Method Research And Prototype Implementation Of Infrasound Data Generation Based On GANs

Some abnormal events,such as chemical explosions in human activities,typhoons and thunderstorms in natural activities,will bring certain harm to the production and living of human beings.So it is of practical significance to identify these events.The occurrence of abnormal events will produce infrasonic waves,and the generating mechanism of different events is different,so infrasonic waves can be used to identify abnormal events.Infrasound data and machine learning algorithms are used to identify infrasound events in the existing infrasound recognition system for abnormal events.However,the problem of too small sample size in training set appears in the research,which leads to poor generalization ability of the algorithmic models.Therefore,in this thesis,the infrasound data generation technology based on GANs is proposed to solve the problem of few samples and provide sufficient data for the system.Moreover,it will also produce certain reference value for the generation of time series.The original data of this project includes abnormal event(including lightning,chemical explosion and typhoon)data set and no-event(i.e.micro-pressure fluctuation)data set.First of all,in terms of data set processing,different data preprocessing methods are adopted for different event types.The time length of data is unified and then the data is standardized.Next,the models of typical GAN,DCGAN and WGAN are constructed as comparative experiments to generate infrasound signals.Combined with Euclidian Distance and DTW value of sequence similarity,and signal feature analysis used for verification,the experimental results show that the generation effect of DCGAN is the best,and that of WGAN is the second.Then,LSTM and GRU networks are used as the discriminators of DCGAN and WGAN respectively to build new models,and the same evaluation process is used to evaluate.The experiments show that the generation effect of infrasonic data based on typical DCGAN is the best.Finally,in this thesis,the expanded data of DCGAN,SMOTE and Bootstrap are mixed with the original data respectively,and then the LSTM classification model is built and trained.The classification accuracy of the data generated by each method is compared,which proves the validity and availability of the data generated by DCGAN.A set of infrasound data generation management prototype system is also completed,which realizes the generation management of infrasound data and provides some management functions that can be extended to the abnormal event infrasound identification system.The system mainly includes four modules:data generation management,monitoring site management,department management and user management.