结合LSTM与CNN的野外车辆声信号分类

针对野外环境下微声传感器采集的小型轮式车、大型轮式车和履带车3种车辆声信号受风噪影响严重、分类性能较低的问题，提出了一种长短时记忆网络（LSTM）与多尺度、多层次特征融合卷积神经网络（CNN）相结合的分类算法——野外车辆识别算法（FVNet）。该算法先采用一层LSTM网络提取声信号的时序特征，充分利用声信号的长时依赖关系；再用CNN并行提取多尺度特征，避免网络加深过程中特征的流失；引入通道注意力机制进行多尺度和多层次特征融合，增强多尺度、多层次关键特征信息；最后在相同数据集上进行验证。实验结果表明，FVNet算法对3种车辆的总识别率可达94.95%，与传统方法相比，其总识别率提高了14.61%，取得了较好的分类效果。

Acoustic Signal Classification of Field Vehicles Based on Combination of LSTM and CNN

In order to solve the problem of low classification performance and serious influence of wind noise on the acoustic signals of small wheeled vehicles, large wheeled vehicles and tracked vehicles collected by micro acoustic sensors in the field, a classification algorithm network for field vehicle classification (FVNet) based on the combination of long short term memory modal(LSTM) and the multi scale and multi level feature fusion convolutional neural network(CNN) is proposed. Firstly, an one layer LSTM network is used to extract the temporal features of the acoustic signal, which makes full use of the long term dependence of the acoustic signal. Then the CNN is used to extract multi scale features in parallel to avoid the loss of features in the process of network deepening. The channel attention mechanism is introduced to fuse multi scale and multi level features to enhance the multi scale and multi level key feature information. Finally, it is verified on the same data set. The experimental results show that the total recognition rate of FVNet algorithm for three types of vehicles can reach 94.95%, which is 14.61% higher than that of traditional methods, and achieves better classification effect.