In order to discover fire accidents in chemical industrial parks in time and reduce accident losses，this study used CNN to establish a real-time fire detection system for chemical industrial parks. Based on CNN，the YOLOv5 algorithm was used to calculate chemical industrial park fire data sets and ordinary fire data sets. The loss value，recall rate，precision and mean average precision of the two data sets were compared. Among them，the loss value and recall rate of the chemical industrial park fire data set are slightly lower，but the precision and mean average precision were higher than that of an ordinary fire data set，which shows the feasibility of detecting fire. In addition，based on fire detection results，this study further designed the flame image recognition software system for the chemical industry park with the help of the PyQt5 program framework，realized the application of fire image and video recognition in the chemical park，and expanded the application scope of the method. The results show that the YOLOv5 target detection algorithm based on convolutional neural network can detect fires in chemical industrial parks in real-time. This detection method is portable，and the results are reliable，which can help improving the safety management level of the chemical industrial park.

To evaluate the influence of upper arch deformation on the safety of train operation on ballastless track inside high-speed railway tunnels，the upper arch deformation characteristics of different ballastless tracks were obtained based on measured data. Subsequently，a dynamic calculation model of the coupled system of uplift arch filler layer-ballastless track-vehicle was established. The impact of different upper arch deformation characteristics on the dynamic response of the vehicle was determined by using the calculation model，thereby assessing the safety of high-speed train operation. The results show that the measured wavelength is mainly distributed in 3 to 30 m and the amplitude is mainly distributed in 0 to 8 mm，among which the upper arch deformation is the most prominent with a wavelength of 10 m and an amplitude of 2 mm. When the upper arch wavelength is less than 6 m，the peak value of wheel-rail force is larger，and when the upper arch wavelength is about 12 m，the peak value of vertical acceleration of the vehicle is larger. The upper arch，with a wavelength of about 6 m，has a significant impact on high-speed driving safety and has a certain derailment risk，which should be paid more attention.

To fill the research gap，a HVRP for hazardous materials transportation was proposed in this study，aiming at optimizing the three objectives as total risk，total cost and total carbon emissions. Firstly，in the model construction stage，the measurement methods of total cost and total risk were improved，including establishing a loading-dependent risk quantification method for explosion scenarios，and designing a specific soft time window for penalty cost calculation，which could preferentially reduce the waiting time of vehicles with large loads at customers. Then，to better solve the above model，the NSGA-II was improved in two aspects. A hybrid crossover method with new crossover operators was designed to improve the global search efficiency，and a two-stage variable neighborhood search (VNS) algorithm was adopted to improve the local search ability. Finally，a numerical example was given to verify the effectiveness of the model and algorithm. The results indicate that，compared with the original NSGA-II，the convergence curve of the improved algorithm decreases faster，and the average values of the three optimization objectives of total cost，total risk and total carbon emissions are further reduced by 3.36%，12.16% and 6.96% respectively. In a fleet with limited number of vehicles，the carrier can have different degrees of influence on each target by choosing different vehicle types.

In order to explore the spontaneous combustion mechanism of high-voltage power cables under the effect of core heating when it was overloaded or short circuited，heating rods was used to simulate the internal heat source for experimental research. The influence of the power of the internal heat source and the cross-sectional area of the inner core of the cable on the key parameters of the spontaneous combustion characteristics，such as cable temperature rise and ignition time，were analyzed. It was found the whole process of cable from heating to ignition to combustion can be divided into four stages: initial heating，pyrolysis gas escaping，ignition，combustion and extinguishment. The results show that the dripping of insulating layer melt leads to the change of flame shape. The heating rate of the surface temperature of each structural layer of the cable increases with the increase of internal heat source power and the decrease of the distance to the surface of the internal heat source. The increase in the power weakens the influence of the cross-sectional area of the inner core of the cable on the heating rate of each layer. The ignition time of the cable is linearly decreasing with the power of the internal heat source.

In order to investigate the effect of smoke evacuation patterns under ambient winds on the safe evacuation of people，taking a terminal building as the subject of the study，the multi-floor area was selected as the data acquisition range，33 control conditions were divided by controlling the smoke exhaust window，and fire dynamics simulator (FDS) software was used to obtain the smoke exhaust effect in the building under different conditions. Further construct wind direction intervals with north winds shifting to east winds via northeast winds to obtain the curve law of wind direction and time to reach hazards，and 0-90° as the high side window opening angle interval to obtain the suitable window opening range，according to the environmental wind to determine and optimize the safe evacuation. The results show that for the smoke exhaust effect in the building，if the high side windows are fully opened，it is conducive to smoke exhaust under no wind and low wind speed. If the wind speed is high，the wind direction perpendicular to the floor alignment is more effective for smoke evacuation. In the wind direction range，the wind direction of 0-70° will promote the smoke exhaust of the high side window，and the wind direction of 30 and 60° is the best，and the wind direction along the floor trend will have smoke backflow. Within the high side window opening angle，a 5-30° down-hung window or full opening window is appropriate. Combined with the time to reach hazards of the wind direction，the evacuation of floor personnel is optimized，and the evacuation time after adjustment is reduced by 10% compared with the unoptimized time.

In order to optimize the anti-explosion design of an underground unventilated kitchen，the finite element model of a house with an underground unventilated kitchen was established in FLACS software. Based on this model，the effect of gas cloud size，ignition position，shape of obstacle，position，and size of obstacle on gas explosion pressure was investigated. The extent of explosion damage to the unventilated kitchen structure was obtained according to simulation results. The results show that when the gas cloud size increases in the unventilated kitchen，the explosion damage to the building is more severe. The maximum pressure peak values are 41.9，19.5，and 3.25 kPa when ignition is in the kitchen，ventilation shaft，and living room. When an obstacle exists，a greater amount of pressure is generated. The peak pressure is much higher when the obstacle cross-section shape is square than when it is circular and rectangular. The closer the obstacle is to the ignition position，the more intense the pressure peak. With the increase in the obstacle section size，the peak pressure increases continuously. The greatest rise in peak pressure is near the elevator. Shock waves from gas explosions in unventilated kitchens have a greater impact on the floor space than in traditional kitchens.

In order to establish a scientific，accurate and normative urban safety risk management system，this paper makes some proposals based on the characteristics of urban safety risk. First，the urban safety risk was classified into four types according to the causal chain model which was established based on the definition and connotation of risk. Then，the characteristics of these four types of risk were analyzed considering the relevant static and dynamic urban characteristics. Last，the general and specific urban safety risk management proposals were made considering the current practice. This research shows that the characteristics of urban safety risk include accumulation of conventional risk，emergency of emerging risk，intensification of complex risk，and frequent occurrence of extreme risk. Some measures should be implemented to enhance the urban safety risk management system. The general framework of urban safety risk management should be established，and the specific risk assessment process and risk treatment measures should be implemented. Further，they should be integrated into the regular emergency management system of a city.

In order to effectively prevent and control fire accidents of hazardous chemical tanks during transportation，it is necessary to identify the type of tank fire quickly and correctly. The statistical data of domestic hazardous chemicals tanker fire accidents from 2013 to 2020 was analyzed with Python software. According to the characteristics of fire accidents，a regression model with multiple logistic regression theory was established. The three common fire types of hazardous chemicals tankers，namely，jet fire，pool fire and BLEVE(Boiled Liquid Evaporate Vapor Explosion)fire，were regarded as the categories of explained variables. Through the analysis of accident data，a total of 15 fire factors in five aspects，namely，space and road characteristics，time and meteorological characteristics，personnel factors，vehicle and equipment factors，and hazardous chemicals factors，were determined. A multiple logistic regression equation of fire probability of hazardous chemicals tankers was constructed，and the explicitness and accuracy of the equation were tested. The determined model was proved by examples. The probability of three types of fire in hazardous chemicals tankers was obtained. The accident prevention measures were proposed. The results show that it is the most likely a pool fire for a liquefied natural gas tanker，with a probability of 0.55. The second one is jet fire，and the last one is the boiling liquid expanding vapor explosion fire. To reduce the probability of pool fires，measures such as improving the mechanical strength of the container and reducing the collision strength could be adopted to prevent leakage holes of large size in the container. Measures such as slowing down the leakage rate，controlling the leakage range and efficient scientific fire extinguishing are adopted to decrease sequences of tank fire.

In order to help fire investigators determine the location of the fire origin，improve the chain of evidence in accident investigations and identify the cause of the fire quickly and accurately，the researches on fire origin determination based on fire traces were reviewed in present paper. First，fire traces were classified，including burn marks，smoke marks，collapse marks and electrical wiring marks，with emphasis on soot deposition traces. Then，multiple fire origin determination methods at home and abroad were reviewed and classified into three categories: determining the fire origin directly using experience，determining the fire origin using numerical reconstruction techniques，and determining the fire origin using machine learning algorithms. The advantages and limitations of each method were analyzed respectively. Finally，the future research tendency of fire origin determination technology was prospected. The results show that numerical simulation of soot deposition traces combined with machine learning for fire origin determination has a good perspective for application.