Intelligent construction site real-time safety monitoring system and method based on artificial intelligence

Abstract

The invention discloses an intelligent construction site real-time safety monitoring system and method based on artificial intelligence, belongs to the technical field of construction engineering management, and particularly relates to a method for acquiring video streams, images and environment data through a high-definition camera, an IoT sensor and an unmanned aerial vehicle, integrating and improving YOLOv8, LSTM and Transformer algorithms based on edge computing, and monitoring the safety of a construction site in real time through the improved YOLOv8, LSTM and Transformer algorithms. And real-time monitoring of PPE wearing, behavior abnormity and environmental risks is realized. The system delay adopted by the invention is lt; the false alarm rate is 1s; through implementation verification, the accident rate is reduced by 35%, and the method is suitable for a complex construction site environment.

Description

Technical Field

[0001] This invention relates to a real-time safety monitoring system and method for smart construction sites based on artificial intelligence, belonging to the field of construction engineering management technology. Background Technology

[0002] Traditional safety monitoring mainly relies on manual inspections, which has problems such as limited coverage, low efficiency and slow response, making it difficult to meet the real-time monitoring needs of modern construction sites.

[0003] Currently, among existing technologies, the Chinese patent with publication number CN120517996A, entitled "Smart Construction Site Tower Crane Safety Monitoring Method, System and Equipment," has achieved some data collection functions, but it is mostly passive recording and lacks intelligent analysis capabilities, failing to automatically identify dynamic behavioral risks or predict potential hazards. While most smart construction site digital management platforms on the market possess data integration functions, they have not effectively integrated multimodal AI algorithms to achieve end-to-end safety monitoring closed loops. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a real-time safety monitoring system and method for smart construction sites based on artificial intelligence. By integrating and optimizing the real-time monitoring system and method of YOLOv8, LSTM and Transformer algorithms, it aims to significantly improve the efficiency of construction site safety management through high-precision risk identification and rapid response.

[0005] To achieve the above objectives, the technical solution adopted by this invention is a real-time safety monitoring system for smart construction sites based on artificial intelligence, including... The data acquisition module is used to collect video streams, images, and environmental data from the construction site. The AI ​​processing module runs on an edge computing server and integrates an improved YOLOv8 target detection algorithm, an LSTM time series analysis model, and a Transformer multimodal risk prediction model to analyze personnel behavior, equipment status, and environmental risks in real time. The early warning and decision-making module is equipped with a rules engine that generates multi-level alerts based on detection results, including voice broadcasts, APP push notifications, and platform log recordings. The integrated management module is used to interface with the smart construction site digital management platform via API and generate visual dashboards and automated safety reports.

[0006] Preferably, the data acquisition module is equipped with a high-definition camera, an IoT sensor, and a drone inspection device. The high-definition camera supports night vision and wide-angle coverage, and the IoT sensor includes an infrared thermal imaging sensor, a vibration sensor, a temperature sensor, a humidity sensor, and a noise sensor, with a data acquisition frequency of no less than 10 times per second.

[0007] Preferably, the YOLOv8 target detection algorithm of the AI ​​processing module is trained based on a proprietary construction site video dataset, with a detection accuracy of over 95%, and is used to detect PPE wearing status. The LSTM time series analysis model optimizes trajectory tracking through Kalman filtering, reducing the false alarm rate by 20%. The AI ​​processing module's Transformer risk prediction model integrates historical accident data and real-time weather input, achieving a prediction accuracy of 90% and supporting the output of future risk probabilities 5-10 minutes in advance.

[0008] Preferably, the early warning and decision-making module supports custom threshold configuration, triggers an alarm when the PPE missing confidence level is below 0.8, and has a three-level response strategy: low-level push notification, medium-level device suspension, and high-incidence level emergency shutdown.

[0009] Preferably, the integrated management module provides risk heatmaps, equipment status charts, and historical trend analysis, supports CSV and JSON data export, and has a storage capacity of no less than 10TB.

[0010] A real-time safety monitoring method for smart construction sites based on artificial intelligence is implemented according to the following steps. S1. Collect on-site data through the data acquisition module, including real-time video streams, images and environmental data of the construction site collected by high-definition cameras, IoT sensors and drones; S2. Denoise and enhance the acquired video data, and normalize the sensor data. S3. Input the processed data into the improved YOLOv8 target detection algorithm model to detect construction site objects, and input the data into the LSTM time series analysis model to analyze the behavior and calculate the risk score. ,when When the value is >0.7, the Transformer multimodal risk prediction model is triggered to predict the probability of potential hazards in the next 5-10 minutes and determine the risk level. S4. Determine whether to implement a low-level push notification, a medium-level equipment suspension, or a high-level emergency shutdown based on the risk level. S5 records response effects and actual accident data, feeds them back to the AI ​​processing module model for online learning, updates parameters daily, and achieves adaptive optimization.

[0011] Preferably, in step 3, a comprehensive risk score is calculated. The formula used is: ,in: The detection confidence score for the YOLOv8 object detection algorithm ranges from 0 to 1, reflecting the risk of object detection. The score is the behavioral analysis score of the LSTM time series analysis model, ranging from 0 to 1, reflecting the persistence of abnormal behavior. Environmental risk weights, ranging from 0 to 1, are predicted based on the Transformer risk prediction model. These are weighting coefficients, dynamically adjusted according to the type of construction site. On sites with a lot of equipment, the α proportion is higher; on sites with a lot of people, the β proportion is higher; and on sites with many high-rise buildings or inclement weather, the γ proportion is higher, satisfying the following conditions: ; The time interval since the last risk assessment is in minutes. This is the time decay constant, expressed in minutes, reflecting the characteristics of risk decay over time.

[0012] Preferably, in step 3, when When the value is >0.7, the Transformer model is triggered to predict the probability of potential hazards in the next 5-10 minutes. when When the risk level is determined to be low, a notification is sent. when At that time, the risk level was determined to be medium, and the equipment was suspended. when If the risk level is determined to be high, work should be stopped immediately and the emergency response team should be notified.

[0013] Compared with existing technologies, this invention has the following technical advantages: By integrating computer vision, deep learning, and edge computing technologies, this invention achieves real-time, multi-dimensional monitoring of personnel behavior, machinery and equipment status, and environmental risks at construction sites. Addressing the shortcomings of existing technologies in dynamic behavior recognition, predictive early warning, and integration with digital management platforms, this invention aims to significantly reduce construction site accident rates (target reduction of over 30%), shorten response time (target <10 seconds), and seamlessly adapt to the company's smart construction site digital management platform through high-precision AI algorithms and end-to-end closed-loop design, thereby improving overall safety management efficiency. Attached Figure Description

[0014] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0015] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0016] A smart construction site real-time safety monitoring system based on artificial intelligence includes a data acquisition module, an AI processing module, an early warning and decision-making module, and an integrated management module.

[0017] The data acquisition module is equipped with a high-definition camera, IoT sensors, and drone inspection equipment to collect video streams, images, and environmental data from the construction site. The high-definition camera supports night vision and wide-angle coverage, while the IoT sensors include an infrared thermal imaging sensor, a vibration sensor, a temperature sensor, a humidity sensor, and a noise sensor, with a data acquisition frequency of no less than 10 times per second.

[0018] The AI ​​processing module, running on an edge computing server, integrates an improved YOLOv8 target detection algorithm, an LSTM time series analysis model, and a Transformer multimodal risk prediction model. It is used to analyze personnel behavior, equipment status, and environmental risks in real time. The YOLOv8 target detection algorithm, trained on a proprietary construction site video dataset, achieves a detection accuracy of over 95% and is used to detect PPE wearing status. The LSTM time series analysis model optimizes trajectory tracking through Kalman filtering, reducing the false alarm rate by 20%. The Transformer risk prediction model integrates historical accident data and real-time weather input, achieving a prediction accuracy of 90% and supporting the output of future risk probabilities 5-10 minutes in advance.

[0019] The early warning and decision-making module is equipped with a rule engine that generates multi-level alarms based on detection results, including voice broadcasts, APP push notifications, and platform log recordings. It allows for customizable threshold configurations, triggering an alarm when the PPE missing confidence level is below 0.8, and has a three-level response strategy: low-level push notification, medium-level device suspension, and high-incidence emergency shutdown.

[0020] The integrated management module interfaces with the smart construction site digital management platform via API, generating visual dashboards and automated safety reports. It provides risk heatmaps, equipment status charts, and historical trend analysis, supports CSV and JSON data export, and has a storage capacity of at least 10TB.

[0021] The specific plan is to be implemented according to the following steps. S1. Collect on-site data through the data acquisition module, including real-time video streams, images and environmental data of the construction site collected by high-definition cameras, IoT sensors and drones; S2. Denoise and enhance the acquired video data, and normalize the sensor data. S3. Input the processed data into the improved YOLOv8 target detection algorithm model to detect construction site objects, and input the data into the LSTM time series analysis model to analyze the behavior and calculate the risk score. ,when When the value is >0.7, the Transformer multimodal risk prediction model is triggered to predict the probability of potential hazards in the next 5-10 minutes and determine the risk level. when When the risk level is determined to be low, a notification is sent. when At that time, the risk level was determined to be medium, and the equipment was suspended. when If the risk level is determined to be high, work should be stopped immediately and the emergency response team should be notified.

[0022] In calculating the comprehensive risk score When using this formula: ,in: The detection confidence score for the YOLOv8 object detection algorithm ranges from 0 to 1, reflecting the risk of object detection. The score is the behavioral analysis score of the LSTM time series analysis model, ranging from 0 to 1, reflecting the persistence of abnormal behavior. Environmental risk weights, ranging from 0 to 1, are predicted based on the Transformer risk prediction model. These are weighting coefficients, dynamically adjusted according to the type of construction site. On sites with a lot of equipment, the α proportion is higher; on sites with a lot of people, the β proportion is higher; and on sites with many high-rise buildings or inclement weather, the γ proportion is higher, satisfying the following conditions: ; The time interval since the last risk assessment is in minutes. This is the time decay constant, expressed in minutes, reflecting the characteristics of risk decay over time.

[0023] S4. Determine whether to implement a low-level push notification, a medium-level equipment suspension, or a high-level emergency shutdown based on the risk level. S5 records response effects and actual accident data, feeds them back to the AI ​​processing module model for online learning, updates parameters daily, and achieves adaptive optimization.

[0024] The following description uses specific real-world cases as examples. PPE Inspection and Emergency Response Applications This system was deployed at a high-rise building construction site in Xiahe, Shanxi Province, to verify its effectiveness in personnel safety protection and emergency response. The construction site covers an area of ​​approximately 50,000 square meters and is equipped with 10 high-definition cameras (night vision wide-angle, 30fps), 5 sets of IoT sensors (including infrared thermal imaging and vibration sensors), and a drone equipped with a 4K camera. Data is transmitted to an edge computing server via a 5G network.

[0025] Step 1: Data Acquisition. The system initiates 24 / 7 monitoring in the core areas of the construction site (tower crane operating area and aerial work platform). Cameras capture 30 frames of video per second, and sensors record temperature (range 15-35°C), humidity (50-80%), and vibration data 10 times per second. Drones conduct inspections once per hour to cover blind spots.

[0026] Step 2: Preprocessing. The video frames are enhanced for low light using a GAN model to remove noise from nighttime data. Sensor data is normalized and converted to the 0-1 range to ensure consistency.

[0027] Step 3: AI Analysis A. The optimized YOLOv8 model was used to detect workers' PPE wearing status. It was trained on the company's historical construction site video dataset (containing 100,000 labeled frames) and achieved a detection accuracy of 95%. One worker was identified as not wearing a safety helmet with a confidence level of 0.75.

[0028] B.LSTM model analysis of 10 consecutive seconds of frame data (300 frames), combined with Kalman filtering to optimize the trajectory, confirmed that the worker had been in the high-altitude work area for more than 3 seconds, and the behavioral risk score was determined. = 0.85.

[0029] C. Integrating environmental data (70% humidity, 25°C), the Transformer model predicts slip risk. A comprehensive risk score is calculated. : , set up = 0.3、 = 0.4、 = 0.3 (weight of high-rise construction sites). = 0.75、 = 0.6 (based on humidity weight) = 5 minutes = 10 minutes, substituting, we get:

[0030] because <0.7, initially judged as low risk; however, after continuous monitoring for 2 minutes, = 7 minutes The value rose to 0.72, triggering an alert.

[0031] Step 4: Early Warning Response, System Detection = 0.72, judged as a low-level risk, automatically suspending the operation of nearby tower cranes via relay (response time 8 seconds), and pushing a notification to the site supervisor's APP (including worker location coordinates and risk details). After the supervisor confirms, the safety officer is arranged to intervene.

[0032] Step 5: Feedback and optimization. Record the intervention results (risk eliminated after workers wore safety helmets) and feed them back to the AI ​​model. The system analyzes the false alarm rate (0% in this case) and response time (8 seconds), updates the YOLOv8 and LSTM parameters daily, and optimizes them. , , Weights.

[0033] Effects of implementing the present invention: Safety Improvement: In the first month of deployment (October 2025), a total of 120 potential PPE defects were identified, a 50% increase compared to manual inspections (approximately 80 times per month), and the accident rate decreased by 35% (from 0.5% to 0.325%).

[0034] Response efficiency: The average response time has been reduced from 5 minutes for manual inspection to 8 seconds, meeting the target of <10 seconds.

[0035] Model optimization: Learning through feedback, The formula's false positive rate decreased from the initial 8% to 4%, and the prediction accuracy improved to 92%.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the scope of the present invention.

Claims

1. An artificial intelligence-based intelligent construction site real-time safety monitoring system, characterized in that: Comprising a data acquisition module for collecting video streams, images and environmental data of the construction site; an AI processing module based on edge computing server, integrating improved YOLOv8 target detection algorithm, LSTM time series analysis model and Transformer multi-modal risk prediction model, for analyzing personnel behavior, equipment status and environmental risk in real time; an early warning and decision-making module equipped with a rule engine, generating multi-level alarms based on detection results, including voice broadcast, APP push and platform log recording; an integrated management module for interfacing with the intelligent construction digital management platform through API interface, and generating visual dashboards and automated safety reports.

2. The real-time safety monitoring system for intelligent construction site based on artificial intelligence according to claim 1, characterized in that: The data acquisition module is equipped with high-definition cameras, Internet of Things sensors and unmanned aerial vehicle inspection equipment. The high-definition cameras support night vision and wide-angle coverage. The Internet of Things sensors include infrared thermal imaging sensors, vibration sensors, temperature sensors, humidity sensors and noise sensors, with a collection frequency of not less than 10 times per second.

3. The real-time safety monitoring system for intelligent construction site based on artificial intelligence according to claim 1, characterized in that: The YOLOv8 target detection algorithm of the AI processing module is trained based on a proprietary construction video dataset for detecting PPE wearing; The LSTM time series analysis model optimizes trajectory tracking through Kalman filtering; The Transformer risk prediction model of the AI processing module integrates historical accident data and real-time weather input to predict the risk probability in the next 5-10 minutes.

4. The real-time safety monitoring system for intelligent construction site based on artificial intelligence according to claim 1, characterized in that: The early warning and decision-making module supports custom threshold configuration, triggering an alarm when the PPE absence confidence is below 0.8, and has a three-level response strategy: low-level push notification, medium-level suspension of equipment, and high-level emergency shutdown.

5. The real-time safety monitoring system for intelligent construction site based on artificial intelligence according to claim 1, characterized in that: The integrated management module provides risk heat maps, equipment status charts and historical trend analysis for data storage and export.

6. The method of claim 1-5, wherein the method is characterized in that: The following steps are followed, S1, collect construction site data through the data acquisition module, including high-definition cameras, Internet of Things sensors and unmanned aerial vehicle real-time collected construction video streams, images and environmental data; S2, denoise and enhance the collected video data, and normalize the sensor data; S3, input the processed data into the improved YOLOv8 target detection algorithm model to detect the construction site objects, and input the LSTM time series analysis model to analyze the behavior and calculate the risk score When >0.7, trigger the Transformer multimodal risk prediction model to predict the hidden danger probability in the next 5-10 minutes, and determine the risk level; S4, determine to execute low-level push notification, medium-level suspension of equipment or high-level emergency shutdown according to the risk level; S5, record the response effect and actual accident data, and feed back to the AI processing module model for online learning, daily parameter update and adaptive optimization.

7. The artificial intelligence-based intelligent construction site real-time safety monitoring method of claim 6, characterized in that: In step 3, the overall risk score is calculated using the formula: wherein: The detection confidence of the YOLOv8 target detection algorithm ranges from 0 to 1, reflecting the target detection risk. The behavior analysis score for the LSTM time series analysis model, ranging from 0-1, reflects the abnormal behavior persistence; is the environmental risk weight, ranging from 0-1, based on the Transformer risk prediction model prediction; respectively, are weighting coefficients, which are dynamically adjusted according to the site type. In a site with more equipment, the proportion of a is high. In a site with more people, the proportion of β is high. In a site with more high-rise buildings and bad weather, the proportion of γ is high, satisfying ; is the time interval since the last risk assessment in minutes; is the time decay constant, in minutes, reflecting the risk decay characteristic over time. 8.The real-time safety monitoring method of the intelligent construction site based on artificial intelligence of claim 7, characterized in that: In step 3, when > 0.7, triggering the Transformer model to predict the hidden danger probability in the next 5-10 minutes; When the risk level is determined to be low, push notifications, When the risk level is determined to be medium, the device is suspended, When the risk level is determined to be high, an emergency shutdown is performed and the emergency team is notified.

Images