Risk Identification Assessment And Control Method Of Tunnel Construction Based On Human Representation And Scene Understanding

The safety management of tunnel construction is an important foundation to support the high-quality development of tunnel industry in China.In order to cope with the complex operational environment and engineering conditions,the safety risk management of tunnel construction site is the key and core.To improve the working mode of traditional safety inspection in construction,which is labor-based,time-consuming,labor-intensive,process-subjective and management-passive,this thesis,based on the perception of risk factors with deep learning for computer vision,deeply studies the structural representation and understanding algorithm aiming at "Man,Machine,Medium and Management" on the construction site,explores a dynamic inspection method for structured and ontology-based safety rules adapted to visual objects,and develops the vision-based automatic hazards identification and management system(VAHIMS)to establish the data archive,safety labels and portraits of tunnel construction scenes,and to provide decision-making assistance for safety managers to take targeted and differentiated control measures and management strategies according to the classification of scene safety features.Specific research contents and contributions are as follows:(1)Analysis of safety risk factors of tunnel construction and its mechanism.Based on the construction characteristics of tunnel engineering,this dissertation studied the risk factors of tunnel construction in combination with the connotation and extension of construction hazard sources in tunnel scenarios,then established the multi-level progressive influence structure among key risk factors by applying the Interpretative Structural Modeling method,and summarized the working mode and the challenges of tunnel safety risk management to indicate the research objectives of tunnel safety risk management and control.(2)Risk assessment of tunnel construction based on structural representation of human body.In order to reduce the safety risk caused by irregular operation behaviors of construction workers,a non-invasive and non-contact operation behavior risk assessment method under visual perception was studied.Firstly,the CNN-based human pose estimation model under complex construction environment was constructed to obtain the structural representation of 2D human pose data under the conditions of uneven illumination,background difference and partial occlusion.Secondly,a scheme to automatically extract the motion phase state features of human skeleton movements was proposed,which evaluated the working behavior standardization by the synchronous and asynchronous motion relationship between the human body joints while working.Thirdly,the ergonomics-centered REBA evaluation method was improved to eliminate the jumping interference effects at critical joint angle by fuzzy logic optimization,and the REBA score accuracy has been promoted.Finally,through the case study of climbing operation,the automatic identification and assessment process of working posture based on the structured information of human skeleton information was demonstrated.The results show that this method can effectively prevent the construction risk caused by improper working posture,prevent work-related musculoskeletal disorders and further evolution through improving the standard of working posture by safety training and education,and reduce the probability of accidents caused by its impact on the physical and mental state of workers.(3)On-site hazards identification and graded early warning based on visual semantic information.Considering that it is relatively inefficient for the overall safety management to discuss the worker behavior in isolation way without the information of tools and environments on the construction site,this dissertation presented the deep learning-based algorithms to extracted the visual information and semantic relationship features of "Man,Machine,Medium " through the two-stage scene graph generation method of entity object detection and visual relationship classification,and then proposed the pipeline of risk automatic reasoning module based on scenario rule graph.Firstly,the object detection results are fed to entity object pairing preprocessing based on domain prior knowledge,which improved the visual relationship prediction efficiency.Secondly,human-oriented visual relationship detection was realized by fusing visual image,positional relationship,human posture,semantic prior and other features,and the strategy of training data adjustment and Logit Adjustment-based loss function optimization during training process was adopted to suppress the influence of long tail effect to improve the accuracy of visual relationship prediction detection results.In addition,the safety inspection strategy based on the triplet formatted safety checking rules on-site was proposed to identify risks and grade early warnings.Finally,the experimental case was analyzed to demonstrate the automatic reasoning process of hazards based on the comparison of scene information and rule semantics on construction site,which improved the intelligent level of safety risk inspection.(4)Establishment of tunnel construction safety risk management and control system integrating scene safety rules.In order to adapt to the dynamic optimization requirements of safety management knowledge and rules which are constantly iteratively updated,a safety rules and knowledge ontology-based processing framework of construction was presented,combined with the safety label and scene portrait classification derived from visual safety checking results to support the targeted and differentiated decision-making output of risk management and control.Firstly,a structured model of safety rule text content based on natural language processing technology was studied,which can automatically extract key semantics in construction rule documents and represent the data in triples.Secondly,a ontology-based safety rules knowledge base was constructed to effectively manage all kinds of explicit and tacit safety knowledge of the text form in the construction process,which made up for the defects of manual extraction of text rules,and improved the construction risk automatic inference machine based on scenario-rule matching and comparison mechanism.On this basis,a visual safety inspection archive of construction site was established,and safety labels and scene portrait were extracted through K-means clustering algorithm to provide intuitive risk control decision-making basis for safety managers.Meanwhile,PDCA-SDCA dual cycle management mode was implemented,adopted the feedback and optimized dynamic adjustment mechanism to promote the steady improvement of tunnel construction safety management.(5)Case study of tunnel construction safety management and risk control.In regards to the defects of subjective dependence,labor-intensive,time-consuming,and laborintensive on-site safety inspection mainly based on manual observation of front-line safety personnel,this dissertation has proposed a unified VAHIMS framework for hazards identification,risks evaluation and control,and decision-making assistance based on visual monitoring data of construction sites.Then through the practical application and two-staged comparison experiment of safety management during Fuzhou water conveyance tunnel project,the architecture design,development scheme and application logic of the VAHIMS system were introduced.The results revealed that the vision-based safety hazards inspection can effectively identify various kinds of risks in the scene,reduce the workload of safety personnel,and promote the safety awareness of construction site workers,meanwhile combined with the classified management strategies of safety label and scene portrait can improve the overall safety level.Finally,based on my own experience,put forward suggestions for the development of tunnel construction safety risk management and control.Aiming at hazards identification and risks control tasks in the tunnel construction safety management,in this dissertation,visual sensing and deep learning technology are applied to inspect the visually measurable risk factors on-sites.The safety labels and scene portraits based on the inspected records data are established.Given all this,the research takes targeted and classified risk control measures to effectively improve the safety attributes and promote the development of tunnel safety management.