382963 Causality Detection Based on Transfer Entropy for Industrial Alarm Data Analysis

Monday, November 17, 2014: 2:05 PM
403 (Hilton Atlanta)
Fan Yang and Weijun Yu, Automation, Tsinghua University, Beijing, China

Alarms are designed for industrial processes to remind operators to react timely and solve the root cause to prevent further consequences. As industrial processes become more complex and more alarm variables are configured with interactions among them, it is more difficult for operators to monitor the whole system and identify the root cause immediately once an abnormal event occurs, in particular with an overwhelming number of alarms in a short time period, called an alarm flood. Therefore, causality detection based on alarm series becomes a very important problem for alarm management, which is beyond the correlation analysis [1,2].

Transfer entropy (TE), which has receive great attention since it was first introduced by Schreiber in 2000 [3], has been widely used in industrial processes to detect causality between different variables, as well as in other areas, such as in the neuroscience area to identify connectivity between neurons and is also proved effective [4]. Since the original definition of TE is known to be computationally intensive and can be calculated at only a single time delay with a single time bin, many have extended its definition to adapt different situations and make it useful in more areas, such as using it to distinguish direct causality from indirect causality [5]. In spite of this, most only use TE with continues time bins, such as process data series, which is computational costly.

Different from process data series, which can completely describe the system, when using alarm data that is event series (binary) in nature, there may be some information lost. However, practical methods are still needed to identify the causality between variables using alarm data, similarly to the method in neurosciences [5], because alarm data analysis is straightforward for alarm management with less computational cost. Motivated by this, we improve the method to adapt to the industrial situations with more noise. We present a framework for causality detection by TE using alarm data, which is much more computational-friendly. The proposed framework is evaluated using several numerical cases and simulated cases and the method is proved to be effective.


[1]       M Noda, F Higuchi, T Takai, H Nishitani. Event correlation analysis for alarm system rationalization. Asia-Pacific Journal of Chemical Engineering, 2011, 6: 497-502

[2]       F Yang, SL Shah, D Xiao, T Chen. Improved correlation analysis and visualization of industrial alarm data. ISA Transactions, 2012, 51(4): 499-506

[3]       T Schreiber T. Measuring information transfer. Physical Review Letters, 2000, 85(2): 461.

[4]       S Ito, ME Hansen, R Heiland, et al. Extending transfer entropy improves identification of effective connectivity in a spiking cortical network model. PLoS One, 2011, 6(11): e27431.

[5]       P Duan, F Yang, T Chen, SL Shah. Direct causality detection via the transfer entropy approach. IEEE Transactions on Control Systems Technology, 2013, 21(6): 2052-2066.

Extended Abstract: File Uploaded
See more of this Session: Advances in Data Analysis: Theory and Applications
See more of this Group/Topical: Computing and Systems Technology Division