| Infrasound refers to the sound whose frequency is lower than the range of human hearing.Many man-made events such as mining and blasting,nuclear weapon test and rocket launch will produce strong infrasound.In addition,many natural events such as aurora,volcanic eruption and earthquake will also produce infrasound.Infrasound widely exists in nature,and its low-frequency characteristics enable it to propagate over a long distance.For these reasons,CTBTO member states have established a series of infrasound monitoring arrays,which can monitor nuclear tests,volcanic eruptions,tsunamis and other events for disaster prevention and mitigation.This thesis mainly studies the recognition algorithm of natural events based on infrasound.Natural events refer to the objective facts directly caused by objective phenomena rather than human activities.Effective identification of natural events in infrasound events can warn some natural disasters;At the same time,it is helpful to separate the man-made events in infrasound events and improve the recognition rate of man-made events.This thesis studies the infrasound event recognition algorithm using Support Vector Machine,Dynamic Time Warping,Recurrent Neural Network,Convolutional Neural Network and other technologies,and designs and implements the infrasound monitoring system using Spring,Netty,message queue and other technologies.Firstly,using the Support Vector Machine model,some simple waveform features are established through feature engineering,and the features are input into the model for classification.Secondly,try Dynamic Time Warping to identify infrasound events.The core of the algorithm is to compare the differences of waveforms and ignore the original acoustic characteristics of infrasound.Based on the above experimental results,it is concluded that the ability of network mining deep-seated features needs to be improved.Thirdly,try the deep learning model,and use the Long ShortTerm Memory neural network,Fully Convolutional Networks and other models to identify infrasound events.The deep learning model can dig out deeper features that are difficult to design manually.Finally,aiming at the problem that the above model can not be effectively used for classification of ultra long infrasound events,based on the above model,the network structure is modified and fused,and a new model structure(SWFG)is proposed,which comprehensively compares the accuracy and recognition speed of all the above algorithms.The main research work and achievements of this thesis are as follows:1.Basic research on infrasound event classification algorithmIn this thesis,the conventional time-domain features of infrasound events are extracted,including rectified average,standard deviation,root mean square,kurtosis,skewness factor and waveform factor.The above features are input into Support Vector Machine to realize the classification of infrasound events.Then try to classify using the Dynamic Time Warping algorithm based on waveform difference,and compare the above results to lay a foundation for subsequent research.2.Research on infrasound event classification algorithm based on deep.learningAfter trying the Long Short-Term Memory network and Gated Recurrent Unit model,and achieving certain results.In order to further improve the ability of the network to mine deep-seated information,Convolutional Neural Network is used.The structure is modified to onedimensional convolution so that it can process time series data.Based on the above,the Fully Convolutional Networks is used to classify infrasound events for the first time,and a good classification effect is achieved.3.A fusion algorithm model for ultra long infrasound events is proposedSome long-lasting events will be encountered in the identification of infrasound events,and the above algorithm model can not effectively classify the events.For this event,a hybrid algorithm model(SWFG)integrating sliding window,Fully Convolutional Networks and Gated Recurrent Unit is proposed,and the effectiveness of the model is verified by the actual data of Tonga volcanic eruption.4 The infrasound monitoring data acquisition and analysis subsystem is designed and implementedAs a part of a large-scale monitoring system,this thesis mainly introduces the design and implementation of data acquisition and analysis subsystem.In the demand analysis stage,clarify the functions of the subsystem,including data acquisition,data cleaning,data analysis and equipment discovery,and draw use case diagrams for the above functions.In the outline design stage,the topology diagram and deployment diagram of the system are drawn to clarify the architecture of the system as a whole.In the specific implementation stage,activity diagrams are drawn to illustrate some complex logic.Finally,through the test case and the actual operation effect diagram,it shows that the subsystem realizes the above requirements. |
