AI-based enterprise safety management method and system

Abstract

The invention provides an enterprise security management method and system based on AI, and relates to the technical field of security management, and the method comprises the steps: configuring a storage path and authority of a data storage unit, and initializing operation parameters of a calculation processing unit; acquiring images and / or video streams of the enterprise security control key area in real time; the real-time data is preprocessed; analyzing the preprocessed image and / or video stream by using the trained AI model, and identifying a preset target and action features thereof in enterprise production operation; comparing the recognized preset target and the action characteristics thereof with a preset early warning rule, and when the deviation exceeds a dynamic deviation threshold value, judging that the action is a wrong action, generating an early warning event and triggering an alarm mechanism; according to the method, data related to an early warning event is stored and examined, an optimization model and rules are fed back through an examination result, a security risk trend is analyzed, improvement suggestions are generated, and intelligentization, precision and high efficiency of enterprise security management are realized.

Description

Technical Field

[0001] This application relates to the field of security management technology, and in particular to AI-based enterprise security management methods and systems. Background Technology

[0002] In the process of enterprise production and operation, safety management is the core link to ensure the safety of personnel and avoid property loss. Traditional enterprise safety management mainly relies on manual inspection, fixed threshold monitoring and post-event traceability, which has significant limitations. The manual inspection mode is limited by the energy and experience of the inspectors, resulting in limited coverage, delayed response and high missed inspection rate. Especially in high-risk operation areas (such as chemical workshops and heavy machinery operation areas), manual inspection itself may also bring safety risks.

[0003] In recent years, AI-based technologies such as target detection and behavior analysis have provided new means for safety management. However, existing technologies still have the following key shortcomings: Current AI safety management systems typically employ fixed model accuracy and computing power allocation strategies, failing to dynamically adjust according to real-time scenarios. For example, applying the same processing accuracy to high-risk targets (such as personnel intruding into dangerous areas) and low-risk targets (such as regular vehicle movement) can lead to delayed warnings in high-risk scenarios due to insufficient computing power, while wasting computing resources in low-risk scenarios. Traditional safety management systems rely on fixed action templates and thresholds, making it difficult to adapt to individual differences in actions (such as different worker operating habits) and new types of violations (such as undefined abnormal behaviors). Enterprise security data involves sensitive information (such as monitoring of critical areas), but existing systems typically use simple role-based access control, lacking dynamic path generation and auditing mechanisms, making them prone to data leaks or read / write failures due to incorrect access configurations. Most AI safety systems only provide single warnings and lack feedback and optimization processes for misjudgments, leading to recurring similar problems. Summary of the Invention

[0004] Based on the above problems, this application proposes an AI-based enterprise security management method and system, which can integrate multi-dimensional risk assessment, dynamic resource scheduling, and intelligent threshold optimization to solve problems such as insufficient recognition accuracy, unreasonable resource allocation, and weak risk prediction ability in existing technologies, thereby achieving intelligent, precise, and efficient enterprise security management.

[0005] The objective of this application is achieved through the following technical solution:

[0006] Firstly, this application provides an AI-based enterprise security management method, the method comprising:

[0007] Configure the storage path and permissions of the data storage unit, and initialize the operation parameters of the computing unit;

[0008] Real-time acquisition of images and / or video streams from key areas of enterprise security management; and preprocessing of real-time data;

[0009] The trained AI model is used to analyze preprocessed images and / or video streams to identify preset targets and their motion characteristics in enterprise production operations.

[0010] The identified preset targets and their action characteristics are compared with preset early warning rules. When the deviation exceeds the dynamic deviation threshold, it is judged as an erroneous action, an early warning event is generated, and an alarm mechanism is triggered.

[0011] The system stores and reviews data related to early warning events, uses the review results to optimize models and rules, analyzes safety risk trends, and generates improvement suggestions.

[0012] Preferably, the analysis of the preprocessed images and / or video streams using the trained AI model includes:

[0013] The GPU-accelerated AI inference framework performs target detection and classification on a single frame of image or video, and outputs structured data of the target, including coordinates, category labels and confidence scores.

[0014] Based on the real-time reasoning results of multiple consecutive frames, action semantic tags are generated through a temporal behavior analysis model;

[0015] The model accuracy and GPU computing power allocation are dynamically adjusted based on the target risk level.

[0016] Preferably, the GPU-accelerated AI inference framework performs target detection and classification on a single frame of image or video, outputting structured data of the target, including:

[0017] The video frames are processed by a TensorRT-accelerated object detection model, and the object detection results are output.

[0018] A target tracking model is used to assign a unique ID to the detected target and associate it with cross-frame trajectories.

[0019] Preferably, the GPU-accelerated AI inference framework performs target detection and classification on a single frame of image or video, outputting structured data of the target, including:

[0020] The shallow feature maps of the object detection model are reused in the object tracking model;

[0021] The target tracking model integrates the target's motion trajectory vector and appearance feature vector, and maintains the continuity of the target ID through weighted similarity calculation; the appearance feature vector originates from the shallow feature map.

[0022] Preferably, the generation of action semantic tags based on real-time inference results from multiple consecutive frames through a temporal behavior analysis model includes:

[0023] A safety rule knowledge graph is established, where graph nodes represent standard action types and edges between nodes represent action association constraints. The safety rule knowledge graph is associated with preset early warning rules. A spatiotemporal self-attention network is used to perform weighted processing on continuous video frames, focusing on key action frames and extracting frame features whose weights exceed a set threshold.

[0024] The continuous trajectories of the same target are input into a spatiotemporal convolutional network, and the frame features and trajectory dynamic information are fused to generate a spatiotemporal feature vector.

[0025] Based on the action analysis template associated with the preset early warning rules in the safety rule knowledge graph, the similarity between the spatiotemporal feature vector and the template is calculated, and the action deviation value is output.

[0026] Preferably, the target risk level is determined by the following methods:

[0027] A preset risk assessment matrix is ​​provided, which includes target category weights, action risk coefficients, and environmental sensitivity.

[0028] The target category is obtained based on the target detection results, and the corresponding action risk coefficient is queried by combining the action semantic tags.

[0029] Environmental sensitivity is determined by combining the spatial distance between the target and the hazardous area with the degree of environmental disturbance;

[0030] The target risk level is determined based on the target category weight, action risk coefficient, and environmental sensitivity.

[0031] Preferably, comparing the identified preset targets and their motion characteristics with preset warning rules includes:

[0032] An adaptive template evolution mechanism is adopted to periodically analyze historical misjudgment cases, automatically identify reasonable action variation types not covered in the safety action templates, and generate new template sub-libraries through reinforcement learning and merge them with the original template library;

[0033] A dual-loop threshold optimization algorithm is adopted. The inner loop dynamically fine-tunes the deviation threshold based on the real-time identified action deviation distribution, while the outer loop combines safety accident statistics and expert evaluation results to calibrate the threshold adjustment strategy.

[0034] Preferably, the dynamic deviation threshold setting is determined by the following methods:

[0035] Establish a basic threshold library, which includes minimum safe distance and maximum permissible speed parameters for different target categories;

[0036] An environmental correction factor is introduced, which is dynamically adjusted based on parameters such as light intensity, visibility, and population density.

[0037] The final dynamic deviation threshold is determined by using the base threshold, environmental correction factor, and risk level coefficient.

[0038] Preferably, the storage and review of data related to early warning events, and the feedback of review results to optimize models and rules, includes:

[0039] Video clips related to the early warning event are extracted, stored in a standard format, and the structured metadata of the early warning event is recorded in the database.

[0040] The database is updated based on the occurrence time of each warning event, the correlation between each warning event, the frequency of occurrence of each warning event, and the severity of the warning event.

[0041] The warning events are reviewed through the review and playback interface to obtain the review results; the AI ​​model and dynamic deviation threshold are iteratively optimized based on the review results.

[0042] On the other hand, this application provides an AI-based enterprise security management system for implementing the AI-based enterprise security management method of the embodiments of this application. The system includes:

[0043] The configuration module is used to configure the storage path and permissions of the data storage unit and initialize the operation parameters of the computing unit.

[0044] The acquisition module is used to acquire images and / or video streams from key areas of enterprise security management in real time; and to preprocess the real-time data.

[0045] The identification and analysis module is used to analyze pre-processed images and / or video streams using a trained AI model to identify preset targets and their motion characteristics in enterprise production operations.

[0046] The early warning module is used to compare the identified preset targets and their action characteristics with preset early warning rules. When the deviation exceeds the dynamic deviation threshold, it is judged as an erroneous action, an early warning event is generated, and an alarm mechanism is triggered.

[0047] The storage module is used to store and review data related to early warning events, and to optimize models and rules based on the review results, analyze security risk trends, and generate improvement suggestions.

[0048] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored on and executable on the memory, wherein the processor executes the program to implement the steps of any of the methods described in this application.

[0049] The beneficial effects of this invention include:

[0050] From data acquisition and preprocessing (adaptive noise reduction, key area clipping) to AI analysis (dynamic precision inference, time-series behavior analysis), and then to early warning response (tiered alarms, data storage optimization) and model iteration (optimizing thresholds and templates based on review results), a complete closed loop of "perception-analysis-decision-optimization" is formed. Database update strategies such as time decay weighting and correlation clustering ensure efficient reuse of high-value risk data, providing reliable support for security risk trend analysis and helping enterprises shift from passive response to proactive prevention. Dynamic deviation thresholds are calculated in real-time using a linkage of "basic threshold × environmental correction factor × risk level coefficient," adapting to environmental factors such as light intensity and personnel density, maintaining accuracy even under harsh conditions. The security rule knowledge graph automatically expands with new violations, combined with a dual-closed-loop optimization mechanism, enabling... The early warning rule update cycle has been shortened from weekly to hourly, enabling rapid response to new work specifications or sudden risk scenarios. This is particularly suitable for multi-position, highly dynamic industrial production environments. The technology of reusing shallow features from the target detection model to the tracking model, combined with the weighted fusion of motion trajectory vectors and appearance feature vectors, effectively maintains the continuity of target IDs in occluded and rapidly moving scenarios, solving the ID jump problem in traditional tracking and providing stable input for motion feature analysis. A multi-dimensional risk assessment matrix (integrating target category weights, motion hazard coefficients, and environmental sensitivity) breaks through the dependence on a single parameter, making risk level judgments more closely aligned with actual scenarios. An adaptive template evolution mechanism generates new template sub-libraries through reinforcement learning, and a dual-closed-loop threshold optimization algorithm dynamically calibrates the judgment criteria, significantly reducing misjudgments caused by insufficient template coverage or rigid thresholds. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of an AI-based enterprise security management method provided in an embodiment of this application.

[0052] Figure 2 This is a schematic diagram of an AI-based enterprise security management system provided in an embodiment of this application. Detailed Implementation

[0053] The present application will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0054] See attached document Figure 1 This application provides an AI-based enterprise security management method, the method comprising:

[0055] Configure the storage path and permissions of the data storage unit; initialize the computing parameters based on historical computing power demand prediction, including setting GPU resource allocation, model accuracy, and frame rate threshold;

[0056] Install and debug the acquisition equipment at the preset location; ensure that the image covers the work area; and acquire images and / or video streams of key areas for enterprise safety management in real time through the acquisition equipment.

[0057] The acquired images and / or video streams are preprocessed, including denoising, contrast enhancement, cropping key regions, frame sampling, and resolution scaling, to improve data quality to meet the needs of subsequent AI analysis. Among these processes, an adaptive denoising algorithm is used to dynamically adjust the denoising parameters based on the noise characteristics of the images or videos.

[0058] The trained AI model is used to analyze preprocessed images and / or video streams to identify preset targets and their motion characteristics in enterprise production operations, and compare them with preset early warning rules; wherein, the preset targets include specific objects such as personnel and vehicles; the motion characteristics include relevant dynamic information such as action behavior and movement trajectory;

[0059] When the deviation between the action feature and the preset warning rule exceeds the dynamic deviation threshold, it is determined to be an erroneous action; the deviation threshold is dynamically optimized based on job safety specifications and historical data statistical analysis.

[0060] Based on the erroneous action, an early warning event is generated and an alarm mechanism is triggered. The warning is issued on-site through an audible and visual alarm device, whose volume and flashing frequency are automatically adjusted according to the ambient noise and light. At the same time, an alarm notification containing information such as the time and location of the event is sent to relevant management personnel.

[0061] The working principle of the above technical solution is as follows:

[0062] First, by configuring the path and permissions of the data storage unit, the standardized storage and access control of secure data are ensured. At the same time, based on historical computing power demand prediction, the GPU resource allocation, model accuracy, frame rate threshold and other computing parameters are initialized to provide suitable hardware resource support for subsequent AI analysis. The acquisition equipment is installed and debugged in the preset position to ensure that the lens covers the key areas of operation (such as the operating table in the factory workshop, the dangerous goods storage area in the warehouse, etc.), providing basic hardware guarantee for data acquisition.

[0063] The acquisition device captures images or video streams of key areas in real time. These streams are then fed into the video stream via RTSP, the standard real-time streaming protocol for network cameras. Access tools include FFmpeg, an industry-standard multimedia framework used to stably pull video streams from various RTSP sources. The video streams are decoded (H.264 / H.265 -> raw frames). Key preprocessing is performed, including frame sampling (extracting analysis frames at a set frequency, such as 5-10 FPS) and resolution scaling (scaling to the AI ​​model input size, such as 640x640). The processed frame data is then passed to the AI ​​analysis engine. A load balancer is deployed on the server side to distribute the RTSP streams from different cameras to multiple processing nodes, ensuring system throughput and stability.

[0064] After preprocessing and optimization, the raw data enters the AI ​​analysis stage. Through adaptive denoising algorithms (dynamically adjusting parameters according to image noise characteristics, such as enhancing and smoothing particulate noise caused by workshop dust, and improving edge preservation ability for low light noise in night monitoring), contrast enhancement, key area cropping (focusing on core areas such as operating consoles and passageways), and frame sampling (balancing real-time performance and computing power consumption), the data quality is improved, laying the foundation for accurate analysis by AI models.

[0065] The trained AI model analyzes the preprocessed data. On the one hand, it identifies preset targets in production operations (such as operators and transport vehicles) and their motion characteristics (such as personnel's operating postures and vehicle driving trajectories). On the other hand, it compares these characteristics with preset warning rules (such as personnel must not touch machine tools without protective gloves and vehicles must not speed in passageways). During the comparison process, a dynamic deviation threshold (dynamically optimized based on job safety regulations and historical data) is used as the judgment standard. When the difference between the actual motion characteristics and the rules exceeds this threshold, it is judged as an erroneous motion.

[0066] For example, in the operating area of ​​a chemical plant reactor, the system collects video streams of operators not wearing protective masks; the preprocessing stage uses an enhancement algorithm to highlight the facial features of the personnel; the AI ​​model identifies the personnel target and the action characteristics of not wearing a mask, compares them with the preset warning rule that operators must wear protective masks, and calculates the deviation value (here, the complete deviation between not wearing and wearing is required). If the deviation exceeds the dynamic threshold (because the reactor area is a high-risk position, the threshold is set strictly), it is judged as an erroneous action.

[0067] After an erroneous action triggers a warning event, the system activates a multi-layered alarm mechanism. The on-site audible and visual alarm devices dynamically adjust according to the environment (e.g., automatically increasing the volume to 90 decibels when there is high noise in the workshop, and increasing the flashing frequency to 3 times per second in strong light environments) to ensure that on-site personnel can quickly perceive the event. At the same time, a notification containing the event time and location (e.g., the No. 3 reactor area, 14:23) is pushed to the management personnel to achieve a multi-dimensional response.

[0068] The effects of the above technical solution are as follows:

[0069] The adaptive denoising algorithm in the preprocessing stage is dynamically optimized for noise in different scenarios to reduce the impact of environmental interference on data quality. The AI ​​model, combined with dynamic deviation thresholds (optimized based on job specifications and historical data), avoids false alarms (such as appropriately relaxing the tolerance for action deviations for low-risk positions) or false alarms (strictly controlling the deviation range for high-risk positions) caused by fixed standards, significantly improving the accuracy of erroneous action identification.

[0070] Precise deployment of data acquisition equipment ensures no blind spots in key areas; audible and visual alarm devices dynamically adjust according to environmental noise and lighting to solve the problem of traditional alarms failing in complex industrial environments (such as insufficient alarm volume in high-noise workshops or inconspicuous flashing under strong light); the initial configuration of storage and computing power is optimized based on historical data to avoid resource waste (such as reducing model accuracy in low-risk areas to save GPU computing power).

[0071] The dynamic deviation threshold is continuously iterated based on historical data, enabling the early warning rules to be constantly adapted to actual production scenarios; the multi-layered alarm mechanism triggered by erroneous actions (on-site warning + management personnel notification) shortens response time and reduces the probability of accidents; at the same time, standardized data storage and access control provide reliable data support for subsequent safety risk analysis (such as mining high-frequency risk points through historical early warning events), forming a closed-loop management of identification-early warning-optimization.

[0072] Automated AI analysis replaces traditional manual monitoring, reducing manpower input; dynamic configuration of computing power and storage (such as storing only key frames for low-frequency early warning events) reduces hardware resource consumption; preset early warning rules and templated deployment adapt to the needs of different positions, reducing system customization costs and facilitating rapid implementation in various types of enterprises (such as factories, warehouses, and construction sites).

[0073] In one possible implementation, the storage path and permissions of the configuration data storage unit include:

[0074] An intelligent permission matrix algorithm is introduced to automatically generate a storage path hierarchy structure and permission allocation scheme based on the enterprise's organizational structure and job functions.

[0075] By using a pre-defined job-permission mapping database, the system automatically matches the corresponding storage path access permissions when adding a new device or user, and generates a permission configuration report for administrator review.

[0076] Blockchain technology is used to record permission change logs to ensure that permission adjustments are traceable and to prevent unauthorized modifications.

[0077] In one possible implementation, the operational parameters of the initialization computation processing unit include:

[0078] Develop a dynamic computing power scheduling model and construct a computing power demand prediction model by analyzing the computing power consumption characteristics of different scenarios in historical job data.

[0079] The model automatically recommends computational parameters based on the current job type and monitors the load of the computing unit in real time. When the load exceeds the threshold, it automatically adjusts the parameter allocation and migrates some non-real-time tasks to the cloud computing resource pool to achieve elastic collaboration between edge and cloud computing power.

[0080] The working principle and effects of the above technical solution are as follows:

[0081] An intelligent permission matrix algorithm is introduced, which automatically constructs a multi-dimensional storage path hierarchy structure based on the enterprise's organizational structure (such as department levels and job sequences) and job functions (such as safety administrators, front-line operators, and system maintainers). For example, directories are divided according to factory area → workshop → equipment area → data type (video stream / structured data), ensuring that the data storage logic is consistent with the enterprise's management architecture. At the same time, the algorithm automatically assigns corresponding storage path access permissions (read / write / modify) to different roles based on a preset job-permission mapping relationship library (such as safety administrators can access data throughout the factory, while operators can only access data in their own workshop), avoiding the subjectivity and oversight of manual configuration.

[0082] When a new device (such as a new surveillance camera) or user is added, the permission matching mechanism is automatically triggered. Based on the device's region or the user's role, the corresponding permission template is retrieved from the mapping database, storage path access rules are generated, and a permission configuration report (including permission scope, effective time, and associated data type) is generated simultaneously for administrator review and confirmation before taking effect. During the permission change process, blockchain technology records the entire operation log (including the person making the change, the content of the change, and the timestamp), forming an immutable distributed ledger to ensure that any permission adjustment is traceable and to prevent unauthorized modifications (such as an operator unilaterally elevating data access permissions).

[0083] The dynamic computing power scheduling model analyzes historical job data (such as video stream frame rates, target density, and AI model inference time in different workshops and time periods) to extract computing power consumption characteristics for various scenarios (e.g., the computing power consumption in a welding workshop is twice that of a normal workshop due to the high frame rate required for spark recognition). This data is then used to construct a computing power demand prediction model. During system initialization, the model automatically recommends computational parameters based on the current job type (e.g., early shift production, equipment maintenance), including GPU core allocation ratio, model inference accuracy (FP32 / FP16), and video stream frame sampling frequency, ensuring that the initial configuration matches the actual computing power requirements.

[0084] During operation, the monitoring module collects real-time load metrics of the computing units (such as GPU utilization, memory usage, and task queue length). When the load exceeds a preset threshold (e.g., GPU utilization ≥ 85%), the dynamic computing power scheduling model automatically initiates a resource migration mechanism: it filters non-real-time tasks (such as historical data backtracking analysis and offline trajectory optimization of low-priority targets) and migrates them to the cloud computing resource pool, releasing the computing power of edge nodes; once the load on the edge nodes drops below the threshold (e.g., GPU utilization ≤ 60%), it then migrates the cloud tasks back according to task priority, achieving elastic collaboration between edge and cloud computing power and avoiding overload of a single node.

[0085] In one possible implementation, the analysis of the preprocessed images and / or video streams using the trained AI model includes:

[0086] Real-time inference computation: Through a GPU-accelerated AI inference framework, target detection and classification are performed on single-frame image / video frames, and the structured data of the target is output, including coordinates, category labels and confidence scores;

[0087] Action semantic parsing: Based on real-time reasoning results from multiple consecutive frames, action semantic tags are generated through a temporal behavior analysis model;

[0088] The model accuracy and GPU computing power allocation are dynamically adjusted based on the target risk level.

[0089] The real-time inference computation includes:

[0090] Video frames are processed using a TensorRT-accelerated object detection model (such as YOLOv8) to output object detection results;

[0091] A target tracking model (such as DeepSORT) is used to assign a unique ID to the detected target and associate it with cross-frame trajectories.

[0092] In one possible implementation, the target risk level is determined by the following methods:

[0093] A preset risk assessment matrix is ​​provided, which includes target category weights, action risk coefficients, and environmental sensitivity.

[0094] The target category is obtained based on the target detection results, and the corresponding action risk coefficient is queried by combining the action semantic tags.

[0095] Environmental sensitivity is calculated by combining the spatial distance between the target and the danger zone and the degree of environmental interference. Specifically, the distance between the target and the danger zone is calculated using the spatial coordinates of the camera.

[0096] The target risk level is determined based on the target category weight, action risk coefficient, and environmental sensitivity.

[0097] In one possible implementation, a risk level coefficient is calculated by weighting the target category weight, the action hazard coefficient, and the environmental sensitivity, and the target risk level is determined based on the level coefficient.

[0098] The working principle and effects of the above technical solution are as follows:

[0099] First, an object detection model (such as YOLOv8) scans the image pixel matrix to locate the spatial coordinates of preset targets (personnel, vehicles, etc.), and outputs target category labels (such as "person wearing a safety helmet" or "forklift") and confidence scores (reflecting recognition reliability, ranging from 0 to 1). The classification model further refines target attributes (such as whether the vehicle is speeding or whether the person is carrying tools), and finally integrates them into structured data, with an example format: {coordinates: (x1, y1, x2, y2), category: "forklift", confidence score: 0.92, speed: 5 m / s}. The entire process is implemented in parallel computation using the CUDA cores of the GPU.

[0100] Motion features between frames are extracted using a 3D convolutional network (C3D) (such as the swing amplitude of a person's arm and changes in the direction of a vehicle's travel). Combined with the self-attention mechanism of the Transformer, key action frames are focused (such as the moment a person exerts force when climbing a guardrail), and scattered action fragments are mapped to high-level semantic labels (such as "illegal climbing", "normal operation", "vehicle driving in the wrong direction"). For example, when the feature combination of "person's hand touching high-voltage equipment + body leaning forward + both feet off the ground" is detected in consecutive frames, the model automatically generates the semantic label "illegal electric shock risk action", realizing the transformation from pixel-level features to behavior-level understanding.

[0101] Combining a pre-set risk assessment matrix, the risk level coefficient (range 0-1) is calculated by weighting the target category, action semantic tags, and environmental sensitivity (such as the distance of the target from the danger zone) obtained through real-time inference, and then mapped to three levels of risk: high, medium, and low. A resource scheduling mechanism is triggered based on the risk level: high-risk targets (e.g., risk level coefficient ≥ 0.7) automatically activate FP32 full-precision inference mode, allocating dedicated GPU cores to ensure detection accuracy; medium-risk targets (e.g., risk level coefficient between 0.4 and 0.7) use FP16 mixed precision, sharing GPU resources; low-risk targets (e.g., risk level coefficient ≤ 0.4) switch to INT8 lightweight precision to reduce computational power consumption. For example, "personnel without protective masks" (high risk) in the chemical plant reactor area triggers full-precision inference, while "stationary vehicles" (low risk) outside the warehouse are processed in lightweight mode, achieving a dynamic balance between accuracy and efficiency.

[0102] The rules for determining the target category weights include:

[0103] Based on the principle of "personal safety first, critical equipment second", the analytic hierarchy process is used to divide the first-level categories (personnel, special equipment, general equipment / items) and assign basic weights. Under the first category, the second-level categories are carried out according to the differences in the impact of the target on safety risks (such as personnel according to the risk exposure intensity of the job, equipment according to the operating status, and items according to the hazard attributes). The weights are dynamically adjusted in combination with historical accident data to achieve differentiated quantification of the basic risk value.

[0104] Primary categories: Personnel (weight e.g. 0.6), Special equipment (e.g., forklifts, cranes, weight e.g. 0.3), General equipment / items (e.g., tools, materials, weight e.g. 0.1).

[0105] Secondary subdivision:

[0106] Personnel: Based on job safety responsibilities and risk exposure intensity, personnel in high-risk positions have a higher risk due to direct contact with dangerous operations (such as welding and high-pressure operations), while external visitors have a lower risk due to unfamiliarity with the environment, and the risk of ordinary operators is relatively stable.

[0107] Special equipment: Based on the equipment's operating status, there are real-time risks associated with the equipment's energy release and mechanical movement during operation, and the risk of shutting down the equipment is significantly reduced.

[0108] Ordinary items: According to the inherent hazardous properties of the items themselves, flammable materials have inherent risks such as spontaneous combustion and explosion, while conventional tools have no special hazardous properties.

[0109] Personnel: Operators (weight, e.g., 0.6 × 1.0), visitors (weight, e.g., 0.6 × 0.8), personnel in high-risk positions (e.g., welders, weight, e.g., 0.6 × 1.2);

[0110] Special equipment: Equipment in operation (weight, for example, 0.3 × 1.2), equipment out of service (weight, for example, 0.3 × 0.5);

[0111] Common items: flammable materials (weight, for example, 0.1 × 1.5), and common tools (weight, for example, 0.1 × 0.5).

[0112] Based on the company's historical accident data, if the proportion of accidents caused by a certain type of target within the first preset time period (e.g., the past year) exceeds the first proportion (e.g., 40%), its weight will be increased by the first percentage (e.g., 20%) (e.g., if forklift accidents occur frequently in a certain factory area, the weight of special equipment will be adjusted from 0.3 to 0.36); conversely, if there are no related accidents within the first preset time period (e.g., two consecutive years), the weight will be decreased by the second percentage (e.g., 10%).

[0113] Risk factor of action calculated using a multi-dimensional weighted method:

[0114] The risk factor of an action includes multiple dimensions such as the severity of the consequences, frequency of occurrence, urgency, and history of violations. It can be scored on a percentage scale and then normalized to the 0-1 range. For example, the specific calculation method is shown in Table 1:

[0115] Table 1

[0116]

[0117]

[0118] Environmental sensitivity is based on the spatial distance between the target and the hazardous area (the closer the distance, the higher the sensitivity) and the degree of environmental interference (such as the impact of lighting, visibility, and obstruction on risk perception). It is calculated by multiplying the distance sensitivity coefficient and the environmental correction factor, and comprehensively reflects the degree of influence of environmental factors on risk assessment.

[0119] Environmental sensitivity is a weighted average of the distance sensitivity coefficient and the environmental disturbance coefficient;

[0120] Among them, the environmental interference coefficient (used to quantify the degree of interference of environmental factors such as lighting, visibility, and occlusion on the "accuracy of risk perception" (the stronger the interference, the larger the coefficient)) is determined based on multi-parameter normalization fusion, as follows:

[0121] By grading, quantizing and weighting the illumination, visibility and occlusion rates, the output coefficient values ​​are in the range of [0, 2].

[0122] Based on the target occlusion rate (the proportion of the target area that is occluded) output by the target detection algorithm, the environmental interference coefficient Koccl is calculated according to the occlusion rate.

[0123] For example: Occlusion rate ≤10%: Koccl=0 (no interference, motion features complete);

[0124] Occlusion rate 10%-50% (partial occlusion, such as the upper body of a person being blocked by equipment): Koccl = (occlusion rate - 10%) / 40% (e.g., Koccl = 0.5 when there is 30% occlusion);

[0125] Occlusion rate > 50% (severe occlusion, loss of key action features), Koccl = 1.0.

[0126] For example, the distance sensitivity coefficient (calculated based on Euclidean distance):

[0127] Hazardous areas (such as high-voltage electrical boxes and chemical storage tanks) should have a pre-set safety radius R (usually 5-10 meters):

[0128] Distance ≤ 0.3R → coefficient 1.0 (extremely high sensitivity);

[0129] 0.3R < distance ≤ 0.6R → coefficient 0.7 (high sensitivity);

[0130] 0.6R < distance ≤ R → coefficient 0.4 (medium sensitivity);

[0131] Distance > R → coefficient 0.1 (low sensitivity);

[0132] In another possible implementation, the distance sensitivity coefficient is dynamically mapped based on the spatial distance between the target and the danger zone, with a value range of [0,1].

[0133] When the distance is greater than or equal to the safety critical distance (e.g., 10m), the distance sensitivity coefficient is 0 (distance does not affect sensitivity).

[0134] When entering a dangerous area, the distance sensitivity coefficient = 1 (distance influence is maximized);

[0135] The intermediate distance is mapped by a linear or exponential function.

[0136] In terms of target category weights, the principle of "personal safety first, key equipment second" is followed and dynamically adjusted in conjunction with historical accident data. This not only adheres to the basic logic of safety management but also precisely quantifies the basic risks of different targets through secondary subdivision, making risk assessment more in line with actual scenarios and providing a reasonable basis for subsequent calculations.

[0137] The risk factor of an action is determined by weighting multiple dimensions such as the severity of the consequences and the frequency of occurrence, breaking through the limitations of a single dimension and comprehensively considering the risk of an action, making the assessment of the degree of action risk more accurate and effectively identifying high-risk actions.

[0138] A weighted average formula is used to fuse the distance sensitivity coefficient and the environmental interference coefficient, and full-process traceability is achieved through clear hierarchical quantification rules. The environmental interference coefficient is based on the normalized weighted fusion of illumination, visibility, and occlusion rate, and the output range is strictly controlled within [0,2]. The distance sensitivity coefficient is locked within [0,1] through piecewise segmentation or linear mapping of the safety radius. After weighting, the two coefficients naturally converge to a reasonable range through implicit constraints (coefficient non-negativity and weight normalization), avoiding extreme values ​​that could lead to sensitivity distortion and providing a stable quantitative basis for risk level assessment.

[0139] The distance sensitivity coefficient accurately captures the physical proximity risk between the target and the dangerous area using Euclidean distance (such as an extremely high sensitivity setting within 0.3R), while the environmental interference coefficient further subdivides the impact of illumination, visibility, and occlusion rate on the "risk identification accuracy" (such as a fine quantification of Koccl=0.5 when there is 30% occlusion). The dual-dimensional collaborative coverage of "physical distance risk" and "environmental perception interference" solves the one-sidedness of single-dimensional assessment (such as ignoring the risk omission caused by fog and occlusion when only looking at distance).

[0140] The distance sensitivity coefficient supports two modes: segmented threshold (e.g., multi-interval division of the safety radius R) and function mapping (linear / exponential). It can be flexibly switched according to the characteristics of hazardous areas (e.g., high-voltage electrical boxes require strict radius control, while chemical storage tanks require gradual sensitivity). The environmental interference coefficient achieves a smooth transition through a segmented formula for the occlusion rate (e.g., linear calculation within the 10%-50% interval), avoiding coefficient jumps caused by small fluctuations in the occlusion rate. The combination of these two features enables the environmental sensitivity to respond in real-time to target movement (distance changes) and sudden environmental changes (e.g., sudden dense fog, temporary occlusion), adapting to the complex and dynamic risk environments in industrial settings.

[0141] The weighted averaging mechanism avoids the dominance of extreme values ​​for any single coefficient in the results: for example, when a target is far from a danger zone (distance sensitivity coefficient = 0.1) but encounters extreme occlusion (environmental interference coefficient = 2.0), the sensitivity can still reflect the impact of environmental interference through weight allocation; conversely, when a target enters a high-risk area (distance sensitivity coefficient = 1.0) but the environment is favorable (environmental interference coefficient = 0), the sensitivity remains high. This design ensures that a single factor (such as excessive distance or an excellent environment) will not cause the risk assessment to fail, thus improving the robustness of the system.

[0142] By combining the three factors to calculate the target risk level, the system achieves the organic integration of multi-dimensional risk factors, accurately classifies risk levels, and provides a reliable basis for subsequent dynamic adjustment of model accuracy and GPU computing power allocation. This helps to optimize resource allocation while ensuring security, and improve the intelligence and precision of enterprise security management.

[0143] In one possible implementation, the real-time inference computation includes:

[0144] The shallow feature map output by the object detection model is reused in the object tracking model;

[0145] The motion trajectory vector and appearance feature vector of the target are fused, and the continuity of the target ID across frames is maintained by weighted similarity calculation.

[0146] The working principle and effects of the above technical solution are as follows:

[0147] When processing video frames, object detection models extract shallow feature maps (such as low-level visual features like edges and textures) through convolutional neural networks (CNNs). These maps are not only used for object localization and classification but also directly reused in object tracking models via feature sharing interfaces. This reuse avoids the tracking model from repeatedly extracting the same features (e.g., when detecting and tracking simultaneously in the same frame, there's no need to calculate edge features twice), reducing redundant computation while ensuring that detection and tracking are based on common features, thus improving the consistency of cross-model collaboration.

[0148] The target tracking model maintains ID continuity through two core vectors:

[0149] Motion trajectory vector: Based on the trajectory information such as the target's motion speed and direction in the previous frame, predict the possible position in the current frame to form spatiotemporal motion constraints;

[0150] Appearance feature vector: From the reused shallow feature map, higher-order features (such as target color and shape details) are further extracted through CNN to form a unique appearance identifier of the target.

[0151] Both methods use a weighted similarity algorithm (the motion vector weights are dynamically adjusted according to the stability of the target's motion, such as increasing the weight to 0.6 when moving quickly and decreasing it to 0.3 when stationary) to calculate the inter-frame matching degree. When the similarity exceeds the threshold, it is determined to be the same target and the ID remains unchanged, thus solving the ID jump problem in scenarios such as occlusion and temporary disappearance.

[0152] In one possible implementation, the real-time inference computation further includes:

[0153] The shallow feature map reused in the target tracking model is processed by the feature attention module. The feature attention module locates the key regions of the target through a self-attention mechanism and performs weighted enhancement on the shallow features to highlight highly discriminative detail features.

[0154] The target tracking model also includes a feature layer adaptive selector, which automatically switches the reused feature level according to the scene complexity. In simple scenes, shallow features are reused, and in complex scenes, mid-level features are called to balance details and efficiency.

[0155] The weights of the weighted similarity algorithm are dynamically adjusted through a deep reinforcement learning model. The deep reinforcement learning model uses the inter-frame ID matching accuracy as a reward signal to optimize the weight ratio of motion trajectory vectors and appearance feature vectors in scenarios such as target acceleration, occlusion, and intersection.

[0156] In one possible implementation, the weight adjustment strategy of the deep reinforcement learning model also incorporates real-time analysis results from the spatiotemporal attention network, specifically including:

[0157] The spatiotemporal attention network extracts multi-dimensional features from the target motion state (velocity, acceleration), appearance feature confidence (feature entropy value), and scene occlusion rate (proportion of the target occluded area) in the current frame and the preceding N frames (N≥3).

[0158] Based on the extracted features, a fusion weight coefficient between the motion trajectory vector and the appearance feature vector is dynamically generated and output to the weighted similarity algorithm.

[0159] When the scene occlusion rate is greater than 50%, the spatiotemporal attention network automatically increases the weight ratio of the appearance feature vector; when the target's motion state is unstable (absolute value of acceleration > preset threshold), the weight ratio of the motion trajectory vector is increased.

[0160] In one possible implementation, the weighted similarity algorithm further includes a conflict detection mechanism. When the deviation between the target position predicted by the motion trajectory vector and the target position matched by the appearance feature vector exceeds a preset threshold, local feature re-extraction is triggered to correct the matching result.

[0161] The target tracking model has a target feature memory, which stores the typical appearance features of the target in a stable state. When the target occlusion ends, the target ID is re-associated by comparing the features in the memory with the current features.

[0162] The target tracking model integrates the target's motion trajectory vector, appearance feature vector, and behavior feature vector to construct a three-dimensional matching space, and learns the dynamic correlation weights of the three in different scenarios through a graph neural network.

[0163] The target tracking model mines the ID invariance features of the target through self-supervised learning. When conventional feature matching fails, the ID invariance features are used to maintain the continuity of the target ID across frames.

[0164] The working principle and effects of the above technical solution are as follows:

[0165] The shallow feature maps extracted by the target detection model are first processed by the feature attention module, which focuses on key target regions (such as people's faces and vehicle markings) through a self-attention mechanism. High-discrimination features (such as helmet color and equipment texture) are weighted and enhanced, while background interference is weakened. At the same time, the feature layer adaptive selector dynamically switches the feature level according to the scene complexity (such as personnel density and occlusion frequency): shallow features are reused in simple scenes (such as empty warehouses) to save computing power, while medium-level features are called in complex scenes (such as cross-operations in workshops) to retain more details. Finally, the optimized features are input into the target tracking model, which reduces redundant calculations and improves feature targeting.

[0166] The target tracking model integrates motion trajectory vectors (reflecting target speed and direction), appearance feature vectors (derived from reused features), and behavioral feature vectors (such as personnel operation postures and vehicle driving patterns) to construct a 3D matching space. The deep reinforcement learning model uses inter-frame ID matching accuracy as a reward signal, combined with real-time analysis results from a spatiotemporal attention network (target motion state, appearance confidence, and scene occlusion rate), to dynamically adjust the weight ratio of these three factors: for example, automatically increasing the weight of appearance features when the occlusion rate is >50%, and increasing the weight of motion trajectory when the motion state is unstable (such as sudden stops or changes of direction). If the deviation between the predicted motion position and the appearance matching position exceeds a certain limit, a conflict detection mechanism triggers local feature re-extraction (such as recalculating local detail features of the target) to correct the matching result.

[0167] The model constructs a target feature memory to store typical appearance features of the target in a stable state (such as full-body features when there is no occlusion). When the target reappears after a brief occlusion, the original ID is quickly associated by comparing features from the memory. At the same time, ID-invariant features (such as the height ratio of people and the wheelbase of vehicles) are mined through self-supervised learning. These features are not affected by temporary changes (such as changing clothes or partial occlusion) and maintain ID continuity when conventional feature matching fails (such as long-term occlusion), ensuring the robustness of cross-frame tracking.

[0168] In one possible implementation, the action semantic parsing includes:

[0169] Action templates are generated based on a security rule knowledge graph, and graph nodes are dynamically expanded based on the results of manual review.

[0170] A spatiotemporal self-attention network is used to weight and focus on consecutive frames, and features of frames with weights exceeding a set threshold are extracted for template matching.

[0171] In one possible implementation, the action semantic parsing further includes:

[0172] Input the continuous trajectory of the same target into a spatiotemporal convolutional network to extract spatiotemporal features of the action;

[0173] The action deviation value is calculated based on the similarity between the feature vector and the safety template.

[0174] The action semantic parsing also includes:

[0175] A safety rule knowledge graph is established, where graph nodes represent standard action types and edges between nodes represent action association constraints. The safety rule knowledge graph is associated with preset early warning rules. A spatiotemporal self-attention network is used to perform weighted processing on continuous video frames, focusing on key action frames and extracting frame features whose weights exceed a set threshold.

[0176] The continuous trajectories of the same target are input into a spatiotemporal convolutional network, and the frame features and trajectory dynamic information are fused to generate a spatiotemporal feature vector.

[0177] Based on the action analysis template associated with the preset early warning rules in the safety rule knowledge graph, the similarity between the spatiotemporal feature vector and the template is calculated, and the action deviation value is output.

[0178] When reviewing and marking new types of violations, the system automatically expands the graph nodes, generates corresponding action analysis templates, and synchronously updates the associated preset warning rules.

[0179] In one possible implementation, the preset warning rule is visualized through the nodes and associated constraints of the security rule knowledge graph. When the deviation value output by the action semantic parsing exceeds the dynamic threshold, the corresponding warning is triggered according to the warning rule associated with the action type node in the graph.

[0180] In one possible implementation, the spatiotemporal self-attention network performs weighted processing on consecutive video frames, focusing on key moments of action, and filters redundant background frames to reduce invalid feature extraction.

[0181] In one possible implementation, the spatiotemporal feature vector generated by the spatiotemporal convolutional network contains temporal logical information of the action, which is determined based on the dynamic change law of the target's continuous trajectory.

[0182] The working principle and effects of the above technical solution are as follows:

[0183] The action semantic parsing process (based on real-time reasoning results from multiple consecutive frames, generates action semantic tags through a temporal behavior analysis model, and achieves accurate understanding and risk quantification of target actions in image / video streams through a closed-loop mechanism of "knowledge graph-driven - spatiotemporal feature extraction - dynamic early warning linkage") is as follows:

[0184] A safety rule knowledge graph is constructed as the core logical carrier: graph nodes define standard action types (such as "operating equipment correctly" and "entering a dangerous area without authorization"), and edges between nodes describe action association constraints (such as the temporal dependency of "safety belts must be worn before working at heights"). The graph is directly associated with preset warning rules (each action type node corresponds to a specific warning threshold and response strategy). When a new type of violation is identified (such as "unauthorized removal of protective devices" not found in the initial graph), a new graph node is automatically added, a matching action analysis template is generated, and the associated preset warning rules are updated synchronously to ensure that the template library and rule library dynamically evolve with actual scenarios.

[0185] For continuous video frames, the spatiotemporal self-attention network learns the temporal importance of actions (e.g., the weight of "the moment of contact with high-voltage equipment" is higher than that of "still frames without action"), automatically focuses on key action frames, and extracts only the frame features whose weight exceeds a set threshold (filtering redundant background frames and reducing invalid computation). At the same time, the continuous motion trajectory of the same target is input into the spatiotemporal convolutional network, and the static features of key frames (e.g., limb posture) and the dynamic information of the trajectory (e.g., changes in movement speed and direction) are fused to generate a spatiotemporal feature vector containing the temporal logic of the action (e.g., the order of "climbing first and then crossing"), thus fully characterizing the spatiotemporal attributes of the action.

[0186] Based on the action analysis template of the associated early warning rules in the knowledge graph, the similarity between the spatiotemporal feature vector and the template is calculated, and the action deviation value is output (the higher the deviation value, the greater the deviation of the action from the safety standard). When the deviation value exceeds the dynamic threshold, the system calls the preset early warning rule associated with the action type node, triggering the corresponding level of early warning (such as audible and visual alarm, order dispatch), realizing the seamless connection of "action analysis - risk quantification - early warning response".

[0187] Spatiotemporal self-attention networks focus on keyframes to improve the feature signal-to-noise ratio. Spatiotemporal convolutional networks, after fusing trajectory information, improve the accuracy of action sequence logic recognition, solving the problem that traditional single-frame analysis cannot capture action continuity (such as misjudging "incomplete standard operations" as violations). The dynamic expansion mechanism of the knowledge graph allows the system to quickly adapt to new types of violations (expansion response time is reduced from days to hours in traditional manual configuration), making it particularly suitable for industrial environments with rapid process iterations and complex operating scenarios.

[0188] The keyframe filtering mechanism reduces redundant frame feature calculations, and the reuse of trajectory information by the spatiotemporal convolutional network further reduces the computing power requirement, while maintaining real-time performance when supporting parallel analysis of multiple video streams.

[0189] The design that links safety rule knowledge graphs with early warning rules makes early warning logic more concrete (e.g., associating the "high-altitude operation" node with "deviation value > 0.6 triggers an emergency early warning"), avoiding the rigidity of traditional hard-coded rules. Dynamic thresholds combined with the quantified output of action deviation values ​​can accurately trigger early warnings based on differences in action type (e.g., higher thresholds for high-risk actions), reducing false alarms for low-risk actions and missed alarms for high-risk actions, thus improving early warning accuracy.

[0190] In one possible implementation, the analysis of the preprocessed data using the trained AI model further includes:

[0191] A multimodal feature fusion network is constructed to extract visual features, infrared thermal imaging features, and motion trajectory features of images. When a single feature is disturbed by the environment, other modal features are automatically weighted and fused for compensation.

[0192] The working principle and effects of the above technical solution are as follows:

[0193] The network constructs multiple feature extraction branches for the preprocessed data (image / video streams, infrared thermal imaging data, target trajectory sequences):

[0194] Visual feature branch: Extracts appearance features of images (such as target shape, color, and texture) through convolutional neural networks (CNN), focusing on detailed information under visible light;

[0195] Infrared thermal imaging feature branch: Utilizes a lightweight infrared feature extractor to capture the temperature distribution features of the target (such as overheated areas of equipment and thermal outlines of the human body), adapting to visual failure scenarios such as low light and smoke.

[0196] Motion trajectory feature branch: The continuous motion trajectory of the target is encoded by a recurrent neural network (RNN) or Transformer, and dynamic features such as velocity, acceleration, and direction change are extracted to supplement the temporal logic of the action.

[0197] By calculating feature entropy values ​​(the lower the entropy value, the more stable the feature) and consistency with historical features (e.g., consistency decreases when visual features suddenly become blurred), the degree of environmental interference on a single modality is quantified (e.g., strong light causes a sharp increase in visual feature entropy, and heavy fog causes a decrease in the signal-to-noise ratio of infrared features). Based on the evaluation results, the network dynamically adjusts the weights of each modality through a self-attention mechanism: when a certain modality is disturbed (e.g., visual features fail), the weight ratio of other modalities is automatically increased (e.g., the weight of infrared features increases from 0.3 to 0.6, and the weight of motion trajectory features increases from 0.2 to 0.4), achieving complementary compensation between features.

[0198] The fused multimodal feature vector integrates the advantages of visual, infrared, and motion information, and is directly input into the subsequent action semantic parsing module (such as a spatiotemporal convolutional network), providing a more comprehensive input for the generation of spatiotemporal feature vectors and enhancing the accuracy of action temporal logic and risk quantification.

[0199] The complementary mechanism of multimodal features enables the integrity of features to be maintained through other modes even when a single mode is disturbed (such as visual failure at night or infrared saturation in high temperature environments), thereby improving anti-interference ability and ensuring stable resolution in complex environments.

[0200] In one possible implementation, comparing the identified preset target and its action characteristics with preset warning rules includes:

[0201] An adaptive template evolution mechanism is adopted to periodically analyze historical misjudgment cases, automatically identify reasonable action variation types not covered in the safety action templates, and generate new template sub-libraries through reinforcement learning and merge them with the original template library;

[0202] A dual-loop threshold optimization algorithm is adopted. The inner loop dynamically fine-tunes the deviation threshold based on the real-time identified action deviation distribution, while the outer loop periodically (e.g., quarterly) calibrates the threshold adjustment strategy by combining safety accident statistics and expert evaluation results.

[0203] In one possible implementation, the comparison of action features with preset warning rules further includes an adaptive template evolution mechanism, specifically:

[0204] A two-tier template library architecture is constructed. The basic template library stores standard safety action characteristics that conform to job specifications, while the variant template library stores verified reasonable individual action differences.

[0205] When the deviation between the action feature and the base template is in the first interval, such as the interval [0.7σ, 1.5σ] (σ is the standard deviation of the base template feature distribution) in multiple consecutive comparisons (e.g., 5), the action sample is automatically input into the Generative Adversarial Network (GAN) to generate a variant template adapted to individual differences.

[0206] The Wasserstein distance metric is used to measure the feature similarity between the new variant template and the existing template library. When the distance similarity reaches a threshold (e.g., 0.3), the new template is added to the variant template library and the version number is updated. At the same time, it is associated with the adaptation range of the corresponding base template.

[0207] The working principle and effects of the above technical solution are as follows:

[0208] The basic template library, serving as the cornerstone of standardized actions, stores action characteristics that strictly conform to job safety standards, providing a fundamental reference for action comparison. The variation template library, on the other hand, focuses on action variations arising from reasonable individual differences in actual work, serving as a supplement and expansion of the basic template library.

[0209] The system continuously monitors the deviation between the action features and the basic template. When the deviation falls within a specific range (e.g., [0.7σ, 1.5σ], where σ is the standard deviation of the basic template feature distribution) in multiple consecutive comparisons (e.g., 5 times), it indicates that the action has a reasonable possibility of variation, and the action sample is automatically extracted. The setting of this range is not arbitrary, but based on the statistical analysis of a large amount of historical action data, aiming to accurately capture those action variations that deviate from the standard but are within a reasonable range.

[0210] The extracted action samples are input into a Generative Adversarial Network (GAN), which leverages the powerful generative capabilities of GANs to generate variant templates adapted to individual differences. Subsequently, the feature similarity between the new variant template and the existing template library is measured using the Wasserstein distance. When the similarity meets a set threshold (e.g., 0.3), the new template is included in the variant template library and simultaneously associated with the corresponding base template, forming a complete template evolution loop.

[0211] The inner closed loop dynamically adjusts the deviation threshold based on the real-time distribution of identified motion deviations. For example, if the deviation values ​​for similar actions are generally high within a certain period, the inner closed loop will appropriately loosen the threshold to accommodate short-term fluctuations in the environment or operating habits; conversely, if the overall deviation values ​​are low, the inner closed loop will tighten the threshold to improve the accuracy of the warning. This adjustment is not done overnight, but rather gradually in small steps based on the changing trend of the real-time deviation distribution, ensuring the smoothness and accuracy of the threshold adjustment.

[0212] The outer closed loop takes a macro-level approach, combining safety accident statistics and expert assessments to calibrate the threshold adjustment strategy quarterly. Safety accident data directly reflects the actual safety situation under the current threshold settings, while expert assessments incorporate specialized knowledge and experience. For example, if a certain type of accident occurs frequently, and expert assessments reveal that the threshold settings are too broad, leading to some potential hazards not being detected in time, the outer closed loop will adjust the threshold adjustment logic of the inner closed loop. This could involve changing the mapping relationship between deviation distribution and threshold adjustment magnitude, making the inner closed loop more precise in subsequent dynamic fine-tuning.

[0213] In one possible implementation, the inner closed loop dynamically fine-tunes the deviation threshold based on the real-time identified action deviation distribution, including:

[0214] The density clustering algorithm is used to group the actions in the enterprise's production operations. The density clustering algorithm combines the semantic correlation of actions in the safety rule knowledge graph to assign semantic weights, and groups actions with similar action features and the same risk level into the same action group.

[0215] For each action group, a three-dimensional threshold matrix is ​​constructed, comprising spatial bias threshold, temporal bias threshold, and feature bias threshold, where:

[0216] The spatial deviation threshold is the distance deviation threshold between the moving target and the danger zone;

[0217] The timing deviation threshold is the deviation threshold of the action completion time;

[0218] The feature deviation threshold is the deviation threshold of the angle or trajectory curvature of the moving limb;

[0219] The inner closed loop dynamically fine-tunes the thresholds in the three-dimensional threshold matrix based on the real-time action deviation distribution of each action group. The real-time action deviation distribution is modeled by kernel density estimation, and the fine-tuning range is determined based on the mean, standard deviation and 95% confidence interval of the distribution. The threshold adjustment range of high-risk action groups is greater than that of low-risk action groups.

[0220] When the inner closed loop dynamically fine-tunes the thresholds in the three-dimensional threshold matrix, it also includes:

[0221] When two strongly correlated actions are detected to occur simultaneously, the thresholds of the two action groups are simultaneously fine-tuned based on the action correlation correction factor; the action correlation correction factor is determined according to the correlation edge weights between action nodes in the security rule knowledge graph.

[0222] When multiple concurrent actions cause threshold adjustment conflicts, the threshold adjustment priority is determined by a risk priority arbitration algorithm. The risk priority arbitration algorithm calculates the action risk product based on the action's risk level, probability of occurrence, and scope of impact, and prioritizes satisfying the threshold adjustment needs of action groups with higher action risk products.

[0223] The working principle and effects of the above technical solution are as follows:

[0224] The mechanism for dynamically fine-tuning the deviation threshold in the inner closed loop is based on a three-layer logic of "action clustering and grouping - multi-dimensional threshold modeling - real-time feedback adjustment," enabling precise control of multiple action deviations in complex production scenarios.

[0225] Based on the density-based clustering algorithm (DBSCAN), this approach assigns semantic weights to actions using a safety rule knowledge graph, overcoming the limitations of traditional clustering that relies solely on feature similarity. For example, although the actions of "welding without a protective mask" and "cutting without a protective mask" differ in their specific characteristics, they are grouped into the same action group because of their high semantic correlation in the lack of protection and their shared high-risk status. During clustering, semantic weights are quantified through the association edge attributes of action nodes in the knowledge graph (e.g., a weight coefficient of 1.2 is assigned when the association strength is ≥0.8), ensuring that the grouping results simultaneously satisfy feature similarity and risk homogeneity.

[0226] For each set of actions, a three-dimensional threshold matrix is ​​constructed, encompassing spatial, temporal, and feature dimensions, to achieve multi-dimensional constraints on action deviations.

[0227] Spatial dimension: Calculate the Euclidean distance deviation between the moving target and the dangerous area (such as high-voltage equipment, chemical storage area) based on the coordinate data of target detection. When the deviation exceeds the preset threshold (such as ≤30cm for high-risk group), an early warning is triggered.

[0228] Temporal dimension: Based on the timestamp sequence of video frames, the deviation between the actual completion time of the action and the standard time is statistically analyzed (e.g., hoisting operation timeout deviation > 5 min);

[0229] Feature dimensions: Extract limb angles (e.g., body tilt angle > 60° during high-altitude operations) or trajectory curvature (e.g., abnormal bending of the robotic arm's motion trajectory) through skeletal key point detection, and quantify feature bias.

[0230] The three-dimensional thresholds are independent of each other but work together; exceeding the limit in any dimension is considered an over-deviation.

[0231] The inner loop uses a 5-minute sliding window period and employs kernel density estimation (KDE) to model the distribution of real-time acquired motion bias values, extracting the mean, standard deviation, and 95% confidence interval ([L,U]) as adjustment criteria.

[0232] When the upper limit of the 95% confidence interval U is greater than the current threshold × 1.2, the threshold is adjusted upward (e.g., from 30cm to 35cm) to avoid false alarms caused by environmental interference (e.g., detection deviation caused by sudden changes in light intensity).

[0233] When the lower limit of the confidence interval L is less than the current threshold × 0.8, the threshold is lowered (e.g., from 30cm to 25cm) to improve sensitivity to minor risks. For strongly correlated actions (e.g., disassembling valves and pressurizing pipelines), the correlation correction factor is calculated through the weight of the knowledge graph's related edges (e.g., a weight of 0.9 corresponds to a factor of 1.1), and the two sets of thresholds are adjusted simultaneously to avoid the risk superposition of related actions.

[0234] When multiple actions conflict concurrently, arbitration priority is determined based on the product formula of risk level × probability of occurrence × scope of impact. For example, when hot work (risk level 5) conflicts with tool placement violation (risk level 2), the threshold requirement of the hot work group is guaranteed first.

[0235] By grouping actions, thresholds are avoided for each individual action, reducing the number of threshold parameters. Simultaneously, a 3D matrix covers the multi-dimensional features of actions, improving the accuracy of deviation detection and solving the problem that traditional single thresholds cannot fully describe the risks of complex actions. Real-time distributed modeling and dynamic fine-tuning mechanisms can quickly respond to environmental changes (such as automatically relaxing space thresholds during peak personnel hours) and equipment status fluctuations (such as dynamically adjusting the trajectory deviation tolerance of aging robotic arms), reducing false alarm rates. This is particularly suitable for complex production scenarios with multiple alternating processes. Synchronous adjustment and priority arbitration mechanisms for strongly correlated actions resolve threshold conflicts when multiple actions occur concurrently, ensuring that high-risk actions (such as operations in confined spaces) are always under strict threshold control.

[0236] In one possible implementation, the method for fine-tuning the deviation threshold includes:

[0237] When a preset number of consecutive frames (e.g., 3 frames) are detected to show a preset amplitude (e.g., ±0.1) fluctuation in the entropy value of a certain modality, the reliability assessment threshold of that modality will be automatically relaxed by a preset ratio.

[0238] The modal reliability score is used to make graded adjustments. When there is mild interference, the weights of other modalities are increased by the third proportion (e.g., 10%-20%). When there is severe interference, the weights of other modalities are increased by the fourth proportion (30%-50%). The increase is dynamically adjusted each frame based on the real-time score, and the fourth proportion is greater than the third proportion.

[0239] When the consistency score of the multimodal fusion feature reaches the first preset score (e.g., 0.8), the deviation threshold is executed according to the base value; if the consistency score is lower than the second preset score (e.g., 0.6), the threshold is tightened according to the fifth ratio, where the first preset score is higher than the second preset score.

[0240] The threshold fine-tuning is driven by real-time feedback. After each frame, the quality index of each modality feature and the stability of the fused feature are calculated to trigger the threshold adjustment. The adjustment delay does not exceed a preset time (e.g., 10ms).

[0241] The threshold adjustment effect is analyzed offline according to a preset period (e.g., 24 hours), and the fine-tuning rules are optimized a second time.

[0242] The working principle and effects of the above technical solution are as follows:

[0243] When a modality (such as the visual modality) experiences fluctuations in its feature entropy value due to sudden changes in illumination, a dual judgment mechanism using a preset number of frames (3 frames) and an amplitude (±0.1) automatically relaxes the reliability assessment threshold for that modality, preventing instantaneous interference from a single modality from causing misadjustment of the overall threshold. For example, when a dusty environment causes abnormal fluctuations in the visual feature entropy value, the threshold judgment standard for the visual modality is temporarily relaxed to reduce false warnings caused by image blurring.

[0244] Through a modal reliability grading adjustment mechanism, the weights of other modes are dynamically increased when a single mode is disturbed (10%-20% increase for mild disturbances, 30%-50% increase for severe disturbances), achieving adaptive fusion of "weak mode weakening and strong mode strengthening". For example, when smoke obstruction causes the infrared mode to fail, the weights of the sound mode (such as abnormal equipment noise) and vibration mode are automatically increased to ensure that the fused features can still accurately reflect the action deviation, so that the accuracy of multimodal collaborative evaluation remains above 85% in interference scenarios.

[0245] The consistency score threshold is dynamically adjusted based on multimodal fusion features: when consistency is high (≥0.8), the basic threshold is applied to ensure evaluation accuracy in standard scenarios; when consistency is low (<0.6), the threshold is tightened (e.g., the fifth ratio of 5%-10%) to avoid the risk of fusion feature distortion by increasing the severity of judgment. This mechanism strongly binds threshold adjustment to feature quality, reducing false negatives caused by fusion bias.

[0246] Feature quality calculation and threshold adjustment are completed within 10ms after each frame ends to meet the real-time requirements of high-dynamic scenarios (such as high-speed assembly line operations); at the same time, the adjustment effect is evaluated through offline analysis on a 24-hour cycle, and the fine-tuning rules (such as feature entropy fluctuation threshold and weight increase ratio) are optimized a second time, so that the threshold adjustment strategy can continuously adapt to scene changes (such as ambient light differences caused by seasonal changes) in long-term operation. The overall early warning accuracy can be improved by 15%-20% after rule iteration.

[0247] Through a two-layer mechanism of "real-time feedback-driven fine-tuning + periodic offline optimization rules", the system can ensure rapid response at the single-frame level (such as instantaneous threshold adjustment during sudden occlusion) and smooth rule oscillations caused by short-term fluctuations through offline analysis (such as eliminating interference from occasional extreme data on fine-tuning parameters). This balance enables the system to capture subtle risks in a timely manner in high-frequency switching multi-process scenarios, while avoiding system instability caused by frequent threshold fluctuations.

[0248] In one possible implementation, the dynamic deviation threshold setting is determined by the following methods:

[0249] Establish a basic threshold library, which includes minimum safe distance and maximum permissible speed parameters for different target categories;

[0250] An environmental correction factor is introduced, which is dynamically adjusted based on parameters such as light intensity, visibility (meters), and population density.

[0251] The final dynamic deviation threshold is determined by using the base threshold, environmental correction factor, and risk level coefficient.

[0252] For example: Final dynamic deviation threshold = base threshold × environmental correction factor × (1 + risk level coefficient).

[0253] The environmental correction factor is obtained through the following steps:

[0254] The light intensity, visibility, and population density were normalized to obtain the normalized light intensity L. norm Normalized visibility V norm， and the normalized population density D norm ;

[0255]

[0256] Where L is the current light intensity, L max This represents the maximum light intensity.

[0257]

[0258] V represents the actual visibility, V1 represents the lowest effective visibility value, and V2 represents the highest effective visibility value; for example, if the effective visibility range is [50, 1000], then V1 = 50, V2 = 1000;

[0259]

[0260]

[0261] D represents the actual population density. y The threshold for population density; min() is the minimum value; α and β are coefficients, 0 < α < 1, 1 < β ≤ 3; K d Personnel density factor;

[0262] The calculation of the environmental correction factor includes the determination of the light-visibility coupling factor Klv, as follows:

[0263] When both the normalized values ​​of light intensity and visibility are greater than 0, the light-visibility coupling factor is:

[0264] Klv = 1 + L norm ×V norm

[0265] To demonstrate the synergistic interference of low light and low visibility;

[0266] When the normalized value of light intensity or the normalized value of visibility is 0, the light-visibility coupling factor takes the base value, i.e., Klv = 1. At this time, there is no need to use the product method to reflect the state of coordinated interference elimination when there is no interference from a single environmental factor.

[0267] Environmental Correction Factor K env for:

[0268] K env =w1×Klv+w2×K d

[0269] w1 and w2 are weights.

[0270] By normalizing and co-calculating illumination, visibility, and personnel density, the environmental correction factor accurately reflects the combined impact of different environmental conditions. For example, in nighttime foggy scenarios (low illumination, low visibility), the klv value increases due to co-interference. Combining personnel density and weights, the environmental correction factor is adapted to the harsh environment, providing accurate environmental compensation for the dynamic deviation threshold and improving its rationality. Distinguishing between single environmental factors without interference and those with co-interference, when illumination or visibility is good (normalized value 0), Klv is set to 1 to avoid over-amplifying environmental influences and causing misjudgments. When co-interference exists, it is appropriately amplified, ensuring sensitive identification of hazardous environments while reducing false alarms in normal environments, balancing safety and practicality. The formula includes adjustable parameters such as α, β, w1, and w2, which can be optimized based on the actual needs and historical data of different application scenarios (such as industrial plants, traffic roads, etc.) to flexibly adapt to various scenarios and improve the algorithm's applicability and effectiveness in different environments.

[0271] In one possible implementation, the method further includes:

[0272] The system stores and reviews data related to early warning events, uses the review results to optimize models and rules, analyzes safety risk trends, and generates improvement suggestions.

[0273] In one possible implementation, the storage and review of data related to early warning events, the feedback of review results to optimize models and rules, the analysis of security risk trends, and the generation of improvement suggestions include:

[0274] Video clips related to the early warning event are extracted, stored in a standard format, and the structured metadata of the early warning event is recorded in the database.

[0275] The database is updated based on the occurrence time of each warning event, the correlation between each warning event, the frequency of occurrence of each warning event, and the severity of the warning event.

[0276] By analyzing security risk trends through data mining, security improvement suggestions are generated and fed back to the management department.

[0277] The warning events are reviewed through the review and playback interface to obtain the review results; the AI ​​model and dynamic deviation threshold are iteratively optimized based on the review results (manual confirmation / false alarm marking); (the warning events are status determined and false alarms are marked through the warning event review and playback function).

[0278] The database is updated based on the occurrence time of each warning event, the correlation between each warning event, the frequency of occurrence of each warning event, and the severity of the warning event; including:

[0279] Assign a time decay coefficient to early warning events to ensure that recent high-risk events have a higher proportion of impact on the database;

[0280] Construct an association map of early warning events using a graph neural network (GNN);

[0281] The nodes include event type nodes, environment status nodes, and region location nodes;

[0282] The edge weights between event type nodes are the co-occurrence frequencies, and the edge weights between event type nodes and environment state nodes are the environment risk coefficients.

[0283] In response to any two event type nodes having an edge weight greater than or equal to the weight threshold K, and their combined risk probability with the environment state node being greater than or equal to the probability threshold P, a combined risk node is generated and the joint early warning rule is updated.

[0284] Differentiated storage is performed based on the monthly event frequency:

[0285] For events occurring ≥ 3 times, store the original video clips and correlate them with the sensor data stream, while simultaneously initiating root cause analysis;

[0286] For events with a frequency of less than 1 time, compress the keyframe feature vector and event metadata to less than 30% of the original size.

[0287] The calculation method for the composite risk probability is as follows:

[0288] Composite risk probability = Event co-occurrence frequency / Baseline period × max (Environmental risk coefficient)

[0289] Among them, the environmental risk coefficient is determined based on the environmental correction factor and the equipment status parameters; the equipment status parameters include vibration, temperature, operating time and fault history weights.

[0290] Environmental risk coefficient = Environmental correction factor × (1 + a1 × weighted average of normalized values ​​of equipment state parameters), where the environmental correction factor is calculated based on illumination, visibility, and personnel density, and 0. <a1<0.5。

[0291] The historical weights of faults are determined based on the number of equipment failures in the past, the severity of the failures, and the correlation between the failures and the current events.

[0292] `max(environmental risk coefficient)` refers to the maximum value of the environmental risk coefficient among all environmental state nodes that co-occur with the associated event. For example, if two events co-occur 3 times within 1 hour, the environmental risk coefficients at the corresponding 3 times are: `max(environmental risk coefficient)` takes the maximum value from the above sequence, representing the worst combination of environmental equipment state in the scenario where the two events co-occur.

[0293] The working principle and effects of the above technical solution are as follows:

[0294] Upon receiving the warning event, the server immediately commands the corresponding FFmpeg instance to trace back a certain period of time from the current time and forward a certain period of time to extract the key video segment containing the event from the original RTSP stream.

[0295] Segment format: Saved as standard MP4 (H.264 / AAC) format to ensure broad compatibility.

[0296] Segment storage: Extracted MP4 video segments are uniformly stored in a designated directory on the central server. The system does not store complete, uninterrupted surveillance recordings, but only segments associated with warning events.

[0297] MySQL serves as a high-performance cache and message queue. Redis caches frequently accessed data, reducing the pressure on MySQL.

[0298] Message queues (streams / pubsub): Used to decouple the AI ​​engine from subsequent processing modules; the AI ​​engine publishes alert event information to a Redis queue, which is consumed by a dedicated event processing service responsible for segment extraction, MySQL recording, and notification sending. This improves system response speed and reliability.

[0299] The management backend queries the list of alert events from MySQL and displays information such as time, camera, type, status, and snapshot.

[0300] When a user clicks to view the warning report, the backend reads the corresponding MP4 file directly from the server storage directory based on the recorded videopath and plays it back through the front-end video player.

[0301] The web-based management backend handles camera management, alert rule configuration, user permission management, event status updates (confirmation / false alarm), and system monitoring. The backend primarily interacts with MySQL.

[0302] By selecting the maximum value of the environmental risk coefficient within the co-occurrence period of related events, the extreme combination scenario of "the worst environment + the worst equipment condition" can be accurately identified. For example, if only one of the five co-occurrences of two events occurs under the condition of "dense fog + high temperature and vibration of equipment", taking the maximum value can avoid the average calculation of the other four low-risk scenarios masking this extreme risk, making the composite risk probability more reflective of the "risk superposition effect under the most unfavorable conditions", and significantly reducing the missed detection rate of high-risk scenarios.

[0303] The value of max (environmental risk coefficient) forms a complete closed loop with the composition of the environmental risk coefficient (environmental correction factor + equipment status parameters): both the environmental correction factor (light, visibility, etc.) and the equipment status parameters (vibration, fault history, etc.) change dynamically with the co-occurrence time, and the extraction of the maximum value directly corresponds to the peak value of the dual risks of "environment-equipment"; combined with the quantification of "fault correlation" in the fault history weight (such as the matching degree between the current event and historical serious faults), the physical meaning of the maximum value is made clearer (such as "the high risk at this moment is not only due to the poor environment, but also highly correlated with the historical faults of the equipment"), which makes it easier for engineers to trace the root cause of the risk.

[0304] In a graph neural network (GNN) correlation graph, the edge weights of environmental state nodes represent the environmental risk coefficients. The operation of maximizing the weights essentially extracts the "most critical risk features" from massive co-occurring data, providing a precise quantitative basis for the generation of composite risk nodes.

[0305] When the co-occurrence frequency of two events is greater than or equal to the threshold K, if max (environmental risk coefficient) is also greater than or equal to a certain implicit critical value, the probability of composite risk is more likely to exceed the threshold P, ensuring that the generated composite risk nodes (such as "personnel crossing the boundary + equipment malfunction | extreme environment") truly represent high-risk associated scenarios, and reducing the redundant occupation of the database by invalid composite nodes (such as the merging of low-risk co-occurrence events).

[0306] For high-frequency co-occurrence events (monthly frequency ≥ 3 times), the extreme scene video clips corresponding to max (environmental risk coefficient) will be automatically marked as "key risk material", and the original resolution and complete time frame will be retained first to provide high-quality data for root cause analysis (such as locating "whether the equipment malfunction is caused by high temperature" by playing back extreme scene videos).

[0307] For low-frequency events, the extraction of the maximum value can compress invalid data storage (only the key frame corresponding to the maximum value needs to be retained). Combined with the strategy of "compressing to less than 30% of the original size", storage costs can be reduced by about 40% while ensuring that risk features are not lost.

[0308] In industrial scenarios, the environment and equipment status are highly dynamic (such as sudden rainstorms or instantaneous equipment overload). The rules for determining the value of max (environmental risk coefficient) can quickly respond to such changes: even if two events coexist in a low-risk state for a long time, once a single extreme scenario occurs, the system can update the composite risk probability in a timely manner by changing the maximum value, triggering the dynamic adjustment of the joint early warning rules (such as temporarily tightening the safety threshold), thereby improving the robustness of the system in complex dynamic environments by more than 20%.

[0309] In summary, this value definition, by focusing on extreme risks, strengthening logical closed loops, and collaborating with graph analysis, not only improves the accuracy of risk assessment but also optimizes resource allocation efficiency, providing core support for the intelligence and reliability of database updates.

[0310] See attached document Figure 2 This application also provides an AI-based enterprise security management system for the AI-based enterprise security management method described in this application. The system includes:

[0311] The configuration module is used to configure the storage path and permissions of the data storage unit and initialize the operation parameters of the computing unit.

[0312] The acquisition module is used to acquire images and / or video streams from key areas of enterprise security management in real time; and to preprocess the real-time data.

[0313] The identification and analysis module is used to analyze pre-processed images and / or video streams using a trained AI model to identify preset targets and their motion characteristics in enterprise production operations.

[0314] The early warning module is used to compare the identified preset targets and their action characteristics with preset early warning rules. When the deviation exceeds the dynamic deviation threshold, it is judged as an erroneous action, an early warning event is generated, and an alarm mechanism is triggered.

[0315] The storage module is used to store and review data related to early warning events, and to optimize models and rules based on the review results, analyze security risk trends, and generate improvement suggestions.

[0316] The working principle and effect of the above technical solution are the same as those of the aforementioned AI-based enterprise security management method, and will not be repeated here.

[0317] This application also provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described in this application.

[0318] This application describes the invention from the perspectives of purpose, effectiveness, progress, and novelty, and it meets the functional enhancement and use requirements emphasized by the Patent Law. The above description and accompanying drawings are merely preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this application, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, shall fall within the scope of protection of this patent application.

Claims

1. An AI-based enterprise security management method, characterized in that, The method includes: Configure the storage path and permissions of the data storage unit, and initialize the operation parameters of the computing unit; Real-time acquisition of images and / or video streams from key areas of enterprise security management; and preprocessing of real-time data; The trained AI model is used to analyze preprocessed images and / or video streams to identify preset targets and their motion characteristics in enterprise production operations. The identified preset targets and their action characteristics are compared with preset early warning rules. When the deviation exceeds the dynamic deviation threshold, it is judged as an erroneous action, an early warning event is generated, and an alarm mechanism is triggered. Data related to early warning events is stored and reviewed, and the review results are used to optimize models and rules.

2. The enterprise safety management method according to claim 1, characterized in that, The analysis of preprocessed images and / or video streams using the trained AI model includes: The GPU-accelerated AI inference framework performs target detection and classification on a single frame of image or video, and outputs structured data of the target, including coordinates, category labels and confidence scores. Based on the real-time reasoning results of multiple consecutive frames, action semantic tags are generated through a temporal behavior analysis model; The model accuracy and GPU computing power allocation are dynamically adjusted based on the target risk level.

3. The enterprise safety management method according to claim 2, characterized in that, The GPU-accelerated AI inference framework performs object detection and classification on a single frame of image or video, outputting structured data of the object, including: The video frames are processed by a TensorRT-accelerated object detection model, and the object detection results are output. A target tracking model is used to assign a unique ID to the detected target and associate it with cross-frame trajectories.

4. The enterprise safety management method according to claim 3, characterized in that, The GPU-accelerated AI inference framework performs object detection and classification on a single frame of image or video, outputting structured data of the object, including: The shallow feature maps of the object detection model are reused in the object tracking model; The target tracking model integrates the target's motion trajectory vector and appearance feature vector, and maintains the continuity of the target ID through weighted similarity calculation; the appearance feature vector originates from the shallow feature map.

5. The enterprise safety management method according to claim 2, characterized in that, The real-time reasoning results based on multiple consecutive frames generate action semantic tags through a temporal behavior analysis model; including: A safety rule knowledge graph is established, where graph nodes represent standard action types and edges between nodes represent action association constraints. The safety rule knowledge graph is associated with preset early warning rules. A spatiotemporal self-attention network is used to perform weighted processing on continuous video frames, focusing on key action frames and extracting frame features whose weights exceed a set threshold. The continuous trajectories of the same target are input into a spatiotemporal convolutional network, and the frame features and trajectory dynamic information are fused to generate a spatiotemporal feature vector. Based on the action analysis template associated with the preset early warning rules in the safety rule knowledge graph, the similarity between the spatiotemporal feature vector and the template is calculated, and the action deviation value is output.

6. The enterprise safety management method according to claim 2, characterized in that, The target risk level is determined in the following ways, including: A preset risk assessment matrix is ​​provided, which includes target category weights, action risk coefficients, and environmental sensitivity. The target category is obtained based on the target detection results, and the corresponding action risk coefficient is queried by combining the action semantic tags. Environmental sensitivity is determined by combining the spatial distance between the target and the hazardous area with the degree of environmental disturbance; The target risk level is determined based on the target category weight, action risk coefficient, and environmental sensitivity.

7. The enterprise safety management method according to claim 1, characterized in that, The step of comparing the identified preset targets and their action characteristics with preset warning rules includes: An adaptive template evolution mechanism is adopted to periodically analyze historical misjudgment cases, automatically identify reasonable action variation types not covered in the safety action templates, and generate new template sub-libraries through reinforcement learning and merge them with the original template library; A dual-loop threshold optimization algorithm is adopted. The inner loop dynamically fine-tunes the deviation threshold based on the real-time identified action deviation distribution, while the outer loop combines safety accident statistics and expert evaluation results to calibrate the threshold adjustment strategy.

8. The enterprise safety management method according to claim 1, characterized in that, The dynamic deviation threshold setting is determined in the following ways: Establish a basic threshold library, which includes minimum safe distance and maximum permissible speed parameters for different target categories; An environmental correction factor is introduced, which is dynamically adjusted based on parameters such as light intensity, visibility, and population density. The final dynamic deviation threshold is determined by using the base threshold, environmental correction factor, and risk level coefficient.

9. The enterprise safety management method according to claim 1, characterized in that, The process of storing and reviewing data related to early warning events, and using the review results to optimize models and rules, includes: Video clips related to the early warning event are extracted, stored in a standard format, and the structured metadata of the early warning event is recorded in the database. The database is updated based on the occurrence time of each warning event, the correlation between each warning event, the frequency of occurrence of each warning event, and the severity of the warning event. The warning events are reviewed through the review and playback interface to obtain the review results; the AI ​​model and dynamic deviation threshold are iteratively optimized based on the review results.

10. An AI-based enterprise security management system, used to implement any one of the AI-based enterprise security management methods as described in claims 1-9, characterized in that, The system includes: The configuration module is used to configure the storage path and permissions of the data storage unit and initialize the operation parameters of the computing unit. The acquisition module is used to acquire images and / or video streams from key areas of enterprise security management in real time; and to preprocess the real-time data. The identification and analysis module is used to analyze pre-processed images and / or video streams using a trained AI model to identify preset targets and their motion characteristics in enterprise production operations. The early warning module is used to compare the identified preset targets and their action characteristics with preset early warning rules. When the deviation exceeds the dynamic deviation threshold, it is judged as an erroneous action, an early warning event is generated, and an alarm mechanism is triggered. The storage module is used to store and review data related to early warning events, and to use the review results to optimize models and rules.

Images