Data-Driven Approach to Prevent Process Incidents in Complex Systems – Identify Function Interactions and Weak Signals Leading to Hazards
Abstract
Most incidents in complex systems such as process plants follow incubation periods where weak signals exist for a long time. It is necessary to identify and resolve weak signals to prevent incidents proactively. Since “weak signal” was not precisely defined, the study first proposed its definition. Weak signals were defined as performance variabilities of functions whose interactions combine clues or signs giving rise to early prediction of a future incident. However, identifying weak signals based on individuals’ knowledge tends to be intellectually unmanageable due to complex function interactions and noise within the abundance of data in plants.
To recognize weak signals by their interaction effects, it is a prerequisite to systematically understand function interactions that lead to emerging hazards. The study developed a novel framework for process plants. The framework started from simulating function interactions in process plants through the integration of a human performance model, an equipment performance model and a chemical first-principal model based on Functional Resonance Analysis Method (FRAM). It was followed by a data-driven approach to quantify function couplings and identify the interactions leading to emerging hazards based on lift confidence intervals of association rules. The case study of a batch process showed the identified interactions could be graphically represented in FRAM with quantified function couplings, guiding people to understand how emerging hazards occur and take preventive measures.
Given abundance of data, challenges exist from selecting appropriate information for observing weak signals, evaluating their relevance, to responding. Therefore, the study developed a data-driven framework which involves FRAM and machine learning techniques to address the challenges. The case study of the same batch process showed great potentials for applying the framework in real operations. Given information of the potential weak signals that were extracted based on FRAM, probabilities of a selected hazard were predicted with high accuracy by Random Forest (RF) to indicate relevance of underlying weak signals. For interpretability, a Decision Tree (DT) that approximated the RF was developed with high fidelity, unfolding weak signals and corresponding corrective actions.
Citation
Yu, Mengxi (2021). Data-Driven Approach to Prevent Process Incidents in Complex Systems – Identify Function Interactions and Weak Signals Leading to Hazards. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /195180.