Refining India September 2025 Issue

Impulse lines ushing

Extreme trend of anomaly detected by model

Model output or anomaly signal

thresh

Sensor reading

FTA FTB

Time stamp

Figure 5 Plot of anomaly signal values over time, illustrating the detection of an extreme anomaly trend and subsequent drop after impulse line flushing

this anomaly detection model, it was crucial to identify a period of genuinely ‘normal’ operation for the flow transmitters. Historical data analysis revealed that one or both cracked gas compressor third-stage flow transmitters frequently exhibited fluctuations, indicating varying degrees of impulse line plugging. Recognising this, data where fluctuation is minimum was meticulously selected as the ‘normal’ training dataset (see Figure 4 ). This period was characterised by the absence of abnormal fluctuations in both transmitters, confirmed after recent

flushing activities. This dataset specifically covered operational conditions with varying flows, ensuring the model learned normal behaviour under all relevant plant loads. The remaining historical data, encompassing periods of known and suspected anomalies, was then reserved for validating the model’s performance. Offline validation of the model: proving efficacy The offline validation phase demonstrated the model's impressive capability to detect

4

Cleaned after the plant trip

Cleaned again

1

3

3

2

Re-plugging/ ineffective cleaning

Normal. Anomaly signal below threshold

Anomaly signal

1

2

4

thresh

0

Figure 6 Anomaly signal after the trip, showing a rise in anomaly indicating potential re-plugging before a subsequent cleaning averted another trip

Refining India

69

Powered by