Process Disturbance Cause & Effect Analysis Using Bayesian Networks David Leng*, Nina F. Thornhill Centre for Process Systems Engineering, Department of Chemical Engineering, Imperial College London, London SW7 2AX UK * Corresponding author. Tel.: +44(0)20 7589 5111; e-mail: d.leng12@imperial.ac.uk Abstract: Process disturbances can propagate over entire plants and it can be difficult to locate their root causes from observed effects. Bayesian Networks offer a way to represent unit operations, processes and whole plants as probabilistic models which can be used to infer and rank likely causes from observed effects. This paper presents a methodology to use deterministic steady-state process models to derive Bayesian Networks based on alarm event detection. An example heat recovery network is used to illustrate the model building and inferential procedures. Keywords: Alarm, Conditional Probability, Deterministic, Disturbance, Network, Process, Root Cause 1. INTRODUCTION Modern process plants operate in demanding safety, environmental and economic conditions and comprise a range of interconnected chemical, mechanical, electrical and control operations and systems. Process disturbances affect both short-term production and long-term equipment condition, and the complexity of process interactions means that their root causes can be difficult to unravel, isolate and repair (Thambirajah, Benabbas, Bauer, & Thornhill, 2008). Within the process engineering and control communities there are a number of approaches to cause and effect analysis and disturbance diagnosis (for a review see Thornhill and Horch, 2007, and Yang, Duan, Shah and Chen, 2014). These approaches include the use of graph theory to model causal and connectivity relationships using process and instrumentation diagrams (P&IDs) and other drawing information (e.g. Maurya, Rengaswamy and Venkatasubramanian, 2004; Jiang, Patwardhan and Shah, 2009). Graph models are used in systems theory to demonstrate the existence of connectivity paths between distributed causes and effects (e.g. Deo, 1974). They can also be used in reverse to trace paths from effects to possible causes. This reversal property finds application in Bayesian Networks which model the probabilistic relationships between system variables as Directed Acyclic Graphs and are used to rank possible causes using observed effects (Murphy, 2012). Bayesian Network models have been applied to engineering condition monitoring (Marwala, 2012) and process systems (Yang, Duan, Shah and Chen, 2014), where time series data are used to infer the existence of a causal structure connecting process variables and fault states. In addition Yang et al. note that the physical explanation of Bayesian Network probabilities is not straightforward. Medjaher, Gouriveau and Zerhouni (2009) use Bond graphs to model the forward information flow around a dynamic system and generate residuals – which account for the discrepancies between the predicted state of a system and its observed state – to derive a Dynamic Bayesian network for prognostic analysis. Another research direction is based on the concept of structural equations (e.g. Lee, Christensen and Rudd, 1966) which illustrates an intuitive link between process unit operations, causality, and graphs. This paper addresses the question of how to derive Bayesian Networks which can be used for both diagnostic and prognostic analysis from process flow diagrams, deterministic models and structural equations, and gives a physically intuitive meaning to network probabilities based on observing process alarms. Section 2 presents a practical definition of a plant disturbance and explains the basic principles behind Bayesian Networks. A deterministic heat exchanger unit operation model is used to illustrate the derivation of a corresponding Bayesian unit operation. Using these Bayesian unit operations Section 3 builds an example process heat recovery system as a probabilistic model. The corresponding deterministic process model is used to simulate a process disturbance and the ensuing alarms. It is shown that the Bayesian Network can use alarm information to identify the most likely root cause. Finally section 4 presents a brief discussion of the results. Preprint of paper in IFAC Proceedings (IFAC Papers Online) Vol 48, pp 1457-1464 Presented at IFAC SafeProcess, Paris, Sept 2-4, 2015.