Research On Emergency Evacuation Risk Assessment Method Based On Deep Learning

With the continuous occurrence of safety accidents in public buildings,emergency evacuation of personnel has become an important means of emergency response.However,there are also risks in the evacuation process.It is of great practical significance to determine the risk situation of the evacuation process,which can provide early warning and reference for improving the safety of the evacuation process,and provide auxiliary support for evacuation decisions.At present,due to the complexity of the evacuation process and the lack of real evacuation data,the research on evacuation risk assessment is still limited,while traditional risk assessment methods are highly subjective and difficult to meet the requirements of the emergency evacuation.With the development of artificial intelligence,deep learning can explore the inherent relationship of complex evacuation systems,avoid subjectivity,and easily meet timeliness requirements,which will become an effective method to achieve emergency evacuation risk assessment.In summary,this study applied the deep learning method to the field of risk assessment and innovatively proposed a method for evacuation risk assessment based on deep learning prediction models.Based on the analysis of evacuation risk factors,the convolutional neural network(CNN)depth prediction model was proposed,and the training and testing data of the prediction model were obtained by integrating multi-agent evacuation simulation experiments and generative adversarial networks.In addition,taking the university gymnasium as an example,the realization process and feasibility of the evaluation method were illustrated by constructing the evacuation physical environment,personnel,and fire emergency scenarios.The main research contents of this paper include the following parts:1.Proposing an evacuation risk assessment framework based on deep learning.This paper sorts out and summarizes the research status of emergency evacuation and deep learning methods.Emergency evacuation is a kind of complex system with people as the core.Traditional risk assessment methods have limitations such as strong subjective factors and poor timeliness in the application of such problems.Deep learning has the advantages of strong feature extraction ability and easy to meet timeliness in dealing with complex problems.This study proposed a research framework for emergency evacuation risk assessment using deep learning methods,which mainly included establishing risk assessment models based on convolutional neural networks,obtaining training data through evacuation simulation experiments,implementing data enhancement based on generative adversarial networks,and risk assessment of coupled fire disasters.2.Establishing risk assessment models based on convolutional neural network.Convolutional neural network(CNN)is an important deep learning method for inference and prediction through feature extraction and data training.In this study,four representative network structures,Le Net,Alex Net,VGG,and Res Net,were selected for the construction of evacuation risk assessment models.This study analyzed many risk factors in emergency evacuation,selected the important factors as evaluation factors,and determined the input and output parameters of the risk assessment model.On this basis,the Py Torch framework was used to build risk assessment prediction models.3.Establishing evacuation simulation models to obtain training data for the assessment model.For deep prediction models,it is crucial to have sufficient data samples,but real evacuation data are very scarce.Therefore,this study adopted the multi-agent modeling and simulation method and conducted simulation experiments to obtain the training data required for the risk assessment model by using Any Logic simulation platform.In addition,the pedestrian movement speed in the real video was analyzed by the video analysis software Tracker,which provided a real data basis for the simulation model.The orthogonal experimental design method was used to design the relevant data of the simulation experiment such as building environment factors and agent parameters,which could improve the efficiency of the simulation experiment and the comprehensiveness of the data set.4.Data enhancement based on generative adversarial networks.Although a certain number of data samples can be obtained through simulation experiments,the amount of data is still limited.Generative adversarial networks(GAN)have good data generation capabilities and can be used for data enhancement to alleviate the problem of fewer experimental data samples.Therefore,given this situation,WGAN(Wasserstein GAN)model in GAN was used to perform data enhancement on simulation experimental data and expand the number of data samples,which could guarantee the training accuracy of the deep prediction model.5.Risk assessment models training and testing.After obtaining the training data of risk assessment models through the simulation experiments and data enhancement methods,the CNN-based risk assessment prediction models were trained and tested by applying the two data sets before and after data enhancement,which was verified the feasibility of the method proposed in this paper.The results showed that the training accuracy of the four risk assessment models based on Le Net,Alex Net,VGG,and Res Net were significantly improved after data enhancement,which ensured the effectiveness of model evaluation.6.Evacuation risk assessment for coupled fire emergency scenarios.Based on the deep learning evacuation risk assessment method,under the background of fire disaster,this part took a university gymnasium as an example and integrated the methods of fire numerical simulation and multi-agent simulation.The fire simulation software CFAST and the multi-agent simulation platform Any Logic were used to establish an interactive relationship to achieve the dynamic impact of fire smoke on personnel evacuation.On this basis,the experimental design was carried out by using the uniform design method.Through the emergency evacuation simulation experiment under the fire situation,the emergency evacuation data under the fire disaster situation was obtained.Finally,WGAN data enhancement and CNN prediction training were used to achieve evacuation risk assessment in fire emergency scenarios.In summary,this study proposed a new emergency evacuation risk assessment method based on deep learning methods,multi-agent simulation,and data enhancement methods,which provided new methods and ideas for research in the field of risk assessment.