Wasserstein自编码器异常检测模型  

Wasserstein autoencoder anomaly detection models

作　　者：王星 霍纬纲[1] WANG Xing;HUO Wei-gang(College of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学计算机科学与技术学院,天津300300

出　　处：《计算机工程与设计》2020年第11期3249-3254,共6页Computer Engineering and Design

基　　金：国家自然科学基金委员会-中国民用航空局民航联合研究基金项目(U1633110);国家自然科学基金项目(61301245);中央高校基本科研业务费项目中国民航大学专项基金项目(3122019190)。

摘　　要：针对现有半监督深度生成异常检测模型对复杂真实数据分布学习能力不足及模型训练困难的问题,提出基于改进Wasserstein自编码器(Wasserstein autoencoder,WAE)的异常检测模型WAE-AD。使用能够稳定训练的自编码器网络结构,利用Wasserstein距离对模型拟合分布与待检测数据真实分布之间的距离进行度量,学习更加复杂的高维数据分布。使用正常数据构成的训练集训练模型收敛,根据待检测数据在训练好的模型中的异常得分进行异常判定。实验结果表明,WAE-AD模型的精确率、召回率、F1值3项异常检测性能指标较现有半监督深度生成异常检测模型均有明显提高。To solve the problem of insufficient learning ability of the semi-supervised deep generative anomaly detection model for complex real data distribution and model training difficulty,an anomaly detection model WAE-AD based on improved Wasserstein autoencoder was proposed.Autoencoder network structure that could be stably trained was used,and Wasserstein distance was used to measure the distance between the model-fitting distribution and the true distribution of the data to be detected,so that more complex high-dimensional data distribution was learned better.The model was trained to be convergent with normal dataset,and the anomaly was determined according to the abnormality score of the data to be detected in the trained model.Experimental results show that the accuracy,recall rate and F1 value of the WAE-AD model are significantly improved compared with the existing semi-supervised deep generative anomaly detection model.

关 键 词：异常检测 Wasserstein距离 自编码器 深度生成模型 半监督学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]