共享域特征的深度神经网络异常检测方法  

Deep Neural Networks Anomaly Detection Method Via Shared Domain Feature

作　　者：梁鹏[1] 黎绍发[2] 林智勇[1] 郝刚[1] LIANG Peng;LI Shao-fa;LIN Zhi-yong;HAO Gang(School of Computer Science, Guangdong Polytechnic Normal University, Guangzhou 510665, China;School of Computer Science and Engineer, South China University of Technology, Guangzhou 510006, China)

机构地区：[1]广东技术师范学院计算机科学学院,广州510665 [2]华南理工大学计算机科学与工程学院,广州510006

出　　处：《小型微型计算机系统》2018年第6期1234-1238,共5页Journal of Chinese Computer Systems

基　　金：广东省自然科学基金博士启动项目(2015A030310340)资助;广东省工业攻关项目(2016B010126006)资助;广东省科技项目(2017A040405058)资助

摘　　要：基于数据驱动的异常检测方法需要大量标签样本用于训练分类模型,当标签样本数量不足时,此类方法通常难以取得令人满意的预测结果.为此,提出一种结合迁移学习的深度神经网络异常检测方法,用以提高标签样本数量不足情况下的异常检测性能.本文方法的核心思想是:在深度神经网络中通过共享编码层将源域和目标域的数据映射到同一特征空间(即,共享域特征),减少源域数据与目标域数据的概率分布差异,进而从源域和目标域中选择具有共同域特征的数据实现有效的迁移学习.在真实能耗数据集上进行了验证,实验结果表明,与未使用共享域特征的传统深度神经网络方法相比,本文方法可将异常检测准确率提高2%.The data-driven anomaly detection methods need numerous labeled samples to train a classification model,and they usually fail to achieve satisfactory prediction performance when only few labeled samples are availabled. To address this issue,an anomaly detection method by combining deep neural network and transfer learning is proposed,which can be used to improve the performance of anomaly detection with insufficient labeled samples. The key idea of the proposed method is as follows. A shared coding layer is used in the deep neural network to map the data in both the source domain and target domain into the same feature space（ i. e.,with shared domain features）,lowering the discrepancy of probability distribution between the data in the source domain and those in the target domain.Then,effective transfer learning is conducted by using those data which are selected from both the two domains and share the same domain features. The proposed method has been verified on a set of real energy consumption data. And,the experimental results demonstrate that,comparing with the conventional deep neural network methods without using shared domain features,the proposed method can increase the detection accuracy by 2%.

关 键 词：深度神经网络 迁移学习 异常检测 特征空间 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]