基于改进VAE的传感器异常数据检测方法研究  

Research on Detection Method of Sensor Anomaly Data Based on Improved VAE

作　　者：马海娟 杨波 杨思琪 杨鑫 吕沁锐 MA Hai-juan;YANG Bo;YANG Si-qi;YANG Xin;LYU Qin-rui(School of Electrical Engineering,Chongqing University of Science and Technology,Chongqing 401331,China)

机构地区：[1]重庆科技大学电气工程学院,重庆401331

出　　处：《计算机技术与发展》2024年第8期122-127,共6页Computer Technology and Development

基　　金：重庆市科技局自然科学(cstc2020jcyj-msxm0774)。

摘　　要：气体传感器在采样过程中受复杂工业环境影响常常产生异常时间序列数据。传统的时间序列异常检测采用模型预测方法,但没有考虑到时间序列数据的不平衡问题。因此,提出一种基于改进VAE模型的检测方法。首先,将大量正常时序数据与较少且难以标记的异常时序数据进行合并构建成一个不平衡数据集。其次,在传统VAE模型的基础上采用无监督学习方式,在异常检测分类环节引入动态阈值方法增强网络模型的自适应异常检测能力。最后,提出一种时序异常检测的组合损失函数,通过集成交叉熵损失函数和KL散度进一步提升网络参数优化性能。实验结果表明,该方法在精确率、召回率以及F1值等异常检测性能指标上,比原有的方法有所提升。该方法在传感器异常数据检测中有着较好的应用。Gas sensors often produce abnormal time series data due to complex industrial environment during sampling.The traditional anomaly detection of time series mainly adopts the model prediction method,but does not consider the imbalance of time series data.Therefore,a detection method based on improved VAE model is proposed.Firstly,a large amount of normal time series data is combined with a small amount of abnormal time series data which is difficult to label to build an unbalanced data set.Secondly,based on the traditional VAE model,unsupervised learning is adopted to enhance the adaptive anomaly detection ability of the network model by introducing dynamic threshold method in anomaly detection classification.Finally,a combined loss function for timing anomaly detection is proposed to further improve the performance of network parameter optimization by integrating cross-entropy loss function and KL divergence.The experimental results show that the proposed method is better than the original method in terms of the accuracy rate,recall rate and F1 value.It has a good application in sensor anomaly data detection.

关 键 词：传感器 时间序列 异常检测 变分自编码器 动态阈值 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]