A Study Of Acoustic Target Recognition Technology Based On Deep Learning

The effective detection and identification of low-altitude aircraft become an urgent problem to be solved by traditional radar and other monitoring methods.Acoustic target recognition has many advantages such as omnidirectional detection,passive detection and good concealment,which make it to have gained attention and application in low-altitude aircraft detection and identification.With the needs of the real-time monitoring for national defense and public safety,this thesis aims at the accuracy and robustness of continuous identification of helicopters as a typical low-altitude aircraft.Based on the in-depth summary and analysis of the generation mechanism and characteristics of helicopter acoustic signals,the key techniques such as acoustic signal preprocessing,feature extraction and identification in acoustic target identification are studied,and an acoustic target identification framework based on deep learning is proposed.To improve the performance of continuous identification of acoustic targets,a novel combined deep neural network is proposed to extract features and identify helicopters in this thesis.In the framework of the combined deep neural network,a modified convolutional neural network and a long short-term memory neural network are combined primarily in a parallel manner to optimize the representation of helicopter's acoustic characteristics and to implement helicopter type identification.The optimized feature pattern extracted by the combined deep neural network includes the current spectral characteristics and time series information hidden in the input short-term spectrum.It is designed to overcome the lack of time information of the target signal in the conventional acoustic target recognition methods.The proposed method is tested using the real helicopter acoustic signals from the field experiments.The results indicate that the proposed combined deep neural network significantly improves the recognition accuracy and the robustness of the continuous acoustic target recognition when the target is within the detection range.Aiming at the problem of background noise which degrades the accuracy and robustness of continuous identification of helicopters,this thesis also proposes a helicopter acoustic signals preprocessing framework based on convolutional neural network and couples it with the recognition framework.The framework used the convolutional neural network to map the relationship between the pre-processing features and the pre-processing targets,and determines the optimal pre-processing feature and target by experiments.It compensates for the defect that the traditional pre-processing methods destroy the harmonic structure of acoustic signals.At the same time,in order to further improve the identification accuracy,the preprocessing framework and the identification framework are coupled.The experimental results show that the proposed scheme can slightly improve the identification accuracy.