特征表示增强的轻量化异常序列检测方法

Lightweight anomaly sequence detection method with enhanced feature representation

边缘端的异常检测能够明显提高检测的响应速度,轻量化是深度异常检测模型在边缘端运行的解决方案,常采用模型压缩或减少参数量的方法,但参数量减少会减弱特征表示能力,影响检测准确度.为解决以上问题,提出一种层间特征传递增强的轻量化无监督异常序列检测方法,可以在减少模型参数量的同时保证检测的准确性.首先,借鉴密集卷积网络(DenseNet)的结构思想,设计特征层间连接的网络结构,增加层间的连接,加强特征传递的信息量,使提取的序列深度特征更充分;然后将深度可分离卷积应用到该网络结构中,减少参数量,实现轻量化;最后,用提取的序列特征训练支持向量描述分类器(Support Vector Data Description,SVDD),进行异常序列检测.分别在仿真数据集、Google云平台监控日志数据集和边缘端电力变压器油箱的温度数据集上进行验证,结果表明,提出的方法能准确地检测出不同变化的异常序列,与经典的轻量化网络相比,在准确率、参数量和速度上性能更好.

Anomaly detection at the edge significantly improves the response speed and lightweight is a common solution for deep anomaly detection models running at the edge. It usually adopts model compression or reduction of the number of parameters,but the reduction of the number of parameters will weaken the ability of feature representation and affect the detection accuracy. To solve the above problems,this paper proposes a lightweight unsupervised anomaly sequence detection method with enhanced feature transfer between layers,which can ensure the accuracy of detection while reducing the number of model parameters. First,this method draws on the structural idea of DenseNet to design the network structure of the connection between the feature layers,increase the connection between the layers,strengthen the information transmitted by the feature,and make the extracted sequence depth features more sufficient. Then,deep separable convolution is applied to the network structure to reduce the number of parameters and realize lightweight. Finally,the extracted sequence features are used to train the support vector data description classifier to detect abnormal sequences. It is verified on the simulation dataset,Google cloud platform monitoring log dataset and edge end power transformer oil tank temperature dataset,respectively. The results show that the proposed method can accurately detect different abnormal sequences,and has better performance in accuracy,number of parameters and speed compared with the classical lightweight network.