针对野外环境下单通道车辆声信号受风噪影响严重、分类性能较低的问题，提出一种基于声阵列4通道同步采集信号的一维卷积神经网络模型（FVC-CNN）。该模型借鉴注意力机制加权平均的思想对Inception网络结构进行改进，作为输入层有针对性地提取4通道声信号多个不同时间尺度的特征，抑制噪声干扰，再根据不同车辆声信号特征分布特点，分别训练3个特征提取网络SWNet、LWNet和TNet来提取相应车辆的特征，最后对提取的特征进行多分支多维度的融合以供分类。在相同数据集上进行验证，实验结果表明，FVC-CNN模型总识别率可达94.22%，相较于传统方法识别率提高14.08%，取得了较好的分类效果。
Aiming at the problem that single channel vehicle acoustic signal is seriously affected by wind noise and has low classification performance, a one-dimensional convolutional neural network model FVC-CNN (convolutional neural network for field vehicle classification, FVC-CNN) based on four channel synchronous acquisition signal of acoustic array is proposed in this paper. The model uses the idea of weighted average of attention mechanism to improve the structure of Inception network. As the input layer, it extracts the features of four channel acoustic signals with different time scales to suppress noise interference. According to the distribution characteristics of different vehicle acoustic signals, three feature extraction networks, SWNet, LWNet, and TNet, are trained to extract the characteristics of the corresponding vehicle, finally, the extracted features are fused with multi branches and multi dimensions for classification. Verified on the same data set, the experimental results show that the total recognition rate of FVC-CNN model can reach 94.22%, which is 14.08% higher than the traditional method, and the classification effect is better.
[1] 罗向龙,牛国宏,吴潜蛟,等.基于经验模态分解和支持向量机的车型声频识别[J]. 应用声学, 2010, 29(3):178-183.
[2] Kandpal M, Kakar V K, Verma G. Classification of ground vehicles using acoustic signal processing and neural network classifier[C]//2013 INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND COMMUNICATION (ICSC). Noida, India:IEEE, 2013:512-518.
[3] Yang S S, Kim Y G, Choi H. Vehicle identification using discrete spectrums in wireless sensor networks[J]. Journal of Networks, 2008, 3(4):51-63.
[4] 孙铭堃,梁令羽,汪涵,等.基于级联卷积网络的面部关键点定位算法[J].中国科学院大学学报,2020,37(4):562-569.
[5] Zhang H M, McLoughlin I, Song Y. Robust sound event recognition using convolutional neural networks[C]//2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). South Brisbane, QLD, Australia:IEEE, 2015:559-563.
[6] 刘莹. 基于深度学习的轨迹预测[D].成都:电子科技大学,2019.
[7] 赖策. 卷积神经网络中的激活函数分析[J]. 科学技术创新, 2019(33):35-36.
[8] Kalayeh M M, Shah M. Training faster by separating modes of Variation in batch-normalized models[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(6):1483-1500.
[9] 黄琦. 智能传感器侦察网络中的地面目标识别算法研究[D]. 合肥:中国科学技术大学, 2006.
[10] Tiete J, Domínguez F, da Silva B, et al. SoundCompass:a distributed MEMS microphone array-based sensor for sound source localization[J]. Sensors (Basel), 2014, 14(2):1918-1949.
[11] 黄景昌. 野外监控传感网中运动目标声音信号的探测与识别研究[D]. 北京:中国科学院大学,2015.
[12] Szegedy C, Liu W, Jia Y Q, et al. Going deeper with convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA, USA:IEEE,2015:1-9.
[13] 李光海. 基于聚焦算法的机械振动噪声声源定位的研究和应用[D]. 武汉:武汉理工大学,2015.
[14] 胡利萍. 野外监控传感网声探测器的环境适应性研究[D]. 北京:中国科学院研究生院,2011.
[15] 杨旭. 基于盲源分离的声信号降噪算法研究[D]. 北京:中国科学院大学,2012.
[16] Shi W L, Fan X H.Research on armored vehicle classification based on MFCC and SVM[C]//20173rd IEEE International Conference on Computer and Communications (ICCC).Chengdu, China:IEEE,2017:1938-1941.
[17] Wei J S, Lv J C, Xie C Z.A new sparse representation classifier (SRC) based on probability judgement rule[C]//2016 International Conference on Information System and Artificial Intelligence (ISAI).Hong Kong, China:IEEE, 2016:338-342.
[18] 范裕莹,李成娟,易强,等.基于改进TCN网络的野外运动目标分类[J/OL].计算机工程:(2020-09-10)[2021-03-29].https://doi.org/10.19678/j.issn.1000-3428.0058750.
