Dissertation Defence: Decision Support Methods Based on Alarm Management in Industrial Automation Systems
December 12 at 9:00 am - 1:00 pm
Aliakbar (Ali) Davoodi Beni, supervised by Dr. Ahmad Al-Dabbagh, will defend their dissertation titled “Decision Support Methods Based on Alarm Management in Industrial Automation Systems” in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Electrical Engineering.
An abstract for Aliakbar (Ali) Davoodi Beni’s dissertation is included below.
Examinations are open to all members of the campus community as well as the general public. Please email ahmad.aldabbagh@ubc.ca to receive the Zoom link for this exam.
Abstract
Alarm systems are a core component of industrial automation, alerting human operators to abnormal process conditions. Because process units are interconnected and alarm configurations can be suboptimal, a single fault may trigger multiple alarms and, in extreme cases, alarm floods. Excessive alarms can overwhelm human operators and delay corrective actions. To address this challenge, in this thesis, different alarm management methods are proposed, namely for correlated alarm grouping, alarm flood analysis, operator action prediction, and alarm prediction, with the goal of improving decision support for operators.
First, a three-stage alarm grouping framework is proposed, which combines a time-augmented alarm embedding with a density-based hierarchical clustering method to produce correlated alarm groups. Specifically, in the first stage, the embedding fuses timing information from alarm logs (the immediate output of industrial automation systems), making the proposed word embedding robust to variations in the order of alarms. In the second stage, the proposed clustering method reduces computational cost relative to existing methods. Finally, in the third stage, the proposed visualization method presents the grouped alarms to operators, and an online implementation is provided for practical deployment.
Second, a graph-based framework for alarm flood analysis and classification for both supervised and unsupervised cases is proposed. The framework adaptively updates graph models for alarm flood sequences to capture gradual changes in alarm flood patterns, making it more robust than existing methods. Moreover, it can identify previously unseen flood types without prior process knowledge. This overcomes the shortcomings of existing methods, especially when knowledge and information from an industrial process are limited.
Third, a two-stage alarm prediction framework is proposed, which includes two probabilistic models. Stage I uses a multivariate inference-based method to estimate transition probabilities and propose the most probable future alarms. Stage II employs a statistical model to estimate annunciation timing and confidence for those alarms. In addition, an online/adaptive parameter updating method is proposed to make the prediction model robust against changes in alarm sequences/patterns.
Finally, an operator action prediction method is proposed, which is based on a probabilistic formulation that directly predicts corrective actions from ongoing alarm sequences. The probabilistic model represents dependencies among alarms observed before an action, and parameters are learned by maximum likelihood.
All methods are evaluated on the Tennessee Eastman process benchmark. Results demonstrate higher accuracy and robustness than closely related existing approaches, while requiring limited user intervention and providing outputs that are actionable for operations.
