An AI-based video surveillance system for personnel behavior safety in nuclear power plants.

Abstract

This invention discloses an artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants. It includes a front-end device layer that collects physical environment data from the nuclear power plant; an edge computing layer deployed on a local server at the nuclear power plant, which preprocesses and performs preliminary behavioral analysis on the data collected by the front-end device layer, triggering a local alert when an anomaly is detected; a cloud analysis layer that uses deep learning and natural language processing technologies to perform in-depth analysis on the preprocessed data uploaded from the edge computing layer; and an application service layer that generates dynamic disaster avoidance paths based on the results of the cloud analysis layer. This invention reduces false alarm rates by using the edge computing layer to process video and radiation data in real time, combined with a dual-modal fusion decision-making mechanism; and by deeply integrating NLP knowledge graphs and CNN-LSTM behavior recognition in the cloud analysis layer to achieve spatiotemporal matching of personnel behavior with safety procedures, dynamically generating risk scores; thus significantly improving the real-time performance of behavior monitoring in nuclear power plants.

Description

Technical Field

[0001] This invention relates to the field of nuclear power safety video surveillance technology, specifically to an artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants. Background Technology

[0002] Currently, according to statistics from certain journals, human error accounts for a significant proportion of nuclear accidents at nuclear power plant sites. Therefore, strengthening safety oversight of personnel behavior is essential. Existing safety oversight largely relies on video surveillance to identify violations.

[0003] However, existing conventional video analytics struggles to accurately capture violations in complex scenarios, such as not wearing a safety helmet or accidentally entering high-risk areas. Furthermore, it is susceptible to interference from lighting and noise, resulting in a high false alarm rate. In addition, radiation sensor, access control system, and video surveillance data are processed independently, lacking the ability to integrate and analyze data, and failing to correlate personnel behavior with environmental risks, such as violations in areas with excessive radiation levels. Therefore, there is an urgent need for a monitoring system that integrates artificial intelligence, edge computing, and multimodal data analysis to achieve proactive early warning and closed-loop risk management for nuclear power plant personnel. Summary of the Invention

[0004] To address these issues, the present invention provides an artificial intelligence-based video surveillance system for the safety of personnel behavior in nuclear power plants.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants includes a front-end device layer, an edge computing layer, a cloud analysis layer, and an application service layer.

[0007] The front-end equipment layer collects physical environment data of the nuclear power plant.

[0008] The edge computing layer is deployed on a local server at the nuclear power plant. It preprocesses and performs preliminary behavioral analysis on the data collected by the front-end device layer. When an anomaly is detected, it triggers a local early warning.

[0009] The cloud-based analytics layer utilizes deep learning and natural language processing technologies to perform in-depth analysis on preprocessed data uploaded from the edge computing layer, identifying complex human behavior patterns and potential security risks; it can also integrate information from different data sources.

[0010] The application service layer generates a 3D visualized dynamic disaster avoidance path based on the results of the cloud analysis layer and pushes it to the personnel positioning terminal.

[0011] Furthermore, the data collected by the front-end device layer includes real-time video stream data obtained using a radiation-proof high-definition camera, environmental nuclear radiation data obtained using a nuclear radiation sensor, and access data obtained through a biological access control system.

[0012] Furthermore, the specific implementation steps of the edge computing layer are as follows:

[0013] (1) Video stream preprocessing and radiation data preprocessing are the basic processing units of the edge computing layer. The basic processing unit performs gamma correction on the high-definition video stream through the hardware-accelerated image processing unit, dynamically adjusts the brightness and contrast of the picture to adapt to the complex lighting conditions in the nuclear power area, and uses 3D noise reduction algorithm to eliminate electronic noise and environmental interference in the monitoring picture. For radiation monitoring data, the built-in dose rate threshold detection engine verifies the validity of the sensor data in real time, marks the radiation value that exceeds the safety threshold and triggers the early warning link.

[0014] (2) A lightweight spatiotemporal feature extraction network is deployed in the edge computing layer. The lightweight spatiotemporal feature extraction network adopts an improved 3D convolutional neural network architecture, which extracts the spatial features of video frames and the motion features of the time series simultaneously while ensuring computational efficiency.

[0015] (3) The anomaly detection module constructs a dual-modal fusion decision mechanism; in the first aspect, the behavior analysis results based on computer vision are input into LSTM, and the violation operation mode is identified through time series modeling; in the second aspect, the radiation data stream is passed through the Kalman filter for state estimation, and a dynamic early warning baseline is constructed by combining historical dose data.

[0016] Furthermore, the dual-modal data is synthesized at the decision fusion layer using DS evidence theory to generate element detection results that include anomaly type, occurrence location, and confidence level.

[0017] Furthermore: When a local warning is triggered, for minor violations, an audible and visual alarm is issued through the local display terminal of the edge node, and the event timestamp and video clip are recorded; when a medium-risk behavior is detected, the regional broadcast system is triggered to issue a voice warning, and at the same time, a warning pop-up containing personnel identification information is pushed to the central control room; for major emergencies such as excessive radiation, the system immediately activates the emergency response protocol.

[0018] Furthermore, adjacent edge nodes share personnel trajectory data through a custom P2P communication protocol to build a cross-regional behavior correlation analysis network.

[0019] Furthermore, the cloud-based analysis layer will employ NLP to analyze the nuclear safety procedure knowledge graph and deeply integrate NLP with video analysis. The nuclear safety procedure knowledge graph will first undergo deep semantic understanding using BERT encoding technology, outputting a spatiotemporal constraint rule base for operational steps. The surveillance video will be processed and analyzed by a CNN-LSTM behavior recognition model, which can automatically detect personnel behavior in the video, identify safety risk behaviors, and record these safety risk behaviors as an event sequence in chronological order.

[0020] The built-in rule engine in the cloud analytics layer receives spatiotemporal information from the event sequence of the nuclear safety procedure knowledge graph, performs spatiotemporal matching and risk assessment; the rule engine determines whether there is a safety risk based on the spatiotemporal relationship between equipment status and personnel behavior, and calculates the corresponding risk score.

[0021] Furthermore, the event sequence includes the behavior type and records the spatiotemporal information of the behavior occurrence.

[0022] Furthermore: the risk assessment formula of the rule engine is as follows;

[0023]

[0024] Where R is the final risk score; n is the total number of risk factors; i is the risk factor index; Wi is the risk weight associated with the i-th risk factor, set by nuclear safety experts; Di is the real-time radiation dose associated with the i-th risk factor; Dmax is the upper limit of the dose threshold; λ is the behavioral violation coefficient; Ei is the behavioral violation flag associated with the i-th risk factor; Rdi is the radioactive exposure coefficient associated with the i-th risk factor; and k is the risk amplification factor.

[0025] The present invention has the following advantages: The present invention processes video and radiation data in real time through the edge computing layer and reduces the false alarm rate by combining a dual-modal fusion decision mechanism (DS evidence theory); the cloud analysis layer deeply integrates NLP knowledge graph and CNN-LSTM behavior recognition to achieve spatiotemporal matching of personnel behavior and safety procedures and dynamically generate risk scores; it significantly improves the real-time performance, accuracy and emergency response capability of behavior monitoring in nuclear power bases and reduces the risk of human-caused accidents.

[0026] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0027] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0028] Figure 1 This is a system block diagram of an artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants, provided as an embodiment of this application.

[0029] Figure 2 This is a flowchart of the edge computing layer in an artificial intelligence-based video surveillance system for personnel behavior safety in a nuclear power plant, provided as an embodiment of the present invention. Detailed Implementation

[0030] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. 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.

[0031] Please see Figures 1-2 A video surveillance system for personnel behavior safety in nuclear power bases based on artificial intelligence includes: a front-end device layer, an edge computing layer, a cloud analysis layer, and an application service layer.

[0032] The front-end equipment layer is used to collect physical environmental data of the nuclear power plant, including obtaining real-time video stream data using radiation-proof high-definition cameras, acquiring environmental data using nuclear radiation sensors, and obtaining access data through a biological access control system.

[0033] Among them, the anti-radiation high-definition camera adopts an anti-radiation shell and optical lens, which can work stably for a long time in high-radiation environments and capture high-definition, distortion-free video images; the anti-radiation high-definition camera has functions such as automatic exposure and white balance adjustment to ensure that clear video images can be obtained under different lighting conditions; it supports network transmission and transmits video data to the back-end processing layer for analysis in real time.

[0034] Nuclear radiation sensors are deployed in key areas of nuclear power plants to monitor environmental radiation levels in real time. The sensors employ high-precision sensors and advanced signal processing technology to ensure the accuracy and real-time nature of radiation data.

[0035] The biometric access control system integrates biometric technologies (such as fingerprint recognition and facial recognition) to strictly control personnel access to key areas of the nuclear power plant.

[0036] The edge computing layer is deployed on local servers at the nuclear power plant. It is responsible for preprocessing and preliminary behavioral analysis of the data collected by the front-end device layer, so as to reduce the burden on the cloud analysis layer and improve the system response speed.

[0037] The edge computing layer can preprocess video streams and perform real-time behavior analysis on the preprocessed video streams, triggering local alerts when anomalies are detected.

[0038] The specific implementation steps of the edge computing layer are as follows:

[0039] Video stream preprocessing and radiometric data preprocessing are the basic processing units of the edge computing layer;

[0040] The basic processing unit performs gamma correction on the high-definition video stream through a hardware-accelerated image processing unit, dynamically adjusting the brightness and contrast of the image to adapt to the complex lighting conditions in the nuclear power plant area. At the same time, it uses a 3D noise reduction algorithm to eliminate electronic noise and environmental interference in the monitoring image.

[0041] For radiation monitoring data, the built-in dose rate threshold detection engine can verify the validity of sensor data in real time, mark radiation values ​​that exceed the safety threshold, and trigger an early warning link.

[0042] At the real-time behavior analysis level, a lightweight spatiotemporal feature extraction network is deployed in the edge computing layer. The lightweight spatiotemporal feature extraction network adopts an improved 3D convolutional neural network architecture, which simultaneously extracts the spatial features and temporal motion features of video frames while ensuring computational efficiency. Through attention mechanism optimization, the system can focus on key safety elements such as personnel protective equipment and operation gestures, achieving a real-time analysis capability of 25 frames per second.

[0043] The anomaly detection module constructs a dual-modal fusion decision-making mechanism. Firstly, the behavior analysis results based on computer vision are input into a long short-term memory network (LSTM) to identify violation operation patterns, such as not wearing protective equipment or illegally crossing safety warning lines, through temporal modeling. Secondly, radiation data flows through a Kalman filter for state estimation and is combined with historical dose data to construct a dynamic early warning baseline.

[0044] The dual-modal data is synthesized at the decision fusion layer using DS evidence theory to generate a detection result that includes elements such as anomaly type, location of occurrence, and confidence level.

[0045] For minor violations, the system issues audible and visual alarms through the local display terminals of edge nodes and records event timestamps and video clips. When a medium-risk behavior is detected, the system triggers the area broadcast system to issue a voice warning and simultaneously pushes a warning pop-up containing personnel identification information to the central control room. In the event of major hazards such as excessive radiation, the system immediately activates the emergency response protocol, coordinating with the implementing agencies to: activate the air filtration system to purify radioactive materials, seal off the relevant isolation area through the access control controller, and automatically retrieve the most recent 15 seconds of multi-view surveillance video for command personnel to analyze.

[0046] In this embodiment, the edge computing layer achieves collaboration with the upper and lower layers through bidirectional communication links; it connects downwards to the radiation-proof high-definition camera and nuclear radiation sensor of the front-end device layer, uses the GigE Vision protocol to ensure low-latency transmission of video streams, and collects environmental monitoring data through the Modbus TCP protocol; upwards, it pushes structured analysis results to the cloud analysis layer through the MQTT protocol to provide real-time feedback for the iteration of the CNN-LSTM behavior recognition model.

[0047] In addition, adjacent edge nodes share personnel trajectory data through a custom P2P communication protocol, building a cross-regional behavior correlation analysis network, which effectively improves the detection efficiency of group violations.

[0048] The edge computing layer architecture, through hardware and software co-optimization, realizes the unique safety monitoring requirements of nuclear power scenarios at resource-constrained edge terminals. It not only ensures real-time requirements but also ensures the timeliness of emergency response through a hierarchical early warning mechanism, providing intelligent front-end protection for the safety of personnel behavior in nuclear power bases.

[0049] The cloud analytics layer uses deep learning and natural language processing technologies to perform in-depth analysis on the data uploaded from the edge computing layer.

[0050] The cloud analytics layer performs in-depth mining and analysis on the preprocessed data uploaded from the edge computing layer; it uses deep learning and natural language processing technologies to perform in-depth analysis on the data uploaded from the edge computing layer in order to identify complex human behavior patterns and potential security risks.

[0051] Through in-depth data analysis, the system can discover subtle anomalies that may be overlooked, providing strong support for subsequent security management.

[0052] In addition, the cloud-based analytics layer makes full use of the nuclear safety regulations knowledge graph to provide professional knowledge support for behavioral analysis. The nuclear safety regulations knowledge graph contains a wealth of information such as nuclear safety regulations, operating procedures, and historical cases, which can provide authoritative reference for system analysis. By combining the nuclear safety regulations knowledge graph, the cloud-based analytics layer can more accurately determine whether personnel behavior complies with nuclear safety standards, thereby promptly identifying and correcting potential safety hazards.

[0053] Furthermore, the cloud-based analytics layer also possesses the capability for multi-source data fusion; it can integrate information from different data sources, such as video data and sensor data, organically fusing these originally scattered data; through multi-source data fusion, the system can obtain more comprehensive and reliable information, thereby improving the accuracy and reliability of the analysis; this fusion not only helps eliminate data redundancy and conflicts, but also reveals potential correlations between data, providing deeper insights for security management.

[0054] In this embodiment, the cloud analysis layer uses a CNN-LSTM behavior recognition model as a deep learning model. The CNN-LSTM behavior recognition model is a fusion model of convolutional neural network (CNN) and long short-term memory network (LSTM) to perform deep behavior analysis on video streams. CNN is responsible for extracting spatial features in the video, and LSTM is responsible for capturing dynamic changes in the time series. The combination of the two can more accurately identify complex human behavior patterns.

[0055] The specific implementation steps of the CNN-LSTM action recognition model are as follows:

[0056] The CNN-LSTM action recognition model combines the advantages of Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) to effectively process video or action data and achieve accurate action recognition. The following is a detailed implementation of this model:

[0057] The CNN part is responsible for extracting spatial features from video frames. It uses an improved CNN structure (such as ResNet, Inception, etc.) to extract high-level features from video frames layer by layer through multi-layer convolution and pooling operations, such as people's poses and movements.

[0058] The LSTM part is responsible for capturing the temporal features in the video sequence; it takes the spatial features extracted by the CNN as input and processes the temporal information in the video sequence through the LSTM network to understand the dynamic changes in behavior.

[0059] The training process of the CNN-LSTM action recognition model is as follows:

[0060] The training was conducted using a behavioral dataset specific to nuclear power plants, including samples of normal and abnormal behavior. The dataset was carefully labeled to ensure that each sample had a clear behavioral label.

[0061] Supervised learning methods are employed, and model parameters are optimized through backpropagation. To prevent overfitting, techniques such as data augmentation and regularization are used to improve the model's generalization ability.

[0062] Transfer learning is used to accelerate model convergence and improve recognition accuracy by leveraging CNN models pre-trained on large-scale datasets.

[0063] The specific behavior recognition process of the CNN-LSTM behavior recognition model is as follows:

[0064] (1) Receive video sequences or continuous behavioral data from the front-end device layer;

[0065] (2) First, spatial features are extracted through the CNN part, and then temporal features are captured through the LSTM part. During the processing, the CNN-LSTM behavior recognition model analyzes the video data frame by frame and considers the temporal relationship between frames.

[0066] (3) Output behavior recognition results, such as normal behavior, abnormal behavior, etc. At the same time, the CNN-LSTM behavior recognition model can also output the confidence score of the behavior, providing a basis for subsequent decision-making.

[0067] The CNN-LSTM action recognition model incorporates techniques such as attention mechanisms and multi-task learning to improve its robustness to complex environments and noise. For example, the attention mechanism allows the model to focus more on key regions related to the action and ignore background interference.

[0068] Furthermore, specific behavioral recognition applications should be developed based on the actual needs of nuclear power plant sites. Examples include fall detection and violation identification; these can further improve the safety of nuclear power plant sites and reduce the likelihood of accidents.

[0069] In addition, the cloud analytics layer can also fuse multi-source data; that is, it can integrate multi-source data from the front-end device layer, the edge computing layer and the cloud analytics layer itself, and perform comprehensive analysis to improve the accuracy and response speed of the entire video surveillance system in identifying abnormal events.

[0070] Furthermore, the cloud-based analytics layer will employ NLP to analyze the nuclear safety regulations knowledge graph and deeply integrate NLP with video analytics to construct an intelligent personnel behavior safety monitoring system for nuclear power plants. The nuclear safety regulations knowledge graph first undergoes deep semantic understanding using BERT encoding technology, outputting a spatiotemporal constraint rule base for operational steps.

[0071] Meanwhile, the surveillance video is processed and analyzed by the CNN-LSTM behavior recognition model; it can automatically detect human behavior in the video, identify safety risk behaviors, and record these behaviors in chronological order as an event sequence; the event sequence not only includes the behavior type, but also records the time and location of the behavior, etc., providing important data support for risk assessment.

[0072] The built-in rule engine in the cloud analytics layer receives spatiotemporal information from the event sequence of the knowledge graph, performs spatiotemporal matching and risk assessment; the rule engine has a wealth of security rules and risk assessment models, which can determine whether there are security risks based on the spatiotemporal relationship between device status and personnel behavior, and calculate the corresponding risk score.

[0073] The risk assessment formula for the rules engine is as follows;

[0074]

[0075] Where R is the final risk score; n is the total number of risk factors; i is the risk factor index; Wi is the risk weight associated with the i-th risk factor, set by nuclear safety experts; Di is the real-time radiation dose associated with the i-th risk factor; Dmax is the upper limit of the dose threshold; λ is the behavioral violation coefficient; Ei is the behavioral violation flag associated with the i-th risk factor; Rdi is the radioactive exposure coefficient associated with the i-th risk factor; and k is the risk amplification factor.

[0076] Finally, the risk scoring module generates a final risk score based on the matching results of the rules engine; this score intuitively reflects the safety risk level of on-site personnel behavior and provides a basis for safety management personnel to make decisions.

[0077] The cloud-based analytics layer utilizes deep learning and natural language processing technologies to achieve in-depth analysis of human behavior, providing the system with powerful intelligent analysis capabilities.

[0078] The application service layer is used to present the results of the cloud analytics layer to users in a user-friendly way and provide a wealth of application services.

[0079] The application service layer includes a radiation leak simulation alarm and an emergency evacuation guidance system.

[0080] The radiation leak simulation alarm is based on the predicted diffusion path of radiation data and achieves 3D visualization.

[0081] The emergency evacuation guidance system can dynamically generate disaster avoidance routes and push them to personnel positioning terminals.

[0082] In addition, the application service layer can also set up other application services to meet users' needs for system real-time performance, accuracy, and ease of use.

[0083] This invention employs a Convolutional Neural Network (CNN) for image feature extraction and a Long Short-Term Memory (LSTM) network for processing time-series data to improve the accuracy of anomaly detection. Real-time analysis of surveillance video is performed, and once an anomaly is detected, an early warning mechanism is immediately triggered, notifying relevant personnel through an alarm system. Simultaneously, edge computing is used to reduce data transmission latency, ensuring rapid response.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A video surveillance system for personnel behavior safety in nuclear power plant bases based on artificial intelligence, characterized in that, It includes the front-end device layer, edge computing layer, cloud analytics layer, and application service layer; The front-end equipment layer collects physical environment data of the nuclear power plant. The edge computing layer is deployed on a local server at the nuclear power plant. It preprocesses and performs preliminary behavioral analysis on the data collected by the front-end device layer. When an anomaly is detected, it triggers a local early warning. The cloud-based analytics layer utilizes deep learning and natural language processing technologies to perform in-depth analysis on preprocessed data uploaded from the edge computing layer, identifying complex human behavior patterns and potential security risks; it can also integrate information from different data sources. The cloud-based analysis layer will use NLP to analyze the nuclear safety procedure knowledge graph and deeply integrate NLP with video analysis. The nuclear safety procedure knowledge graph will first undergo deep semantic understanding through BERT encoding technology to output a spatiotemporal constraint rule base for operation steps. The surveillance video will be processed and analyzed by the CNN-LSTM behavior recognition model. It can automatically detect personnel behavior in the video, identify safety risk behaviors, and record the safety risk behaviors as an event sequence in chronological order. The application service layer generates a 3D visualized dynamic disaster avoidance path based on the results of the cloud analysis layer and pushes it to the personnel positioning terminal. The specific implementation steps of the edge computing layer are as follows: (1) Video stream preprocessing and radiation data preprocessing are the basic processing units of the edge computing layer. The basic processing unit performs gamma correction on the high-definition video stream through the hardware-accelerated image processing unit, dynamically adjusts the brightness and contrast of the screen to adapt to the complex lighting conditions in the nuclear power area, and uses 3D noise reduction algorithm to eliminate electronic noise and environmental interference in the monitoring screen. For radiation monitoring data, the built-in dose rate threshold detection engine verifies the validity of the sensor data in real time, marks radiation values ​​that exceed the safety threshold and triggers the early warning link. (2) A lightweight spatiotemporal feature extraction network is deployed in the edge computing layer. The lightweight spatiotemporal feature extraction network adopts an improved 3D convolutional neural network architecture, which extracts the spatial features of video frames and the motion features of the time series simultaneously while ensuring computational efficiency. (3) The anomaly detection module constructs a dual-modal fusion decision mechanism; in the first aspect, the behavior analysis results based on computer vision are input into LSTM, and the violation operation mode is identified through time series modeling; in the second aspect, the radiation data stream is passed through the Kalman filter for state estimation, and a dynamic early warning baseline is constructed by combining historical dose data.

2. The artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants according to claim 1, characterized in that, The data collected by the front-end device layer includes real-time video stream data obtained using a radiation-proof high-definition camera, environmental nuclear radiation data obtained using a nuclear radiation sensor, and access data obtained through a biological access control system.

3. The artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants according to claim 1, characterized in that, The dual-modal data of the anomaly detection module are synthesized with confidence scores at the decision fusion layer, ultimately generating element detection results that include anomaly type, location, and confidence score.

4. The artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants according to claim 1, characterized in that, When a local warning is triggered, for minor violations, an audible and visual alarm is issued through the local display terminal of the edge node, and the event timestamp and video clip are recorded; when a medium-risk behavior is detected, the regional broadcast system is triggered to issue a voice warning, and at the same time, a warning pop-up containing personnel identification information is pushed to the central control room; for major risks of radiation exceeding the standard, the system immediately activates the emergency response protocol.

5. A video surveillance system for personnel behavior safety in a nuclear power plant based on artificial intelligence, as described in claim 4, is characterized in that... Adjacent edge nodes share personnel trajectory data through a custom P2P communication protocol to build a cross-regional behavior correlation analysis network.

6. The artificial intelligence-based video surveillance system for personnel behavior safety in nuclear power plants according to claim 1, characterized in that, The rule engine built into the cloud analysis layer receives spatiotemporal information from the event sequence of the nuclear safety procedure knowledge graph, performs spatiotemporal matching and risk assessment; the rule engine determines whether there is a safety risk based on the spatiotemporal relationship between equipment status and personnel behavior, and calculates the corresponding risk score.

7. A video surveillance system for personnel behavior safety in a nuclear power plant based on artificial intelligence, as described in claim 6, is characterized in that, The event sequence includes the behavior type and records the spatiotemporal information of the behavior.

8. A video surveillance system for personnel behavior safety in a nuclear power plant based on artificial intelligence, as described in claim 6, is characterized in that... The risk assessment formula for the rule engine is as follows; ， Where R is the final risk score; n is the total number of risk factors; i is the risk factor index; Wi is the risk weight associated with the i-th risk factor, set by nuclear safety experts; Di is the real-time radiation dose associated with the i-th risk factor; Dmax is the upper limit of the dose threshold; λ is the behavioral violation coefficient; Ei is the behavioral violation flag associated with the i-th risk factor; Rdi is the radioactive exposure coefficient associated with the i-th risk factor; and k is the risk amplification factor.

Images