Refining India September 2025 Issue

FTA (blue) and FTB (red) during that time, highlighting the variance in FTB, leading to a sudden drop. Post-trip data – offline validation : An even more striking validation occurred with post- trip data. After the trip, the impulse lines were cleaned. However, the anomaly signal, as detected by the model during validation, began to increase again beyond the threshold. This suggested that the lines

Threshold- Yellow alert

Anomaly signal Model output

Dip in ow sensor readings

Figure 7 Anomaly signal and sensor data during low capacity, showing a sudden reduction in cracked gas compressor third-stage flows but without signs of impulse line choking

anomalies well in advance, providing crucial lead time for intervention . Detection lead time : The model successfully identified the onset of anomalous behaviours up to 14 days ahead of what would become an ‘extreme anomaly’ or potential failure. This lead time is invaluable for planning and executing preventative maintenance without disrupting operations. Historical plots of anomaly signals clearly showed a gradual increase, crossing the predefined threshold, indicating developing issues. An analysis of sensor readings from before the incident revealed a clear pattern. Variance in flow turbine (FTB) began to appear long ago, steadily increasing over time. Around 10 days before the trip, a high deviation was observed between the two flow transmitters. Notably, at the time of the incident, a sudden drop in flow occurred, causing the machine to trip. Subsequently, maintenance flushing of the impulse tube was done on FTB. Post-flushing, the variations in the transmitter subsided, returning to a normal state. It was seen that after the flushing activity, the anomaly signal also reduced to near the threshold value. This historical validation confirmed the model’s ability to identify such progressive degradation patterns. Figure 5 shows sensor readings for

might not have been fully clean or that new plugging was initiating, creating a ‘potential for trip again’. The model’s early detection of this re-plugging potential, which was subsequently addressed by maintenance-flushing the lines again, effectively averted another significant unplanned shutdown. This demonstrates the model’s robustness and its ability to detect subtle, re-emerging anomalies. Online validation and robustness Following successful offline validation, the model was deployed online, where its robustness was further validated through real- time monitoring and a separate plant incident. In the online deployment, the model triggers every 30 minutes, which fetches real-time data from the historian, pre-processes it, and then runs the model. Validation with other equipment trip : A month after the model was deployed online, another compressor in a different section of the plant tripped. This event led to a sudden and drastic reduction in cracked gas compressor third- stage flows. Crucially, the model’s output for the cracked gas compressor third-stage sensors remained ‘normal’ and below the anomaly threshold. This was the expected and desired outcome, as both third-stage flow sensors

Refining India