利用轻量化深度学习模型和加速度信号的枪击识别方法

Gunshot detection based on lightweight deep learning model and acceleration signal

针对不同类型的枪击加速度信号,采用深度学习的方法,提出一种新的兼顾精度和轻量化的时间序列(ENT)模型进行研究.该架构核心由注意力倒置残差模块与倒置残差模块组成,能够自动提取枪击加速度信号特征,对不同输入时间尺度更具鲁棒性.在识别精确率方面达到97.42%,超越传统枪击识别算法,在公开枪击数据集上与SVM、决策树、随机森林3种传统机器学习模型,以及FCN、ResNet、Inceptiontime、Xceptiontime等4种时间序列深度学习模型对比.实验结果表明:ENT模型更加高效,识别精确率更高.

In contrast to the conventional method,this research employs for the first time a deep learning strategy to examine various types of gunshot acceleration signals.A novel time series classification model called EfficientNettime has been proposed,capable of balancing recognition accuracy and model lightweight.The core of this architecture is composed of MBConvtime and Fused-MBConvtime modules,which can automatically extract the characteristics of gunshot acceleration signals and are more robust to different input time scales.The proposed model is verified and compared with three conventional machine learning models(SVM,Decision Tree,and Random Forest)and four other deep learning time series models(FCN,ResNet,Inceptiontime,and Xceptiontime),using a publicly available gunshot dataset.The experimental results show that the recognition precision of EfficientNettime model reaches 97.42%beyond the traditional gunshot recognition algorithm,and the model is more efficient and has higher recognition precision.