Effective alarm management to improve safety using a data-driven approach based on Bayesian networks

Abstract

Proper fault management is important for safe operations in the process industry. Since a critical alarm can have multiple potential fault modes, the diagnosis of a fault may not serve its intended purpose. Incorrect diagnosis of a fault in a critical system may cause severe consequences. It is, therefore, important to detect and diagnose faults and characterize their priority considering the risk of potentially unsafe scenarios. Risk-based fault diagnosis and prioritization help minimize ignorance of the critical fault and, most importantly, enhance safety. This study presents a data-driven methodology applying Bayesian network (BN) to decide the management priority of faults from a risk perspective. The parameters of the process model are learned from operational data to increase model feasibility and reduce input uncertainty. The challenge of incomplete data is handled using the expectation-maximization (EM) algorithm. The methodology and model are explained using an accidental oil leak in a processing facility. The assessment results reveal that compared to the previous methods, the proposed approach can avoid the oversight and neglect of a critical fault such as an oil leak, and thus prevent large amounts of oil loss.