Research On Multi-Sound Source Intelligent Recognition Method Based On MFF-ResNet

With the rapid changes of battlefield environment and situation in modern war,the concept and means of warfare operations have contained continuous development.Especially in close-range urban combats and anti-terrorism operations,the improvement of individual combat ability is of great significance to the development of the whole battlefield situation.Therefore,the equipment technology of individual combat is in urgent need of breakthrough development.Effective sound source identification in complex combat scenarios,such as gunfire,tank sound,drone sound,armored vehicle engine noise,etc.,is vitally important to the quick identification of the enemy combat weapons,as well as the evaluation of their combat power.In such situation,various sound sources were seriously aliased,and the characteristics of the sound source targets became less obvious.Further,it led to ineffective identification about the sound source targets.To solve the above problems,this dissertation carries out research on multi-source intelligent identification method based on MFF-ResNet,centering on the task of sound source target identification in complex environment.Firstly,the sound source signal studied in this paper is the gunshot signal with low signal-to-noise ratio under strong reverberation and interference.According to sound signal characteristics and sound source identification requirements,this paper studies the theory of acoustic sensor array layout and succeeds in making a series of improvements.That is,the sound array model has been optimized,the microphone probe selection has been completed,the signal conditioning circuit and signal acquisition and storage system has been designed,and the hardware system has been built.Secondly,a neural network model is designed to enhance sound field signal of GAN network,and the characteristic parameters of sound spectrum under strong reverberation are extracted.Thirdly,this paper has proposed a sound source target identification(MFF-ResNet)method with two-level multi-head and multi-attention fusion,based on deep learning,designing models of sound source feature extraction,convolutional network structure,feature fusion layer,as well as the feature classifier.In this paper,ten kinds of gunshot sound sources were simulated and verified on the NIJ Grant 2016-DN-BX-0183 data set.The results show that the identification accuracy of the network on the data set reaches 95.36%.Finally,gunshot sound source target identification system as a whole underwent the outfield experimental research.The identification accuracy of the test reached 92.10%.The findings show that the source target identification system proposed in this paper can realize the acquisition and accurate classification of the source signal,along with the ability to resist reverberation and noise.