Detection of Causality between Process Variables Based on Industrial Alarm Data Using Transfer Entropy

Abstract

In modern industrial processes, it is easier and less expensive to configure alarms by software settings rather than by wiring, which causes the rapid growth of the number of alarms. Moreover, because there exist complex interactions, in particular the causal relationship among different parts in the process, a fault may propagate along propagation pathways once an abnormal situation occurs, which brings great difficulty to operators to identify its root cause immediately and to take proper actions correctly. Therefore, causality detection becomes a very important problem in the context of multivariate alarm analysis and design. Transfer entropy has become an effective and widely-used method to detect causality between different continuous process variables in both linear and nonlinear situations in recent years. However, such conventional methods to detect causality based on transfer entropy are computationally costly. Alternatively, using binary alarm series can be more computational-friendly and more direct because alarm data analysis is straightforward for alarm management in practice. The methodology and implementation issues are discussed in this paper. Illustrated by several case studies, including both numerical cases and simulated industrial cases, the proposed method is demonstrated to be suitable for industrial situations contaminated by noise.

Capturing Causality for Fault Diagnosis Based on Multi-Valued Alarm Series Using Transfer Entropy

“…However, this is not the case for switch variables (e.g., ON/OFF) and status variables [19]. In this work, taking the alarm series that are generated by comparing the corresponding thresholds for continuous variables as an example, we introduce the proposed methods.…”

“…In alarm systems, x is the binary alarm sequence generated by setting the alarm threshold for a process variable p. The binary alarm series x usually has two states [19][20][21]. The first one is 0 when the process variable p is in the normal state; otherwise, x takes the value 1.…”

“…Figure 5 shows the tendency of the TEs and modified MIs based on four simulation variables (A, B, C, and D that will be analyzed later) within the prediction range h = 1 to 50 (according to Yu et al [19], k = l = 1, τ = 1). It is shown that the TE and the modified MI methods obtain a maximum value at a similar h. Therefore, the prediction horizon h can be determined via the modified MI.…”

“…The four parameters need to be determined beforehand in Equation (4). Here, according to Yu et al [19], we select k = l = 1, τ = 1, and the prediction horizon h is determined by the modified MI mentioned above.…”

“…However, the physical meaning of the symbols is not straightforward. Then, Yu et al [19] proposed transfer entropy to detect causality between process variables based on alarm data, which can replace process data with binary alarm series. Its main contribution is a new application of TE that uses binary alarm data for causality detection, while a lot of information contained in the process data is lost.…”