100
Calculated level value is going down to 10 - 15% while level transmitter value showing 65 - 75% and gap increasing
80
60
Half of maximum depletion @ 22:00 hrs
40
Maximum depletion @ 01:00 hrs
20
Calculated level Level transmitter value
0
15:00
18:00
21:00
00:00 (10 June)
03:00
06:00
Timestamp
Timestamp
Figure 6 Actual vs calculated level, revealing LT malfunction
• Mass_Out: The sum of all outlet streams’ flow rate, includ - ing the overheads and bottoms product flows. Flows are converted from volumetric (m³/h) to mass (TPH) using temperature-corrected densities. The model then cal - culates a cumulative accumulation (or depletion) over time. This accumulation, measured in tons, is converted into an equivalent liquid level (%) as ‘calculated level (%)’ based on the column’s geometry and fluid density. The mass balance closure (see Figure 5 ), defined by [(Mass_in-Mass_out) *100/Mass_in], has historically shown a natural variation between 96.8% and 100.1%. To isolate genuine changes in accumulation or depletion, calculations have been adjusted to exclude this expected variability. Therefore, cumulative accumulation or depletion is only added when the closure extends beyond these typi - cal bounds (below 96.8% or above 100.1%). • Performance: During the incident, this model clearly showed its value. While the actual LT was showing a false reading of around 65-75% (see Figure 6 ), the mass balance model correctly identified a continuous depletion of mass from the column bottom starting from 18:00. By 01:30 on Trip-day, the model calculated a cumulative depletion
or ‘noise’ as it responds to minor process fluctuations. A malfunctioning transmitter often loses this characteristic, becoming static, flat-lined, or exhibiting a smooth, unnatu - ral trend. This model quantifies this behaviour. It calculates the rolling standard deviation of the LT signal over a one- hour window. A (CumSum control chart is then applied to this standard deviation data. When the transmitter behaves unnaturally (for example, the steady, smooth rise seen dur - ing the incident), the standard deviation drops, and the CumSum chart quickly flags this deviation from the normal signature. • Performance: When applied to the incident data, the model successfully identified the LT malfunction starting from 16:30 on Trip-Previous-Day, just 30 minutes after the anom - alous behaviour began (see Figure 7 ). This provides a direct and early alert of sensor failure, independent of other process conditions. Model 4: Reboiler upset prediction This model was created to provide a specific warning about the health of the critical thermosiphon reboilers, as shown in Figure 8 , Reb-A (top) and Reb-B (bottom).
of 40 tons. The predicted level from the model showed a steady drop to as low as 10-15%, directly contradicting the physical sensor and accu - rately reflecting the reality of the column being drained. This provides an unambiguous, real-time validation of the pri - mary LT’s health. Model 3: Level transmitter health analytics This model focuses solely on the health of the level trans - mitter itself by analysing its operational ‘signature’. • Methodology: A healthy process transmitter exhibits natural, small-scale variations
Level transmitter health anomaly
1.0
70
0.5
68
LT health
0.0
66
Pre-ale rt threshold
-0.5
64
Pre-ale rt threshold
-1.0
Actual level starts increasing monotonically
62
LT actual
-1.5
-2.0
60
Date-time
Figure 7 LT health signal identifying malfunction via signature analysis
63
PTQ Q4 2026
www.digitalrefining.com
Powered by FlippingBook