A precast beam yard safety violation identification and linkage early warning method based on AI vision

Abstract

The application relates to the technical field of image detection, and discloses a precast beam field safety violation identification and linkage early warning method based on AI vision. The method comprises five progressive steps of process risk mapping, sensing node scheduling, imaging parameter adaptation, quality feedback optimization and violation linkage early warning, realizes dynamic risk distribution coordinate driving sensing node migration to a high-risk operation surface with production progress as a traction, dynamically adjusts a focus point with the change of the operation surface, realizes self-adaptive switching of an imaging mode and exposure gain through comparison and index of a process type and an environmental parameter, adapts configuration, illumination conditions and background complexity, constructs a closed-loop optimization mechanism through quality evaluation and feedback instructions, records optimal configuration strategies under different process environment combinations to form a knowledge base and realize strategy iteration evolution, and realizes a linkage early warning effect of a sound and light warning and equipment interlocking in the field through accurate matching of a risk level, so that monitoring is upgraded from static fixation to dynamic adaptation, from passive response to active prevention and control.

Description

Technical Field

[0001] This application relates to the field of image detection technology, and in particular to a method for identifying and triggering early warning of safety violations in precast beam yards based on AI vision. Background Technology

[0002] Construction safety management is a core aspect of ensuring project quality and the safety of personnel. With the rapid development of intelligent construction technology, precast beam yards, as key production units in bridge engineering, are gradually evolving towards intelligent and visualized construction safety monitoring. AI-based vision-based safety violation identification and early warning methods, which utilize intelligent cameras and environmental sensors deployed in key areas of the beam yard to conduct real-time behavioral analysis of all processes, including rebar processing, formwork installation, concrete pouring, prestressing tensioning, and beam hoisting, have become an important technical means to improve the safety management level of precast beam yards.

[0003] However, existing AI vision-based safety monitoring systems for precast beam yards mostly employ fixed-point deployment of sensing devices, limiting their monitoring coverage to physical locations such as the rebar processing area and beam fabrication platform area. This prevents them from adaptively adjusting to the periodic dynamic shifts of the beam yard work surface. In actual engineering projects, when the risk in the platform area drops sharply after tensioning operations, while the risk in adjacent beam-moving channels increases dramatically, the fixed array cannot dynamically shift the sensing focus. This results in a structural contradiction: the original monitoring point becomes an "empty mirror," while the new work surface remains in a sensing blind spot. Furthermore, the lighting conditions, occlusion relationships, and background complexity vary significantly across different processes, making it difficult for a single sensing configuration to adaptively switch. This leads to both invalid data uploads and missing valid data, severely restricting the recognition accuracy and timely warning of the AI ​​vision analysis system. Summary of the Invention

[0004] This application proposes a precast beam yard safety violation identification and linkage early warning method based on AI vision. It has the advantages of process-driven perception scheduling, adaptive configuration of imaging parameters, quality closed-loop feedback and hierarchical linkage early warning, and solves the problems of mismatch between fixed array and dynamic work surface coverage, incompatibility between single configuration and heterogeneous environment and single early warning method.

[0005] To achieve the above objectives, this application adopts the following technical solution: a method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, characterized by the following specific steps: S1. Process Risk Mapping: Link the precast beam yard production progress management system to obtain the current beam fabrication process stage and work scope, determine the distribution of high-risk areas, and generate risk coordinates for the work surface. S2, Sensing Node Scheduling: Receive the risk coordinates from step S1, control the sensing nodes to migrate to high-risk work areas, solve for the optimal deployment position and viewing angle, and dynamically adjust the sensing focus as the work area migrates. S3. Imaging parameter adaptation: Based on the process type and environmental characteristics identified in step S1, the imaging parameters of the nodes deployed in step S2 are configured, the imaging mode is switched and the exposure gain is adjusted to adapt the configuration to the process environment. S4. Quality Feedback Optimization: Evaluate the quality of the video collected in steps S2 and S3. If the video does not meet the analysis threshold, feed back to step S2 to adjust the position or step S3 to correct the parameters, and record the optimal configuration strategy. S5. Violation Linkage Early Warning: Receive the video data filtered in step S4, perform target detection and behavior recognition, trigger on-site audio and visual warnings according to the risk level, and associate the warning event with the process information in step S1.

[0006] Furthermore, in step S1, the precast beam yard production progress management system and building information model database are connected to extract the current beam fabrication process stage, platform number, and work space coordinates. The process stage includes rebar binding, formwork installation, concrete pouring, prestressing tensioning, beam moving and hoisting, and beam storage and curing. The process sequence nodes are mapped to the beam yard spatial coordinate system to establish a dynamic correspondence between process status and work area, determine the spatial boundary and geometric topological characteristics of the current active work surface, and provide a spatial benchmark for risk simulation.

[0007] Furthermore, in step S1, the process risk association rule library is called according to the current process stage. The rule library stores the high-risk operation types and spatial range templates corresponding to each process stage, deduces the spatial distribution of high-risk areas, assigns regional perception priority weights according to risk levels, establishes mapping rules between risk levels and resource scheduling, generates risk distribution coordinates of the work surface and outputs them to the scheduling link, so that the deployment of perception resources matches the current work hotspots and eliminates the spatial coverage blind spots of the fixed array.

[0008] Furthermore, in step S2, the risk distribution coordinates output in step S1 are received, and the orbital mobile camera, adjustable gimbal, and drone sensing nodes are driven to migrate to the high-risk work area. The optimal deployment position and viewing angle of the sensing nodes are solved based on the coverage optimization algorithm. The coverage optimization algorithm takes maximizing the field of view coverage as the objective function and the shortest node movement path as the constraint condition, so that the sensing coverage focus is dynamically adjusted as the work area migrates periodically, and the nodes in the original low-risk area go into dormancy or are released.

[0009] Furthermore, in step S2, the spatial geometric features of the current high-risk work surface are used as constraints to establish a matching relationship between the field of view coverage of the sensing nodes and the boundary of the work surface. Coverage redundancy is eliminated in the overlapping areas of multiple nodes. For blind areas, the field of view overlap rate threshold is used to determine and trigger the supplementary node scheduling instruction, so that the sensing resources have no blind spots in space and are continuously connected with the process migration in time.

[0010] Furthermore, in step S3, based on the current process type and environmental characteristics identified in step S1, the imaging parameters of the sensing nodes deployed in step S2 are adaptively configured. The visible light and infrared thermal imaging modes are switched and the exposure gain and polarization filter angle are adjusted to match the frame rate resolution and target motion speed. The visible light mode is used for normal daytime lighting scenarios, and the infrared thermal imaging mode is used for nighttime or low-light scenarios, so that the sensing configuration is adapted to the lighting conditions and background complexity of the current process.

[0011] Furthermore, in step S3, a comparison index between process type and environmental parameters is established. The imaging mode is automatically switched according to the current process lighting conditions and background complexity. Near-infrared supplementary lighting is enabled for low-light night scenes, polarization filtering is enabled for strong arc light interference scenes, and multi-frame fusion exposure is enabled for high dynamic range scenes. The multi-frame fusion exposure suppresses overexposure of highlights and underexposure of shadows by fusing short exposure frames and long exposure frames, so that the imaging quality is adapted to the current process environment.

[0012] Furthermore, in step S4, real-time data quality assessment is performed on the video streams acquired in steps S2 and S3. The assessment indicators include image sharpness, target occlusion rate, and signal-to-noise ratio. Image sharpness is evaluated using the variance of the Laplacian operator, and the target occlusion rate is calculated using the foreground mask area ratio. If the quality does not meet the visual analysis threshold, a feedback instruction is generated and sent back to step S2 to adjust the node position or to step S3 to correct the imaging parameters. The optimal configuration strategy under different process environment combinations is recorded, a perception configuration knowledge base is constructed, and iterative optimization of the perception strategy is realized.

[0013] Furthermore, in step S5, the video data filtered in step S4 is received, and target detection, behavior recognition, and violation assessment are performed. The behavior recognition includes wearing a safety helmet, wearing a reflective vest, entering a dangerous area, standing under a suspended object, and smoking in violation of regulations. Based on the risk level, on-site sound and light warnings and graded linkage warnings pushed by mobile terminals are triggered. Equipment interlock control is implemented in high-risk areas. At the same time, the warning events are associated with the process information in step S1 to generate a process risk association report.

[0014] Furthermore, in step S5, confidence levels are calculated for the identified violations, distinguishing between suspected and confirmed violations. Risk levels are assigned based on the type of violation, the area where it occurred, and the stage of operation. Warning events are then correlated in multiple dimensions according to process stage, team information, and equipment status to generate statistical analysis reports on high-frequency violation scenarios and time periods. These reports include a heatmap of violation frequency, a time period distribution curve, and a team violation ranking, providing data support for optimizing management strategies.

[0015] The beneficial effects of this invention are as follows: This application provides a method for safety violation identification and linkage early warning in precast beam yards based on AI vision. Through five progressive steps, namely, process risk mapping, perception node scheduling, imaging parameter adaptation, quality feedback optimization, and violation linkage early warning, the method achieves the following: 1) Dynamic risk distribution coordinates are constructed based on production progress to drive perception nodes to migrate to high-risk work areas and dynamically adjust the focus according to changes in the work area; 2) Imaging mode and exposure gain are adaptively switched by comparing process type and environmental parameters to adapt the configuration to lighting conditions and background complexity; 3) A closed-loop optimization mechanism is constructed based on quality assessment and feedback instructions to record the optimal configuration strategy under different process environment combinations to form a knowledge base for strategy iteration and evolution; 4) Accurate matching of risk levels triggers on-site audible and visual warnings and equipment interlocking hierarchical linkage early warning effects, thus upgrading monitoring from static and fixed to dynamic adaptation and from passive response to proactive prevention and control. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort: Figure 1 This is a flowchart illustrating the method steps of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In the specific scenario of online monitoring of precast beam yards in operation, existing methods have fundamental limitations. The work surfaces of precast beam yards dynamically shift periodically as the beam fabrication process progresses. High-risk areas at each stage—reinforcement binding, formwork installation, concrete pouring, prestressing tensioning, beam relocation and hoisting, and beam storage and curing—constantly change spatially. Existing systems mostly use fixed-point deployment of cameras and sensors, resulting in a spatial mismatch between the monitored area and the current actual work hotspots. Original monitoring points become empty mirrors, while new work surfaces are in blind spots. Simultaneously, differences in lighting conditions, occlusion relationships, and background complexity across different processes, along with strong moisture interference from nighttime concrete pouring and strong arc light interference from daytime reinforcement welding, prevent single-sensor configurations from adaptively switching, creating a structural contradiction of both invalid data uploads and missing valid data. Therefore, there is an urgent need for a precast beam yard safety monitoring method that is production progress-driven, dynamically schedules sensing resources, adaptively configures imaging parameters, optimizes quality through closed-loop feedback, and provides early warning of violations.

[0019] This application provides a method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision. Please refer to [link / reference]. Figure 1 The specific steps are as follows: S1. Process Risk Mapping: Connect with the precast beam yard production progress management system and building information model database to extract the current beam fabrication process stage, platform number, and work space coordinates. Map the process sequence nodes to the beam yard spatial coordinate system, establish a dynamic correspondence between process status and work area, determine the spatial boundary and geometric topological features of the current active work surface, call the process risk association rule library based on the current process stage, deduce the spatial distribution of high-risk areas, assign area perception priority weights according to risk level, establish mapping rules between risk level and resource scheduling, generate work surface risk distribution coordinates, and output them to the scheduling stage.

[0020] The production progress management system stores the work plans and actual start status of each beam support. Data fields include support number, current work code, planned start time, actual start time, and estimated completion time. The system synchronizes work status changes in real time via a database interface or message queue. The building information model (BIM) database uses the beam yard plane coordinate system as a reference, recording the spatial coordinates and geometric boundaries of each support, passageway, hoisting area, and equipment parking area. Work processes include six stages: rebar tying, formwork installation, concrete pouring, prestressing tensioning, beam moving and hoisting, and beam storage and curing. Each stage corresponds to a different operational space and risk characteristics. When mapping work sequence nodes, the center point of the support is used as the spatial anchor point, mapping the work code to the spatial coordinate set of that support and its surrounding associated areas, establishing a dynamic correspondence between work status and operational area. The process risk association rule base is stored on a local server, recording the high-risk operation types and spatial ranges corresponding to each process stage. For example, the high-risk areas during the rebar tying stage are the working surface above the platform and the rebar hoisting channel; during the formwork installation stage, they are the formwork stacking area and the assembly working surface; during the concrete pouring stage, they are the area covered by the concrete pump truck's placement radius and the vibration working surface; during the prestressing tensioning stage, they are the anchorage areas at both ends of the tensioning platform and the jack operating radius; during the beam moving and hoisting stage, they are the area on both sides of the gantry crane's traveling track and the radius area below the hoisted object; and during the beam storage and curing stage, they are the beam storage platform and the beam transport channel. Risk levels are divided into three levels according to the degree of operational hazard: beam moving and hoisting and prestressing tensioning are level one (high risk); concrete pouring and formwork installation are level two (medium risk); and rebar tying and beam storage and curing are level three (low risk). When assigning regional perception priority weights according to risk level, the weight for level one risk areas is 1.0, for level two it is 0.7, and for level three it is 0.4. The mapping rule between risk level and resource scheduling aims to maximize the weight, generate the proportional relationship of the number of sensing nodes deployed in each region, and finally output the risk distribution coordinates of the work surface. The coordinate format is the two-dimensional coordinates of the regional center point, the risk level weight, and the recommended number of sensing nodes.

[0021] Through the above steps, process progress information is transformed into spatial risk distribution in real time, and the deployment of sensing resources is matched with the current work hotspots. This eliminates the spatial coverage blind spots caused by the assumption of a constant work surface in fixed arrays, and solves the spatial matching problem between sensing resource deployment and dynamic work surfaces. In existing technologies, sensing devices are fixed in physical locations such as rebar processing areas and beam fabrication platforms, and cannot be adjusted as processes move. This step, through real-time association between the production progress management system and the building information model, enables sensing scheduling to have process prediction capabilities, solving the structural mismatch problem of spatial coverage at its source.

[0022] S2. Sensing Node Scheduling: Receive the risk distribution coordinates output from step S1, drive the track-mounted mobile camera, adjustable gimbal, and drone sensing nodes to migrate to the high-risk work area, solve the optimal deployment position and viewing angle of the sensing nodes based on the coverage optimization algorithm, establish the matching relationship between the field of view coverage of the sensing nodes and the boundary of the work area with the spatial geometric characteristics of the current high-risk work area as constraints, eliminate the coverage redundancy in the overlapping areas of multiple nodes, and determine the blind area based on the field of view overlap rate threshold and trigger the supplementary node scheduling command, so that the sensing resources have no blind spot coverage in space and are continuously connected with the process migration in time.

[0023] The system includes track-mounted mobile cameras deployed along the longitudinal main passageway of the beam yard. The tracks are made of aluminum alloy, with a single track length of 50 to 100 meters. Driven by stepper motors, the movement speed is 0.1 to 0.5 meters per second, with a positioning accuracy of ±0.1 meters. The cameras are mounted on track sliders, which have built-in encoders that transmit position coordinates in real time. Adjustable gimbals are deployed on fixed poles at the four corners of the beam fabrication platform and the edge of the beam storage area. The gimbals have a horizontal rotation range of 0 to 360 degrees, a pitch angle of -90 to +90 degrees, and an adjustable rotation speed of 0.1 to 10 degrees per second, with a positioning accuracy of ±0.5 degrees. The cameras' viewing angle is remotely controlled via the gimbals. The drones are quadcopters with a flight time of at least 25 minutes and a flight altitude of 30 to 80 meters. Equipped with obstacle avoidance sensors and an RTK positioning module, they achieve a positioning accuracy of ±0.05 meters and are used to supplement the perception of temporary work areas not covered by the tracks and gimbals. The coverage optimization algorithm uses maximizing field-of-view coverage as the objective function and minimizing node movement paths and energy consumption as constraints. It employs either a genetic algorithm or a particle swarm optimization algorithm. Field-of-view coverage is defined as the ratio of the union area of ​​the field-of-view coverage of each node to the area of ​​the current high-risk work area. The path constraint is the sum of the cumulative movement distance of the track-mounted camera along the track and the flight distance of the drone. The energy consumption constraint is that the sum of the power consumption of each node and its movement power consumption must not exceed the current available energy budget. During node scheduling, track-mounted cameras and adjustable gimbals are prioritized for deployment in Level 1 risk areas, adjustable gimbals or drones are deployed in Level 2 risk areas, and nodes in Level 3 risk areas enter a dormant or low-power standby state. Redundancy is eliminated in multi-node overlapping areas through field-of-view overlap rate calculation. When the field-of-view overlap rate of two nodes exceeds 60%, it is considered redundant coverage, and the lower-priority node is released to other blind areas. Blind areas are calculated based on the field-of-view coverage gap area. When the gap area accounts for more than 10% of the work area, a supplementary node scheduling command is triggered, calling upon drones or dormant nodes from adjacent areas to fill the blind spot.

[0024] Through the above steps, the sensing coverage focus is dynamically adjusted as the work surface moves periodically. Nodes in the original low-risk areas go into dormancy or are released, ensuring that sensing resources have no blind spots in space and are continuously connected in sequence as the work process moves, thus solving the problem of spatial coverage blind spots in fixed arrays. In existing technologies, fixed-point cameras cannot achieve dynamic migration of the sensing focus; this step, through the coordinated scheduling of track-based movement, gimbal control, and drone supplementation, upgrades the deployment of sensing resources from static and fixed to dynamic adaptation, eliminating sensing blind spots after the work surface is moved.

[0025] S3. Imaging Parameter Adaptation: Based on the current process type and environmental characteristics identified in step S1, the imaging parameters of the sensing nodes deployed in step S2 are adaptively configured. The visible light and infrared thermal imaging modes are switched and the exposure gain and polarization filter angle are adjusted. The frame rate resolution and target motion speed are matched. A comparison index between process type and environmental parameters is established. The imaging mode is automatically switched according to the current process lighting conditions and background complexity to make the sensing configuration compatible with the current process lighting conditions and background complexity.

[0026] The visible light camera uses a CMOS sensor with a resolution of at least 1920×1080 pixels, an adjustable frame rate of 25 to 60 frames per second, and an adjustable exposure time of 1 to 30 milliseconds. It is used for worker behavior recognition and equipment status monitoring under normal daylight conditions. The infrared thermal imaging camera uses an uncooled focal plane detector with a resolution of at least 640×512 pixels, a frame rate of 30 frames per second, and a temperature detection range of -20 to 150 degrees Celsius. It is used for personnel positioning and equipment overheat monitoring in nighttime or low-light conditions. The polarization filter uses a linear or circular polarizer with adjustable polarization direction, mounted at the front of the visible light camera lens, to eliminate specular reflection interference from aluminum profile surfaces, concrete formwork, and welded steel reinforcement areas. The near-infrared supplementary lighting source uses an LED array with a wavelength of 850 nanometers and an illumination distance of 30 to 50 meters. It is used for auxiliary lighting in low-light nighttime conditions, with the supplementary light intensity automatically adjusted according to the ambient light level. Multi-frame fusion exposure involves capturing short exposure frames and long exposure frames and then weighting and fusing them. The short exposure time of 1 / 1000 second is used to suppress overexposure in strong light, while the long exposure time of 1 / 30 second is used to enhance shadow details. The fusion weights are adaptively allocated based on pixel brightness, ensuring that image details are fully preserved in high dynamic range scenes.

[0027] The imaging mode switching logic is driven by a process type and environmental parameter mapping index, which is stored in a local configuration library, recording the environmental characteristics and recommended imaging parameter combinations for each process stage. During the rebar tying stage, under normal daylight, visible light mode is enabled at a frame rate of 25 frames per second with automatic exposure. During the formwork installation stage, under normal daylight, visible light mode is enabled at a frame rate of 25 frames per second with automatic exposure. During the concrete pouring stage, which includes nighttime operations and strong moisture interference, visible light mode is enabled with near-infrared supplemental lighting, a frame rate of 30 frames per second, and an exposure gain of +3 to +6 dB. During the prestressing tensioning stage, under normal daylight, visible light mode is enabled at a frame rate of 60 frames per second to capture the rapid movements of the tensioning process. During the beam moving and hoisting stage, which includes nighttime operations and strong arc light interference, visible light mode is enabled with polarization filtering, a frame rate of 30 frames per second, and automatic exposure. During the beam storage and curing stage, under low-light conditions at night, infrared thermal imaging mode is enabled at a frame rate of 30 frames per second, with a temperature range of -20 degrees Celsius to 80 degrees Celsius. Parameter configuration is automatically executed by the controller built into the sensing node. The configuration command is triggered by the process identification result of step S1, without the need for manual intervention.

[0028] Through the above steps, the sensing configuration is adapted to the lighting conditions and background complexity of the current process, ensuring imaging quality in low-light conditions at night, strong arc interference, and high dynamic range scenarios, thus solving the adaptation problem between a single sensing configuration and heterogeneous environments. In existing technologies, a single sensing configuration cannot adaptively switch, resulting in excessive noise in nighttime images and overexposure due to arc interference during the day. This step achieves automatic adaptation of imaging parameters through a comparison index of process type and environmental parameters, ensuring stable image acquisition quality under different process environments.

[0029] S4. Quality Feedback Optimization: Real-time data quality assessment is performed on the video streams acquired in steps S2 and S3. The assessment indicators include image sharpness, target occlusion rate, and signal-to-noise ratio. If the quality does not meet the visual analysis threshold, a feedback instruction is generated and sent back to step S2 to adjust the node position or step S3 to correct the imaging parameters. The optimal configuration strategy under different process environment combinations is recorded, a perception configuration knowledge base is constructed, and iterative optimization of the perception strategy is realized.

[0030] Image sharpness is evaluated using the Laplacian operator variance. The variance is calculated after performing Laplacian convolution on video frames. A variance greater than 100 indicates sharpness, 50-100 indicates slight blur, and less than 50 indicates severe blur. Target occlusion rate is calculated using the foreground mask area ratio. The foreground region is extracted using background subtraction, and the ratio of the foreground area to the total field of view is calculated. A ratio less than 30% indicates no occlusion, 30%-60% indicates partial occlusion, and greater than 60% indicates severe occlusion. Signal-to-noise ratio (SNR) is calculated as the ratio of the image grayscale standard deviation to the noise standard deviation. An SNR greater than 20 dB is considered good, 10-20 dB is considered average, and less than 10 dB is considered poor. The visual analysis thresholds are: image sharpness variance not less than 80, target occlusion rate not more than 40%, and SNR not less than 15 dB. All three indicators must be met simultaneously for the image to be considered acceptable; failure to meet any one triggers feedback optimization.

[0031] Feedback instructions include two categories: position adjustment instructions and parameter adjustment instructions. Position adjustment instructions are sent back to step S2, controlling the track-mounted camera to move ±5 meters along the track or the pan-tilt unit to rotate ±15 degrees to avoid obstructions or optimize the field of view. Parameter adjustment instructions are sent back to step S3, controlling the exposure gain to adjust ±1 dB to ±3 dB, the frame rate to a higher level, or the supplementary lighting intensity to increase by 20% to 50%. After each feedback adjustment, video is re-acquired and the quality is evaluated until the threshold is met or the maximum number of adjustments is reached (3 times). The optimal configuration strategy under different process environment combinations is recorded in the perception configuration knowledge base. The knowledge base is stored in the form of a relational database. The table structure includes fields for process stage, ambient light intensity, time, optimal imaging mode, optimal exposure parameters, optimal frame rate, and optimal node position. Records are automatically appended after each quality evaluation, forming a mapping relationship between process environment combinations and optimal configuration strategies. This allows for direct retrieval under the same or similar working conditions, avoiding repeated trial and error and achieving iterative optimization of the perception strategy.

[0032] Through the above steps, video data quality is guaranteed in real time, invalid data uploads are suppressed, missing valid data is compensated for, and the perception strategy is continuously optimized during use, solving the problems of unstable data quality and static, fixed strategies. In existing technologies, perception parameters remain fixed, and image quality cannot be automatically recovered after environmental changes cause a decline. This step, through a closed-loop mechanism of quality assessment and feedback optimization, enables the system to have on-site adaptive and self-evolving capabilities, and the perception configuration knowledge base is improved over time.

[0033] S5. Violation Linkage Early Warning: Receive video data filtered in step S4, perform target detection, behavior recognition, and violation assessment. Behavior recognition includes wearing safety helmets, wearing reflective vests, entering dangerous areas, standing under suspended loads, and smoking in violation of regulations. Based on the risk level, trigger on-site audio-visual warnings and graded linkage early warnings pushed to mobile devices. Implement equipment interlock control in high-risk areas, and associate the early warning events with the process information in step S1 to generate a process risk association report.

[0034] The object detection function employs a deep learning object detection network. The input image size is 640×640 pixels, and the confidence threshold is set to 0.5. Detection categories include personnel, safety helmets, reflective vests, machinery, suspended objects, and open flames / smoke. The output is the object bounding box coordinates and category label. The behavior recognition function extracts the coordinate sequence of key human points based on a human pose estimation network, constructs a human skeleton model, tracks limb movement trajectories, and uses a 30-frame temporal analysis window covering behavior segments from 1 to 1.2 seconds to identify the standardization and continuity of work processes. Violation assessment compares real-time behavioral characteristics with a pre-set safety behavior characteristic database. Violation assessment is determined by the absence of a helmet in the head area for more than 3 seconds, the absence of a reflective vest in the torso area for more than 3 seconds, the intrusion into a dangerous area by the intersection of the personnel target boundary frame and the pre-set restricted area boundary frame for more than 2 seconds, the standing of a person under a suspended load by the distance between the personnel target and the suspended load target boundary frame being less than 2 meters and the suspended load being suspended in the air, and the illegal smoking by the detection of an open flame or smoke target located in a non-designated hot work area.

[0035] The risk level is divided into three levels: Level 1 is a high-risk warning, corresponding to personnel standing under the hoisted load during beam lifting and personnel entering the tensioning zone during prestressing tensioning, triggering a high-decibel audible and visual alarm, on-site broadcast warning, real-time push notifications on mobile devices, and emergency braking interlocks for the gantry crane or tensioning equipment; Level 2 is a medium-risk warning, corresponding to personnel not wearing safety helmets during concrete pouring and not wearing reflective vests during nighttime operations, triggering a medium-decibel audible and visual alarm, on-site broadcast reminders, and push notifications on mobile devices; Level 3 is a low-risk warning, corresponding to brief unauthorized smoking during rebar tying, triggering a low-decibel warning sound and mobile device recording. The audible and visual alarm uses a combination of a red rotating warning light and a high-volume horn, with a sound pressure level of no less than 90 decibels at the high-decibel setting, 70 to 90 decibels at the medium-decibel setting, and 50 to 70 decibels at the low-decibel setting. The equipment interlocking control is implemented through a programmable logic controller. When a first-level warning is triggered, an emergency stop signal is sent to the control circuit of the gantry crane or tensioning equipment to cut off the power supply and lock the brake. The system is manually reset after on-site confirmation by management personnel.

[0036] The warning event is associated with the process information in step S1. Recorded fields include warning time, platform number, current process stage, violation type, risk level, captured image of the violating personnel, and handling status. The process risk association report generates violation frequency heatmaps, time-period distribution curves, and team violation rankings on a daily, weekly, and monthly basis. The violation frequency heatmap uses the beam yard layout as its base map, aggregating the number of violation events by region and rendering them with color gradients. The time-period distribution curve uses 24 hours as the horizontal axis to count the number of violations in each time period. The team violation ranking is sorted by a weighted score of the number of violations and their severity. Report data is stored in a relational database and can be queried and analyzed by managers through a front-end visual interface, providing data support for optimizing management strategies.

[0037] Through the above steps, violations are accurately identified and handled in a tiered manner. Early warning information penetrates highly disruptive work environments to reach the target recipient, upgrading safety management from passive response to proactive prevention and control. This addresses the mismatch between a single early warning method and risk management. Existing technologies employ single and fixed early warning methods, failing to adapt to the handling needs of different risk levels. This step, by combining tiered, linked early warning with equipment interlocking control, ensures precise matching of early warning responses to risk levels, physically preventing high-risk violations from occurring instantly.

[0038] Based on Implementation Method 1, the track-mounted mobile camera in step S2 can use a magnetic track instead of a traditional sliding track. The magnetic track is installed along the top or side wall of the main passageway of the beam yard. The camera is suspended below the track by a magnetic base. The magnetic base has a built-in electromagnet and roller assembly. When powered on, it generates an attractive force to fix it to the track. When powered off, it releases for quick disassembly and relocation. Flexible connecting sections are used at the corners of the track to adapt to the irregular layout of the beam yard. With the above magnetic track configuration, the deployment position of the sensing node can be flexibly adjusted according to the needs of beam yard renovation and expansion without damaging the ground or pouring foundations, thus improving installation and maintenance efficiency.

[0039] Building upon Implementation Method 1, the multi-frame fusion exposure in step S3 can replace the fixed short-long exposure combination with adaptive exposure based on the scene brightness histogram. This involves real-time analysis of the current scene brightness distribution, dynamically determining the optimal exposure time combination, automatically shortening the exposure time for bright areas, and automatically extending the exposure time for shadow areas. During fusion, weights are adaptively allocated based on the local brightness variance. Through this adaptive exposure strategy, the exposure parameters are adjusted in real-time according to the scene content, making it more adaptable to the dynamic changes under complex lighting conditions in the beam yard compared to a fixed combination, thus improving the detail retention capability in high dynamic range scenes.

[0040] Building upon Implementation Method 1, the perception configuration knowledge base in step S4 can replace the static recording mechanism with an incremental learning mechanism. After each successful quality assessment, not only is the optimal configuration strategy recorded, but new process environment combinations are also identified through cluster analysis. When encountering a working condition not recorded in the knowledge base, an exploratory parameter search is automatically triggered. During the search, the quality score of each parameter combination is recorded, and the combination with the highest score is added to the knowledge base. Through this incremental learning mechanism, the knowledge base continuously expands with usage, its adaptability to new working conditions is constantly enhanced, and the optimization of the perception strategy evolves from experience accumulation to autonomous evolution.

[0041] Based on Implementation Method 1, the equipment interlocking control in step S5 can use a wireless communication link to replace the wired control loop. A communication connection is established between the sensing node and the gantry crane or tensioning equipment via an industrial wireless local area network. When an early warning is triggered, an emergency stop signal is issued via wireless command. The equipment is equipped with a dual-channel redundant receiving module, with the main channel and backup channel alternately verifying the integrity of the command. If either channel successfully receives the command, the interlocking action is executed. Through this wireless interlocking configuration, the wiring requirements between the sensing node and the controlled equipment are significantly reduced, making it particularly suitable for scenarios where the beam yard work surface frequently relocates and equipment positions dynamically change, thus improving the deployment flexibility of the interlocking control.

[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, characterized in that, The specific steps are as follows: S1. Process Risk Mapping: Link the precast beam yard production progress management system to obtain the current beam fabrication process stage and work scope, determine the distribution of high-risk areas, and generate risk coordinates for the work surface. S2, Sensing Node Scheduling: Receive the risk coordinates from step S1, control the sensing nodes to migrate to high-risk work areas, solve for the optimal deployment position and viewing angle, and dynamically adjust the sensing focus as the work area migrates. S3. Imaging parameter adaptation: Based on the process type and environmental characteristics identified in step S1, the imaging parameters of the nodes deployed in step S2 are configured, the imaging mode is switched and the exposure gain is adjusted to adapt the configuration to the process environment. S4. Quality Feedback Optimization: Evaluate the quality of the video collected in steps S2 and S3. If the video does not meet the analysis threshold, feed back to step S2 to adjust the position or step S3 to correct the parameters, and record the optimal configuration strategy. S5. Violation Linkage Early Warning: Receive the video data filtered in step S4, perform target detection and behavior recognition, trigger on-site audio and visual warnings according to the risk level, and associate the warning event with the process information in step S1.

2. The method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, as described in claim 1, is characterized in that... In step S1, the precast beam yard production progress management system and building information model database are connected to extract the current beam fabrication process stage, platform number, and work space coordinates. The process stages include rebar binding, formwork installation, concrete pouring, prestressing tensioning, beam moving and hoisting, and beam storage and curing. The process sequence nodes are mapped to the beam yard spatial coordinate system to establish a dynamic correspondence between process status and work area, determine the spatial boundary and geometric topological characteristics of the current active work surface, and provide a spatial benchmark for risk simulation.

3. The method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, as described in claim 2, is characterized in that... In step S1, the process risk association rule library is called according to the current process stage. The rule library stores the high-risk operation types and spatial range templates corresponding to each process stage, deduces the spatial distribution of high-risk areas, assigns regional perception priority weights according to risk levels, establishes mapping rules between risk levels and resource scheduling, generates risk distribution coordinates of the work surface and outputs them to the scheduling link, so that the deployment of perception resources matches the current work hotspots and eliminates the spatial coverage blind spots of the fixed array.

4. The AI ​​vision-based method for identifying and issuing early warnings of safety violations in precast beam yards, as described in claim 3, is characterized in that... In step S2, the risk distribution coordinates output in step S1 are received, and the orbital mobile camera, adjustable gimbal, and UAV sensing nodes are driven to migrate to the high-risk work area. The optimal deployment position and viewing angle of the sensing nodes are solved based on the coverage optimization algorithm. The coverage optimization algorithm takes maximizing the field of view coverage as the objective function and the shortest node movement path as the constraint condition, so that the sensing coverage focus is dynamically adjusted as the work area migrates periodically, and the nodes in the original low-risk area go into dormancy or are released.

5. The AI ​​vision-based method for identifying and issuing early warnings of safety violations in precast beam yards, as described in claim 4, is characterized in that... In step S2, the spatial geometric features of the current high-risk work surface are used as constraints to establish a matching relationship between the field of view coverage of the sensing nodes and the boundary of the work surface. Coverage redundancy is eliminated in the overlapping areas of multiple nodes. For blind areas, the field of view overlap rate threshold is used to determine and trigger the supplementary node scheduling command, so that the sensing resources have no blind spots in space and are continuously connected with the process migration in time.

6. The method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, as described in claim 5, is characterized in that... In step S3, based on the current process type and environmental characteristics identified in step S1, the imaging parameters of the sensing nodes deployed in step S2 are adaptively configured. The visible light and infrared thermal imaging modes are switched and the exposure gain and polarization filter angle are adjusted to match the frame rate resolution and target motion speed. The visible light mode is used for normal daytime lighting scenarios, and the infrared thermal imaging mode is used for nighttime or low-light scenarios, so that the sensing configuration is adapted to the lighting conditions and background complexity of the current process.

7. The method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, as described in claim 6, is characterized in that... In step S3, a comparison index between process type and environmental parameters is established. The imaging mode is automatically switched according to the current process lighting conditions and background complexity. Near-infrared supplementary lighting is enabled for low-light night scenes, polarization filtering is enabled for strong arc light interference scenes, and multi-frame fusion exposure is enabled for high dynamic range scenes. The multi-frame fusion exposure suppresses overexposure of highlights and underexposure of shadows by fusing short exposure frames and long exposure frames, so that the imaging quality is adapted to the current process environment.

8. The method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, as described in claim 7, is characterized in that... In step S4, real-time data quality assessment is performed on the video streams acquired in steps S2 and S3. The assessment indicators include image sharpness, target occlusion rate, and signal-to-noise ratio. Image sharpness is evaluated using the variance of the Laplacian operator, and the target occlusion rate is calculated using the foreground mask area ratio. If the quality does not meet the visual analysis threshold, a feedback instruction is generated and sent back to step S2 to adjust the node position or to step S3 to correct the imaging parameters. The optimal configuration strategy under different process environment combinations is recorded, a perception configuration knowledge base is constructed, and iterative optimization of the perception strategy is realized.

9. The method for identifying and issuing early warnings of safety violations in precast beam yards based on AI vision, as described in claim 8, is characterized in that... In step S5, the video data filtered in step S4 is received, and target detection, behavior recognition, and violation assessment are performed. The behavior recognition includes wearing a safety helmet, wearing a reflective vest, entering a dangerous area, standing under a suspended load, and smoking in violation of regulations. Based on the risk level, on-site sound and light warnings and graded linkage warnings pushed by mobile terminals are triggered. Equipment interlock control is implemented in high-risk areas. At the same time, the warning events are associated with the process information in step S1 to generate a process risk association report.

10. The AI ​​vision-based method for identifying and issuing early warnings of safety violations in precast beam yards, as described in claim 9, is characterized in that... In step S5, confidence levels are calculated for the identified violations, distinguishing between suspected and confirmed violations. Risk levels are assigned based on the type of violation, the area where it occurred, and the stage of operation. Warning events are correlated in multiple dimensions according to process stage, team information, and equipment status to generate statistical analysis reports on high-frequency violation scenarios and time periods. These reports include a heatmap of violation frequency, a time period distribution curve, and a team violation ranking, providing data support for optimizing management strategies.

Images