摘要
为解决交通恐怖袭击行为识别时数据样本有限,难以满足深度神经网络(DNN)对海量训练数据的需求,提出数据增强深度学习的样本有限条件下交通恐怖袭击行为识别方法。首先,设计数据增强策略与方法,获得足够的训练样本;其次,在训练时使用原训练样本和增强训练样本计算代价函数,以抑制过拟合,提高鲁棒性;然后,堆栈多个稀疏自编码(SAE)并添加分类层,构建DNN,对增强后的训练样本进行逐层自学习和有监督反向微调,将特征提取与模式识别融为一体,准确识别袭击行为;最后,通过全球恐怖主义数据库(GTD)数据进行算例分析。结果表明:在有限交通恐怖袭击事件数据样本下,数据增强深度学习算法的特征提取能力和识别结果较对比算法得以提升,识别的平均准确率可达98.75%。
In order to solve the problem that data samples for recognition of traffic terrorist attacks are too limited to meet the demand of DNN for massive training data,based on the data-enhanced deep learning,a terrorist attack recognition approach was proposed. Firstly,data enhancement strategy and method were designed to gain enough training samples. Secondly,cost function was calculated via both original and enhanced training samples during network training process to overcome overfitting and improve robustness. Then, DNN was developed by using stacked multiple sparse auto-encoder( SAE) and classification layers. Moreover,feature extraction and pattern recognition were integrated via layer-by-layer self-learning and supervised and reversed fine-tuning of enhanced samples in order to accurately identify attack behaviors. Finally,the proposed approach was illustrated and demonstrated with attack data from global terrorism database( GTD). The results show that the feature extraction ability and recognition results of the enhanced deep learning algorithm have been improved compared with comparison algorithm with limited data samples,and its average recognition accuracy can reach as high as 98. 75%.
作者
杨黎霞
许茂增
陈仁祥
YANG Lixia;XU Maozeng;CHEN Renxiang(School of Economics and Management,Chongqing Jiaotong University,Chongqing 400074,China;School of Management,Chongqing Radio and Television University,Chongqing 400052,China;School of Mechatronics and Vehicle Engineering,Chongqing Jiaotong University,Chongqing 400074,China)
出处
《中国安全科学学报》
CAS
CSCD
北大核心
2021年第8期30-37,共8页
China Safety Science Journal
基金
国家自然科学基金资助(71471024)
重庆市社会科学规划项目(2018YBGL071)。
