PTQ Q4 2026 Issue

AI/ML-driven process upset analytics: C 3 splitter case

A case study on proactive failure prevention for a C 3 splitter unit produced predictive models that allowed operators to make timely adjustments, yielding significant benefits

NC Chakrabarti, Akalesh Tiwari, Manish Biswal, Sougat Das, Balamurugan Mohan, and Jesse Mallhi Reliance Industries Limited

T he reliable operation of critical infrastructure in the petrochemical industry is paramount to ensuring safety, efficiency, and profitability. Against this back - drop, a comprehensive case study on the development and implementation of an advanced predictive and prescriptive analytical system to prevent unplanned shutdowns in a dis- tillation system is discussed, specifically focusing on the C3 splitter column. A significant process upset, culminating in the trip of the overhead vapour heat pump compressor (HPC), which recycles heat from overhead vapour to the bottom thermo- syphon reboiler, resulted in a 30-hour production down- time, predicating a substantial financial setback. A detailed root cause analysis revealed a cascade of failures initiated by a malfunctioning level transmitter (LT) in the column bottom. To prevent recurrence, a suite of artificial intelligence (AI) and machine learning (ML) models was developed for the early detection of process anomalies. These models lever - age algorithms such as logistic regression (LR) and support vector machines (SVM) for classification tasks, hybrid data, and first principles-based mass balance, along with cumu - lative sum (CumSum) for detecting subtle shifts in process parameters. This system provides real-time analytics on sensor health, dynamic mass balance, reboiler performance, and overall process stability, offering operators a predictive and prescriptive framework to mitigate potential upsets. The system continuously ingests real-time data and provides predictive flags six to eight hours in advance. It also gives prescriptive guidance on what is drifting, why, and which variables to act on. This enhances operational discipline and strengthens margins by preventing avoidable disturbances. In the highly competitive and capital-intensive landscape of the petrochemical industry, operational continuity is a key driver of economic success. Unplanned shutdowns of critical process units not only lead to substantial produc- tion and financial losses but also pose potential safety and environmental risks. The complexity of modern refining and chemical processes, characterised by intricate interdepend- encies between equipment and control systems, makes them susceptible to cascading failures where a seemingly minor fault can escalate into a major plant-wide incident.

Propylene recovery unit performance Examination of an incident within a propylene recovery unit (PRU), a vital component in the production of high-purity propylene, provides further elaboration. The C3 splitter col- umn, the heart of the PRU, experienced a process upset that led to the tripping of its associated HPC.1 The result - ing shutdown halted production for 30 hours. This event underscored the critical need for a more sophisticated and proactive approach to process monitoring, one that moves beyond traditional reactive alarm systems. The post-incident investigation identified the root cause as a malfunctioning level transmitter (LT) in the C3 splitter column bottom. The faulty sensor provided misleading information to the control system and operators, leading to a sequence of events, including the loss of thermosiphon action in the reboil- ers, column bottom drying, and, ultimately, the protective trip of the HPC. The incident highlighted a crucial vulnerability: the lack of a system capable of detecting the initial sensor anomaly and predicting its downstream consequences. In the highly competitive and capital-intensive landscape of the petrochemical industry, operational continuity is a key driver of economic success In response, a project was initiated to develop an ‘intelli - gent system’ leveraging the power of AI and ML. The objec - tive was to create a robust, online near-real-time system for the ‘early event detection (EED)’ of process upsets around the C3 splitter unit. The solution encompasses a multi-fac - eted modelling approach, including: • Process upset prediction: A holistic model to forecast the probability of an impending upset based on a wide array of process parameters. • Sensor health analytics: A dedicated model to monitor the signature of critical sensors, like the column bottom LT, to detect malfunctions in real time. • Dynamic mass balance: A hybrid ‘data’ and ‘first prin - ciples’-based model to track mass flow and accumulation

PTQ Q4 2026