PTQ Q4 2026 Issue

System implementation and operational impact The true value of these individual models is realised through their integration into a unified, online web interface. The models run in near real time, and the insights are shown on an intelligent dashboard, in addition to the mail alerts to stakeholders whenever an anomaly is observed by models. This intelligent system provides a single pane of glass for stakeholders to monitor the health of the C3 splitter unit. The dashboard visualises: • Overall upset probability: A real-time trend of the output from the logistic regression model, with a clear threshold for alerts. • Top contributing factors: When the upset probability rises, the dashboard displays the top process variables that are contributing to the risk, guiding the operator’s diagnos - tic efforts. • Mass balance and level verification: The dashboard shows the real-time mass balance closure and plots the physically measured LT signal alongside the model-cal - culated LT signal, making any discrepancy immediately obvious. • Specific equipment health: Alerts from the LT health model and the reboiler upset models are displayed, allow - ing for targeted intervention.3 The operational impact of this system is transformative. It shifts the paradigm from reactive troubleshooting to proac - tive and predictive maintenance and control. With the mul - tiple layers of early warning provided by the model suite, operators would have had hours, not minutes, of advance notice of the impending shutdown. An alert from the LT signature model at 16:30, or from the mass balance model showing a discrepancy, would have immediately prompted a check on the column level and a correction of the bottoms flow, averting the entire cascade of failures. The system empowers operators with data-driven insights, turning raw process data into action - able intelligence and safeguarding the plant against costly unplanned downtime. Conclusion Model validation against historical data confirms the sys - tem delivers accurate, actionable alerts with a proven lead time of six to eight hours. This predictive window allows operators to make controlled adjustments well ahead of critical inflection points, eliminating emergency trips. Ultimately, this system shifts operations from a reactive to a predictive stance. Operators can now anticipate what will happen next and why, enabling proactive actions that prevent production losses and mitigate severe process upsets around the distillation columns. As discussed, the unplanned shutdown of the C3 split- ter unit served as a powerful catalyst for innovation. The incident, which originated from a single sensor failure (i.e., the level transmitter) and escalated to a full unit outage, highlighted the limitations of traditional process control and alarm systems in managing the complexities of a highly integrated process unit. By pairing a holistic process upset model with targeted modules for sensor health, mass balance, and equipment

Reb-A

1.0

Model alert @ 21:45

0.8

0.6

∆P & ∆T started uctuating after 21:30

0.4

0.2

Probability

0.0

Timestamp

Reb-B

1.0

0.8

Model alert @ 22:30

0.6

∆P & ∆T started uctuating after 22:15

0.4

0.2

Probability

0.0

Timestamp

Figure 8 Reb-A (top) and Reb-B (bottom)

• Methodology: A support vector machine (SVM), a pow- erful classification algorithm effective in high-dimensional spaces, was used. The model was trained on key param - eters for each reboiler, including inlet/outlet temperatures and pressures, flow rates, and the calculated ΔT and ΔP. The SVM learns to find the optimal hyperplane that sep - arates ‘healthy’ reboiler operation from ‘upset’ conditions (i.e., thermosiphon loss). At times, there is not a straight line or simple hyperplane that can separate the classes. For such cases, SVMs use a technique known as the ‘kernel trick’ to project the data into a higher-dimensional space where a hyperplane can effectively separate the classes. The kernel trick allows the SVM to model complex, non-lin - ear relationships between the parameters. • Performance: The model was trained on the past two years of data until the upset day, with the upset period labelled as the ‘bad’ class. In validation, the model success - fully predicted the onset of the upset condition in reboiler Reb-A starting at 21:45 and in reboiler Reb-B starting at 22:30, perfectly aligning with the post-mortem analysis of the process data. This gives operators a specific alert that the column’s heat input is compromised.

PTQ Q4 2026