Reliable dynamic causality analysis for efficient prescriptive maintenance of degraded industrial equipment

Developing efficient prescriptive maintenance strategies is essential in mitigating the performance degradation of equipment in energy-intensive process industries. This paper proposes an innovative approach that integrates data-driven causality analysis and reinforcement learning (RL) to reduce/slow-down the equipment degradation. The degradation is first modeled by constructing representative health indicators (HIs) using deep learning autoencoders. The HIs are then exploited using clustering and interpretable machine learning techniques to identify the degradation's root causes. Afterward, a dynamic causal model is discovered in the form of a Petri net (PN) using process mining techniques. The causal model incorporates the temporal information and sequential relationships between the identified root causes. Finally, an RL agent is integrated with the PN model to recommend the optimal sequence of events that diminishes the performance degradation rate. The proposed approach is tested successfully on a complex case study of a black liquor concentrator in a Kraft pulp mill that is subjected to a decline in operational performance and capacity due to the fouling degradation phenomenon. The obtained results show that the fouling rate was minimized, leading to an approximate 30 % saving in maintenance costs and a reduction of around 4.8 kt CO2/year in greenhouse gas emissions.