Research On Emergency Recognition Technology Based On RCNN

With the explosive growth of information,how to identify the information you want from these complex and diverse data is a matter of great concern.Sudden events are destructive and sudden.The occurrence of this incident will cause indelible harm to society and human beings.The current advanced technology is used to detect and discover the incident,so that early detection and early remediation will alleviate the problem.The harm brought by society has great research significance.For event recognition,many people identify through manual rules and shallow machine learning,but these methods may require a lot of manpower and have the disadvantage of poor portability,with the rapid development of deep learning and its application in society.The resulting good results,more and more people use deep learning to identify events.In deep learning,the two most commonly used models are the convolutional neural network model and the cyclic neural network model.In this paper,an RCNN model combining the two models will be adopted,and the model will be improved so that Better accomplish the task of emergency identification.The main work of this paper is as follows:(1)Construct a burst data set using human-machine mutual assistance.In this paper,by studying the types of emergencies,the key words contained in different types of emergencies are integrated,a trigger vocabulary is established,and the data sets are classified by trigger words to establish a preliminary emergency data set.Manual reclassification based on this data set.(2)Through the research on the working principle of the RCNN model,the RCNN model is a combination of a bidirectional cyclic neural network and a pooled layer.First,the traditional RNN model of the loop part of the model is transformed into LSTM model.Secondly,it is recognized that the Tanh activation function is often used for the RNN model and is not a fast and effective relu activation function.Because the combination of RNN and relu causes the output value to be too large,and the pooling layer has the feature of dimensionality reduction,it is bold.Change the activation function to the relu activation function.Experiments show that by improving the model,the training results of the model tend to be stable,and the appropriate model parameters can be found.The accuracy of the test reaches 83.5%.(3)By comparing the effects of different improved models,positive and negative dataset scales and word vector dimensions on the model,the optimal model is trained,the accuracy rate reaches 90%,and the recall rate and F1 value reach 92.55% and 91.26% respectively.In this paper,the structure of the bidirectional RNN+tanh activation function+pooling layer in the traditional RCNN model is improved to the structure of the bidirectional LSTM+relu activation function+pooling layer,and the results are compared with the currently used models.The model has a good recognition effect.