Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (12): 270-282.DOI: 10.3778/j.issn.1002-8331.2301-0197

• Network, Communication and Security • Previous Articles     Next Articles

Anomaly Detection Method for Industrial IoT Timing Data

XIE Wei, LU Shida, SHI Kuanzhi, WANG Honglan, GU Rongbin, HUANG Jun, LI Jing   

  1. 1.State Grid Shanghai Municipal Electric Power Company, Shanghai 200122, China
    2.Information Communication Company, State Grid Shanghai Municipal Electric Power Company, Shanghai 200072, China
    3.College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China
  • Online:2024-06-15 Published:2024-06-14

面向工业物联网时序数据的异常检测方法

谢伟,卢士达,时宽治,王虹岚,顾荣斌,黄君,李静   

  1. 1.国网上海市电力公司,上海 200122
    2.国网上海市电力公司 信息通信公司,上海 200072
    3.南京航空航天大学 计算机科学与技术学院,南京 211106

Abstract: As an important cornerstone of the fourth industrial revolution, the industrial Internet transforms isolated industrial systems into connected networks and is an important expansion direction for digital industrialization. Anomaly detection in the industrial Internet of things environment is of great significance for automated decision-making. To address the problem that most existing methods fail to effectively consider the complex unknown topological relationships between sensors and the multi-scale patterns inherent in the industrial IoT temporal data, an unsupervised anomaly detection method MSTSAD with multi-scale spatiotemporal feature fusion for industrial IoT temporal data is proposed, which firstly constructs a novel bi-directional spatiotemporal feature extraction module to sequentially captures the correlation and bidirectional dependency between multiple time series. Secondly, a multi-scale gated temporal convolutional neural network is designed to adaptively extract multi-scale temporal features of time series, and a dual affine projection is introduced to realize the cross-fusion of multi-scale temporal features and spatiotemporal features to enhance the feature extraction of the model on the original data. Finally, a variational self-encoder combined with adversarial training is proposed to amplify the reconstruction error of anomalies and enhance the anti-interference ability of the model to train data noise, which enhances the  ability of model to discriminate anomalous data. Experiments are conducted on four publicly available datasets, GPW, ECG5000, Occupancy and SWaT, in comparison with five state-of-the-art methods, and the experimental results show that F1 scores of MSTSAD are enhanced by 0.015~0.047.

Key words: industrial Internet of things, anomaly detection, multi-scale, spatio-temporal information

摘要: 作为第四次工业革命重要基石,工业互联网将孤立的工业系统转化为连接的网络,是数字产业化的重要拓展方向,工业物联网环境中的异常检测对自动化决策具有重要意义。针对现有大多数方法未能有效地考虑传感器间复杂的未知拓扑关系以及工业物联网时序数据内在的不同尺度模式导致异常检测精度有待提升的问题,提出了一种面向工业物联网时序数据的多尺度时空特征融合无监督异常检测方法MSTSAD,构建了一种新颖的双向时空特征提取模块依次捕获多个时间序列之间的相关性和双向依赖性。设计了多尺度门控时间卷积神经网络自适应地提取时间序列的多尺度时序特征,并引入双仿射实现多尺度时序特征和时空特征的交叉融合,增强模型对原始数据的特征提取。提出了结合对抗训练的变分自编码器来放大异常的重构误差并增强模型对训练数据噪声的抗干扰能力,提高了模型对异常数据的区分能力。在GPW、ECG5000、Occupancy和SWaT四个公开数据集上与五种先进方法进行了对比实验,实验结果表明,MSTSAD的F1分数提升了0.015~0.047。

关键词: 工业物联网, 异常检测, 多尺度, 时空信息