Design And Implementation Of Multi Source Data Fusion Subsystem Based On Machine Learning Technology

Infrasound refers to sound waves with a frequency below 20Hz,which are below the audible limit of human ears.Many natural phenomena and human activities,such as tsunamis,lightning,earthquakes,chemical explosions,and missile launches,generate infrasound waves with distinct characteristics.Infrasound waves have low frequencies and slow attenuation rates in the atmosphere,allowing them to propagate over long distances,reaching several thousand kilometers.Monitoring infrasound and identifying corresponding events have significant applications in emergency response and national defense.This thesis mainly studies infrasound event recognition based on multi-source data fusion.In this thesis,multi-source data fusion technology based on data level and decision level,as well as infrasound event recognition algorithms such as random forest,gradient boosted decision trees,and multilayer perceptrons,are used.The multi-source data fusion subsystem was designed and implemented using technologies such as Spring,Netty,Flink,and Flask.First,the infrasound dataare filtered through digital filtering or spectral subtraction filtering to reduce noise,and then the infrasound signal data is processed by data fusion methods of different levels.The features of the infrasound dataare extracted and feature vectors are constructed based on different feature extraction methods.Finally,the infrasound events are classified based on random forests,gradient boosting decision trees,and multilayer perceptron algorithms.The main research work and results of the thesis are as follows:1.Research on infrasound Event Classification Algorithm Based on Multi-Source Data FusionThis thesis explores the classification of infrasound events using a multi-source data fusion approach.Time-domain features,frequencydomain features,and time-frequency domain features are extracted from the infrasound data.Nine time-domain features,including waveform factors,are extracted,while spectral density and power spectrum are used as features in the frequency domain.Limit components and information entropy features are extracted in the time-frequency domain.By designing different levels of data fusion,such as data-level Kalman filtering fusion and subband fusion,the fusion of infrasound data from multiple sources is achieved.The extracted features or fused features are then input into traditional machine learning classification models such as random forests,gradient boosting decision trees,and multilayer perceptrons for infrasound event recognition.Compared to infrasound event recognition without multi-source data fusion,the accuracy is improved by up to 6.8%.2.Backend Implementation of a Multi-Source Data Fusion SubsystemThe multi-source data fusion subsystem described in this thesis is part of a data monitoring software.It focuses on the process of infrasound event recognition and modularizes various training components to design and implement the backend of the multi-source data fusion subsystem.The subsystem is capable of identifying various infrasound events and supports both batch and streaming data inputs.The components are divided based on the recognition process as follows:(1)Data source components(HTTP source,Kafka source,etc.),(2)Data filtering components(digital filtering,spectral subtraction filtering,etc.),(3)Multi-source data fusion components(Kalman filtering fusion,subband fusion,etc.),(4)Feature extraction components(time-domain feature extraction,time-frequency domain feature extraction,etc.),and(5)Classification components(random forest model,multilayer perceptron model,etc.).The implemented subsystem utilizes the Flink computing engine to provide batch-stream integrated computation,allowing customization of model training and usage based on the characteristics of different infrasound events to improve model usability.Additionally,the Spring Boot framework is used to implement supporting functionalities such as process management and result management,enabling data management and subsequent manual analysis.