing tasks. These nodes cover multiple cameras, sensors, and employee badge authentication terminals in the area, processing equipment data in the area. The lightweight AI model deployed on the camera extracts behavioral features from the video stream in real time, such as the wearing status of safety equipment and skeletal key points, reducing the redundancy of raw data. Based on predefined rules, such as static background frames and low-confidence detection boxes, invalid or duplicate data is removed, and only features related to high-risk events are retained, reducing data transmission bandwidth consumption. According to the localized rule engine, a built-in security policy library, such as "no entry without a safety helmet" and "electronic fence boundary crossing judgment threshold," is directly applied. Rule matching is performed at the edge to reduce cloud decision-making latency. For events that trigger rules, such as unauthorized intrusion or missing equipment, local responses are immediately triggered. For events such as audible and visual alarms or access control locks, emergency responses can be completed by waiting for instructions from the central platform. Visual features, location coordinates, and environmental parameters are losslessly compressed and structuredly encapsulated. The system is adapted to a unified transmission protocol for heterogeneous data and synchronizes device clocks based on the NTP protocol. Camera timestamps, employee ID card check-in times, and location data are aligned at the millisecond level, providing a foundation for multimodal fusion on the central platform. Real-time blurring of irrelevant personnel's faces and employee ID card information in the video stream is performed, retaining only clear data of key frames of risk events. Data is categorized and cached according to its sensitivity level, such as facial features and radiation dose. Unnecessary data is retained locally, and only anonymized data of risk events is uploaded to the central platform. AI inference tasks are dynamically allocated based on the computing power status of edge nodes, such as CPU / memory utilization, prioritizing data from high-risk areas. Dual-channel redundant transmission using wired Ethernet and wireless 5G private networks is adopted, automatically switching when the main link is interrupted. Data is automatically cached when the network is interrupted and incrementally uploaded after the connection is restored to ensure data integrity.

[0088] The central processing platform distributes dynamic rules to edge nodes via an incremental update protocol, avoiding bandwidth pressure caused by transmitting the entire rule base. Edge nodes provide update status feedback through an ACK confirmation mechanism. When communication with edge nodes is interrupted, the platform uses its locally cached historical rule base and the permission matrix for the most recent 24 hours to continuously perform basic risk assessments until the link is restored and data is resynchronized, ensuring decision continuity. The multimodal data fusion unit, based on the device clock calibrated using the NTP protocol, performs millisecond-level time-series alignment of employee ID card check-in time, camera video stream timestamps, and UWB positioning coordinates, eliminating time deviations in multi-source data. Through spatial coordinate system mapping, it unifies location information collected by different devices, such as thermal imaging camera contour coordinates and UWB positioning data, constructing a global spatial consistency view. DS evidence theory is used to comprehensively judge multimodal data, including the conflict coefficient. for: , Let be the basic probability assignment function, representing the degree of support for the hypothesis from different pieces of evidence. These are assumptions, such as "authorized area A" or "unauthorized behavior." It is the intersection of assumptions. This is the conflict coefficient, ranging from 0 to 1, which indicates the degree of conflict between pieces of evidence. This indicates no conflict. It is the combined BPA, indicating the fusion of evidence for the hypothesis. The probability of support, Let be an empty set, indicating the no-intersection assumption, and the formula is: For identity and trajectory conflict detection, if the authorized area on the work badge is area A, but UWB positioning shows the person is in area B, it is judged as an unauthorized behavior. The system performs correlation analysis between behavior and environment, combining environmental parameters and personnel actions to calculate the risk level and outputs a fusion decision result, such as "high-risk boundary violation" or "low-risk equipment missing," for risk warning and access control. The dynamic access control unit achieves precise control based on a position-area-time period access matrix, predefining the binding relationship between positions, areas, and time periods, such as "boiler inspectors can enter the boiler core area from 8:00 to 17:00." It supports automatic synchronization and updating of access configurations by the scheduling system, and real-time comparison. The system generates dynamic permission verification logs based on employee ID card permissions and current spatiotemporal coordinates, such as "Maintenance personnel entered the radiation zone at 18:05 - permission expired". For the issuance and revocation of temporary permissions, temporary access codes are issued through a mobile approval process, such as for emergency repair tasks, which restrict areas, time periods, and operation types. When permissions expire or tasks are completed, the area access control is automatically reset to prevent unauthorized personnel from staying. The trajectory analysis unit dynamically generates electronic fences and draws real-time personnel activity heatmaps based on UWB positioning data, automatically generating electronic fence boundaries. Combined with the dynamic permission matrix, it determines out-of-bounds behavior and triggers trajectory deviation alarms, such as "Personnel deviated from authorized path 2".The system monitors the time individuals remain in high-risk areas ("5-meter interval"). If this time exceeds the safety threshold, a tiered evacuation protocol is activated, ranging from broadcast warnings to automatic power limiting and finally, emergency shutdown. It also integrates with environmental parameters, such as a sudden temperature rise, to optimize evacuation priorities and dynamically adjust response strategies. The warning module has a three-level warning mechanism. Level 1 warnings are triggered by not wearing safety equipment or entering unauthorized areas (e.g., a conflict between employee ID and location coordinates). The response involves triggering a red warning light and buzzer for an audible and visual alarm, and recording a snapshot of the violation. Level 2 warnings are triggered by unauthorized stay exceeding the preset safety threshold. In cases of unauthorized use of tools, such as unauthorized operation of equipment, the response is to lock the area access control, prohibiting personnel from entering or leaving, and to coordinate with the power plant's DCS system to execute partial power rationing operations, such as cutting off non-core circuits. Emergency work orders and location information are sent to management personnel. The triggering conditions for a Level 3 warning are significant safety hazards, such as personnel remaining in a fire or leak accident, and high-risk behaviors combined with abnormal environments, such as personnel not wearing protective clothing in high-temperature areas. The response is to trigger the plant-wide emergency shutdown protocol, initiate evacuation broadcasts, simultaneously push the optimal escape route to the AR navigation devices of relevant personnel, and dispatch the nearest rescue personnel in conjunction with the emergency command center. Once an alert is triggered, the central processing platform, based on multimodal data fusion results such as employee ID permissions, behavioral analysis, and environmental parameters, issues response commands to edge nodes. Edge nodes directly execute localized responses, such as audible and visual alarms and access control locking, without waiting for cloud confirmation, ensuring millisecond-level latency. Deeply integrated with the power plant's DCS system, it supports automatic triggering of equipment protection actions based on the alert level, such as power rationing for level two alerts and shutdown for level three alerts. Redundant command transmission is achieved through a 5G private network, automatically switching to a backup channel when the main link is interrupted. Response strategies are dynamically optimized based on environmental parameters such as radiation dose and gas concentration, for example, in high-temperature areas. Prioritizing personnel evacuation over power restrictions, the data tracing and auditing module stores end-to-end data, providing structured storage of original event-related data, including anonymized video clips, employee ID verification records, location trajectories, and environmental parameters. It supports multi-dimensional retrieval by time, personnel, and region, retains data version snapshots, and supports one-click backtracking to ensure the integrity and immutability of audit traceability. It generates compliance reports, automatically extracts security event statistical indicators such as the number of violations and response delays, and generates security management reports compliant with ISO 45001 standards. Custom export formats are supported to adapt to regulatory review requirements, and all warning events are linked to original data.

[0089] The data acquisition module collects raw data and transmits it to edge nodes. The edge nodes perform AI inference, data compression, and anonymization, retaining only the characteristics of high-risk events. The pre-processed standardized data is uploaded to the central processing platform via a 5G private network or Ethernet. After receiving the data, the central processing platform performs multimodal fusion and global decision-making. Based on the fusion results, it issues early warning commands and links the DCS system control equipment. Dynamic rules are distributed to edge nodes via incremental protocols to ensure local rule synchronization. The early warning response results are fed back to the central processing platform, forming a closed loop of "identification-decision-response-audit". The entire chain of data is stored and an ISO45001 compliance report is generated, achieving low-latency, high-precision closed-loop safety management and meeting the active protection needs of high-risk areas in power plants.

[0090] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power plant specific area personnel behavior recognition and early warning system, characterized in that: It includes a data acquisition module, an edge computing module, and a central processing platform. The data acquisition module is responsible for the collection and preliminary processing of multi-source heterogeneous data, including personnel identity, behavior, location, and environmental parameters. The edge computing module performs lightweight processing and localized decision-making on the raw data to reduce the load on the central processing platform. The central processing platform achieves full-process security control of high-risk areas through multimodal data fusion, dynamic access control, and trajectory analysis. The data acquisition module includes: a visual perception unit that captures personnel movements and environmental information through thermal imaging cameras, wide-angle high-definition cameras, and behavior analysis cameras; a work badge authentication terminal that reads work badge information using RFID / NFC and integrates liveness detection to prevent identity theft; a wearable device unit that reports personnel location and abnormal status in real time through a UWB positioning module and an accelerometer; and an environmental perception unit that collects temperature, humidity, gas concentration, and radiation dose and marks high-risk environmental events. The central processing platform calibrates employee ID card check-in time, video timestamps, and UWB positioning coordinates to construct a global spatial consistency view and uses DS evidence theory to comprehensively determine risks. The central processing platform updates permissions based on a job-region-time matrix, generates dynamic verification logs, issues temporary permissions through mobile approval, limits regions, time periods and operation types, draws electronic fences, detects trajectory deviations or timeouts, optimizes response strategies in conjunction with environmental parameters, and pushes risk levels to the early warning module. The system integrates four-dimensional data fusion—including employee badge authentication, liveness detection, trajectory tracking, and environmental perception—to form a closed-loop verification. The employee badge authentication terminal integrates RFID readers and near-infrared cameras, using the ArcFace algorithm to compare employee badge photos with facial features in real time. If the cosine similarity is below a threshold, identity theft is determined. Random blinking / head turning actions are used to prevent forgery attacks. Meanwhile, the UWB positioning module achieves centimeter-level real-time positioning, and combined with a thermal imaging camera to generate personnel contour coordinates, effectively penetrating smoke, dust, and low-light environments. When the deviation between the authorized area of ​​the employee badge and the actual positioning exceeds a threshold, an unauthorized behavior judgment is triggered based on DS evidence theory.

2. The power plant specific area personnel behavior recognition and early warning system according to claim 1, characterized in that: The edge computing module deploys lightweight AI models to extract behavioral features, filter invalid data, and perform data compression and structured encapsulation to adapt to a unified transmission protocol.

3. The power plant specific area personnel behavior recognition and early warning system according to claim 1, characterized in that: The edge computing module has a pre-built security policy library that triggers local responses, reducing reliance on the cloud, blurring the faces and ID cards of irrelevant personnel in real time, and uploading only the de-identified key frames of risk events.

4. The power plant specific area personnel behavior recognition and early warning system according to any one of claims 1-3, characterized in that: The system process is as follows: S1: Multi-source data acquisition Visual perception unit: thermal imaging camera, wide-angle high-definition camera, behavior analysis camera capture personnel outlines, safety equipment wearing status and high-risk behaviors in real time; Employee ID card authentication terminal: Forces verification of employee ID card information at the area entrance and integrates liveness detection to prevent impersonation; Wearable device unit: The smart safety helmet reports the personnel's location and abnormal status in real time; Environmental sensing unit: Temperature, humidity, gas concentration, and radiation sensors monitor environmental risk parameters; S2: Edge computing preprocessing Lightweight AI inference: Deploy models at the camera end to extract behavioral features in real time and filter out invalid data; Localized rule execution: Predefined rules trigger local responses; Data anonymization and compression: The faces of irrelevant persons in the video are blurred, only key frames of risk events are retained, and the data is compressed before uploading; S3: Data Upload The preprocessed standardized data is transmitted to the central processing platform via a dual-channel redundant 5G private network or Ethernet, and is automatically cached and incrementally restored when the network is interrupted. S4: Multimodal Data Fusion Spatiotemporal alignment: calibrate employee ID card check-in time, camera timestamps, and location data to millisecond-level time sequence; Integrated decision-making: Risk is comprehensively assessed using the DS evidence theory, and the risk level is calculated by weighting environmental parameters; S5: Dynamic Permission Verification Personnel permissions are matched in real time based on a job-region-time time permission matrix. Temporary permissions are issued via mobile approval and automatically reset upon expiration. S6: Trial-based early warning trigger Level 1 Warning: Not wearing safety equipment or entering an unauthorized area; Response actions: On-site audible and visual alarms, and push notifications to the individual's wristband; Level 2 Warning: Unauthorized stay exceeding time limit, misuse of tools; Response actions: Lock the area access control, trigger power rationing in the DCS system, and push the work order to the management personnel; Level 3 warning: People are stranded during a fire / leak, and high risks are compounded; Response actions: Initiate a plant-wide emergency shutdown, broadcast evacuation instructions, push AR escape routes, and dispatch rescue personnel; S7: Edge Response Execution The warning command is sent to the edge node through an incremental protocol, directly triggering local actions and deeply integrating with the power plant's DCS system to automatically execute equipment linkage. S8: Full-chain traceability Store de-identified data associated with the events; Supports multi-dimensional searches by time, people, and region, and allows for one-click tracing of the complete event chain; S9: Compliance Report Generation Automatically extract safety indicators and generate reports that comply with ISO45001 standards to meet regulatory review requirements.

Images