化工学报 ›› 2019, Vol. 70 ›› Issue (2): 481-486.DOI: 10.11949/j.issn.0438-1157.20181050

• 过程系统工程 • 上一篇    下一篇

基于LSTM时间序列重建的生产装置异常检测

窦珊1(),张广宇2,熊智华1()   

  1. 1. 清华大学自动化系,北京100084
    2. 浙江航天恒嘉数据科技有限公司,浙江 嘉兴 314201
  • 收稿日期:2018-09-19 修回日期:2018-10-18 出版日期:2019-02-05 发布日期:2019-02-05
  • 通讯作者: 熊智华
  • 作者简介:<named-content content-type="corresp-name">窦珊</named-content>(1994-),女,硕士研究生,<email>dous12@126.com</email>|熊智华(1971—),男,副教授,<email>zhxiong@tsinghua.edu.cn</email>

Anomaly detection of process unit based on LSTM time series reconstruction

Shan DOU1(),Guangyu ZHANG2,Zhihua XIONG1()   

  1. 1. Department of Automation, Tsinghua University, Beijing 100084, China
    2. Zhejiang Aerospace Hengjia Data Technology Co., Ltd., Jiaxing 314201, Zhejiang, China
  • Received:2018-09-19 Revised:2018-10-18 Online:2019-02-05 Published:2019-02-05
  • Contact: Zhihua XIONG

摘要:

工业生产装置通常设置传感器报警阈值进行报警,但是对处于报警阈值以下的时间序列异常难以及时捕捉。基于统计的传统检测方法在解决时间序列异常检测上存在很大挑战,因此提出基于long short term memory (LSTM)时间序列重建的方法进行生产装置的异常检测。该算法首先引入一层LSTM网络对传感器数据的时间序列进行向量表示,采用另一层LSTM网络对时间序列进行逆序重建,然后利用重建值与实际值之间的误差,通过极大似然估计方法对该段序列进行异常概率估计,最终通过学习异常报警阈值实现时间序列异常检测。采用ECG测试数据、能源数据与危险品储罐传感器数据进行了仿真实验,验证了所提方法在不同长度的数据上的有效性。

关键词: 算法, 神经网络, 参数估计, LSTM, 时间序列, 异常检测, 极大似然估计

Abstract:

Industrial production equipment usually sets sensor alarm thresholds for alarms, but it is difficult to capture time series abnormalities below the alarm thresholds. The traditional statistics based detection method has great challenges in these time series anomaly detection. In this paper, an approach to the anomaly detection of process units is proposed by using the long short term memory (LSTM) time series reconstruction. At first, an LSTM network is introduced to vectorize the time series of sensor data, and another LSTM network is utilized to reconstruct the time series in reverse sequence. Then, the errors between the reconstructed values and the actual values are used to estimate the anomaly probability by the maximum likelihood estimation method. Eventually, anormaly detection is achieved by learning the abnormal alarm thresholds. Simulation resutls on the ECG standard testing data, energy data and sensor data of the dangerous goods tank have shown the effectiveness of the proposed method on data with different lengths.

Key words: algorithm, neural networks, parameter estimation, LSTM, time series, anomaly detection, maximum likelihood estimation

中图分类号: