基于双耳表征梅尔频谱特征无人机音频识别  

UAV Audio Recognition Based on Mel Spectrum with Binaural Representation

作　　者：罗吉庆 方虎生 朱经纬 周春华 LUO Ji-qing;FANG Hu-sheng;ZHU Jing-wei;ZHOU Chun-hua(Army Engineering University of PLA,Nanjing Jiangsu 210000,China)

机构地区：[1]陆军工程大学,江苏南京210000

出　　处：《计算机仿真》2024年第10期32-38,73,共8页Computer Simulation

摘　　要：为了提高探测模型根据无人机飞行中产生的噪声在实际场景对多类无人机识别的准确率,提出了基于双耳表征的梅尔频谱和多任务深度神经网络的入侵无人机音频识别模型。针对采用单一声道音频丢失空间信息的问题,提出双耳表征的梅尔频谱提取无人机音频特征作为数据输入,提高对无人机音频的表征能力;为提高模型训练速度和收敛性,在参照AlexNet的基础上增加BN层和Dropout层设计的深度神经网络;针对无人机音频类间相较于环境噪声异质性较小的问题,采用多任务学习的方式训练音频识别模型来提高模型泛化能力。通过手工制作的无人机声纹库上进行一系列对比实验证明了无人机音频识别模型中提出策略的优越性,相比其它无人机音频识别模型能够更好地完成识别任务。In order to improve the accuracy of the detection model in identifying multiple types of UAVs in the actual scene according to the noise generated during UAV flight,an audio recognition model for intrusion UAV based on Mel spectrum and multi-task deep neural network is designed.Aiming at the problem of losing spatial information by using single channel audio,Mel spectrum represented by binaural is proposed to extract the audio features of UAV as data input,so as to improve the representation ability of UAV audio.In order to improve the training speed and convergence of the model,the depth neural network designed by the BN layer and dropout layer is added based on Alex-Net.Aiming at the problem of small heterogeneity between UAV audio classes compared to environmental noise,multi-task learning is used to train the audio recognition model and improve the training ability.A series of comparative experiments are carried out in this paper to prove the superiority of the strategy proposed in the UAV audio recognition model.Compared with other UAV audio recognition models,it can better complete the recognition task.

关 键 词：无人机音频识别 双耳表征 梅尔频谱 多任务深度神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]