In order to accurately identify unsafe behaviors of personnel in complex industrial sites and reduce the occurrence of safety accidents，an improved YOLOv5 unsafe behavior detection model was proposed. Firstly，an attention mechanism was introduced in the backbone of YOLOv5 to enhance the sensitivity of convolutional networks to unsafe behavior features. Secondly，enriching the number of training samples through image geometric transformation and pixel-level processing enhanced the generalization ability of the detection model in different industrial environments. Then，the detection model was distilled，and the network structure parameters were optimized to accelerate the training of the mode. Finally，the model was trained and iterated 200 times to simulate three types of industrial sites: lifting slings，robot-automated production lines，and operating rooms. It detected whether personnel were wearing safety helmets，work clothes and working in safe areas，and determined the level of danger based on their behavior to ascertain whether they were working safely. The results show that the model can detect 12 types of unsafe behaviors of personnel in complex industrial environments，such as dim light，lighting，and occlusion. The accuracy on the unsafe behavior test set is 98.6%，the recall rate is 99.2%，and the average accuracy is 97.58%.

In order to improve the perception level of fault water damage in coal mine，a fault water inrush stage sensing method based on fault activation evolution mechanism and key control factors was proposed. The evolution characteristics of working face floor and fault failure were studied through similar simulation tests of fault water inrush evolution process. The stage characteristics of monitoring parameters and the change of water inflow were revealed by taking the stress in the failure zone of floor，the stress in the fracture zone of fault and the water pressure in the water channel as the stage monitoring parameters. The key controlling factors of water inrush phase transformation were determined by grey correlation analysis method. Then，according to the fault water inrush analysis and research process，the fault water inrush stage perception method was proposed. The study determines that the numerical variation characteristics of monitoring parameters such as stress of floor failure zone，fault fracture zone and water pressure in water channel show obvious stage characteristics during fault activation water inrush. The order of grey correlation degree between control factors and water inflow is as follows: stress in floor failure zone > stress in fault fracture zone > water pressure in water channel，through identification of key control factors. Through the identification of key control factors，the perception of fault water inrush stage can be realized，and the perception method based on grey correlation analysis is feasible in principle.

In order to accelerate the speed and efficiency of autonomous systems testing，the method of generating a scene database for unmanned driving in campus environments was proposed. Firstly，the simulation test scenarios in complex campus environment were analyzed，and the campus scenes were simplified as a combination of road network structure，ground properties，interacting members and environmental factors. Secondly，the method of generating the scene database based on importance indicators was proposed to solve the boundedness of the campus scenario database. Then，the complexity indicators and interest probability indicators were used to describe the importance indicators of scenarios. The fuzzy analytic hierarchy process(FAHP) was used to evaluate the complexity of the scenario. The interest probability of the scenario was calculated by combining the kernel density estimation method and the interested weight calculation method. Next，the parameter space was segmented to obtain the set of similar scenarios，and the scenario sets were sorted according to test priority and importance indicators. The filtered scenarios were gradually added to the test scenario database，and the scenario database with test sequences was generated. Finally，the test evaluations based on the real-world campus scenario database were conducted to verify the effectiveness of the scenario database generation method proposed in this paper. The results show that the campus test scenes can be effectively described using four scene elements and the tree structure. The method proposed in this paper can generate a campus test scene library with high test efficiency，high coverage，conformity to natural probability，and interest interval，which is helpful to improve the efficiency of unmanned simulation test in complex campus environment.

To reduce traffic accidents caused by autonomous vehicles and improve the efficiency of vehicle safety testing in simulation environments，an autonomous driving collision test scenario construction method was proposed based on deep reinforcement learning. Firstly，the vehicle's driving process was mapped to a Markov decision process by setting the state，action，and reward functions. Then，the agent was trained to complete the vehicle collision task and generate the collision test scenarios based on the built simulation platform (CARLA-DRL). Finally，500 random collision simulation tests were conducted to analyze the collision success rate，collision time，and collision energy based on the relative distance between the agent and the autonomous vehicle. The results indicated that the agent generated collision trajectories that conformed to vehicle dynamics and could construct refined and multi-type collision test scenarios. The average collision success rate between the agent and the autonomous vehicle was 62.20%，the average collision time was 127.25 s，and the average collision energy value was 175.98 kJ. The proposed method can construct high-frequency，high-efficient，and high-risk autonomous driving vehicle collision test scenarios，increasing the probability of occasional high-risk scenarios in simulation scenarios and enhancing the efficiency of safety testing for autonomous vehicle collision incidents.

An open-pit mine landslide identification method was proposed based on object-oriented annotation datasets and the Res-U-Net model to realize accurate identification and early warning of open-pit mile landslide disasters. Firstly，the mine landslide image data in the study area were obtained by UAV aerial survey. Secondly，the multi-scale-spectral segmentation method and threshold separation principle were applied to divide and classify the open-pit mine landslide data，and the landslide dataset was developed based on the object-oriented method. Then，the U-Net network was used as the infrastructure to propose a landslide identification semantic segmentation model based on Res-U-Net by integrating the residual module into each convolutional layer. Finally，the datasets constructed by different methods were used to identify landslides，and the Res-U-Net model was compared with the widely used semantic segmentation models，Fully Convolutional Networks (FCN)，and U-net. The results indicated that the landslide data set based on object-oriented annotation had better landslide identification performance when compared to the traditional manual annotation dataset，resulting in improvements in identification accuracy，recall rate，F₁ score，and kappa coefficient of more than 12%. The landslide identification accuracy of the Res-U-Net model was more than 0.8，realizing the accurate landslide open-pit mine disaster identification.

In order to improve the science and effectiveness of traceability and localization of hazardous gas leaks，determining the location and intensity of dangerous gas leaks is the key to emergency response to accidents. The Gaussian plume model was modified by analyzing the mass conservation law and improving the diffusion amplitude of the gas plume with an approximate Gaussian distribution. Additionally，a heuristic algorithm based on the principle of immunization—IA coupled with PSO—was proposed，and the PSO-IA algorithm was applied to source strength inversion. It is concluded that the modified Gaussian plume model has been verified by three classical algorithms (PS，GA and PSO)，resulting in a prediction value error decreased by about 2%. PSO algorithm，which showed a better inversion effect，was selected for comparison with the PSO-IA algorithm. The PSO-IA algorithm has improved the effect of inverting source strength，with a localization error is 1.3 m，a source strength solving error of 0.8%，and a single computation time of less than 1 second. This enables fast and accurate positioning and estimation of source strength.

To solve the intelligent detection problem of fire lane occupancy warning，a lightweight early warning approach based on YOLOv7 was proposed by introducing the principles of area intrusion. Firstly，a research framework for detecting fire lane area intrusions was devised，utilizing the YOLOv7 model. This was accompanied by the compilation of an image dataset that encompassed fire lanes and vehicle detection，sourced from both field investigations and open datasets. Subsequently，the spatial pyramid pooling's multi-stage partial convolution was substituted with a receptive field block module，and the SimAM attention model was incorporated to enhance the network's capability in multi-scale feature extraction and fusion. Furthermore，the Slim-Neck architecture was implemented to minimize the model's computational requirements and parameter count. The interactive interface was then designed and implemented using PyQt5. The algorithm was subsequently validated in a community located in Xi'an，Shaanxi Province. The results show that the accuracy of the model to identify fire lanes and vehicles is over 80%. Compared with the original model，the improved model reduces the number of parameters by 20.5%，the floating-point calculation by 11.3%，and the detection speed by 42.4% to 48.6 f/s. This promotes the development of intelligent detection technology for fire lane occupancy.

In order to assist earthquake rescue personnel in enhancing disaster response speed and adapting to diverse search and rescue needs，an intelligent management method for earthquake rescue equipment information based on a knowledge graph was proposed. Through the top-down knowledge graph construction method，earthquake rescue knowledge was first obtained from various information sources to serve as the basis for knowledge modeling. Next，a rule-based method was used to extract search and rescue knowledge，which was then integrated based on cosine similarity. The integrated knowledge was stored in the form of Resource Description Framework (RDF) triples. Subsequently，the open-source graph database Neo4j was employed to organize the triples into a visualized knowledge graph. Finally，a question-and-answer system was built based on the knowledge graph，allowing users to query the knowledge on the graph using natural language. The results indicate that the knowledge graph includes five categories of entities and relationships: disasters，secondary disasters，environmental factors，rescue needs，and rescue equipment. It facilitates quick matching of equipment based on rescue needs. The knowledge graph-based method can effectively manage and schedule rescue equipment information，improving the efficiency of the preparation phase of rescue operations.

In order to solve the problems of difficulty in quantifying the risk of airport runway incursion events，poor timeliness and low accuracy，and to enhance the capability of predicting runway incursion risks，a DBN model incorporating reinforcement learning for risk prediction was constructed. Firstly，causal inference theory was combined with grey relational analysis to analyze historical runway incursion events and identify the underlying risk factors. Secondly，Bayesian network(BN) theory was applied to explore the correlations among these factors and quantify these correlations using the Pearson linear correlation coefficient. This process helped in constructing a causation correlations network that effectively represented the propagation of risks associated with runway incursions. Then，the triangular fuzzy method and Hidden Markov Models (HMMs) were utilized to further refine and optimize the DBN parameter learning mechanism. Finally，the model's accuracy was validated using historical data. The results demonstrate that the proposed model's predictions of runway incursion risks closely align with the statistical values of historical data，achieving an accuracy rate of 84%，which represents a significant 10% improvement over Bayesian network predictions. Additionally，the use of mutual information to identify key nodes is found to effectively improve accuracy and discrimination compared to the degree value evaluation method.

To alleviate the contradiction between limited traffic police resources and the untimely handling of road traffic accidents，a traffic police resource optimization allocation approach was proposed based on a queuing theory model under a grid management mode of roads. Firstly，license plate recognition data obtained from the city's road network bayonet system was used to extract historical travel trajectories of vehicles and develop a similarity model between road segments. Secondly，the spectral clustering algorithm was adopted to cluster the road segments and form a set with the highest association between the segments，serving as the result of the road network division. Then，for the real-time traffic accidents within the grid，a queuing theory model was further proposed to calculate the minimum number of police officers required for each grid，along with an optimized allocation scheme for police resources. Finally，the proposed method was validated in Yinzhou District of Ningbo City. The results showed that the proposed optimization method for police allocation reduced the number of police officers by 18.18% and patrol mileage by 10.87% compared to the traditional method of dispatching police officers as soon as an accident occurs. Furthermore，the proposed method increased the accident handling response speed by 10.68%，demonstrating excellent optimization performance.