The Problem

The plant may be running in a seemingly stable state for a long time, with no undue alarms or obvious process upsets. Suddenly, for no apparent reason, a major piece of equipment, such as a compressor, trips. Investigation shows that severe mechanical damage has occurred.

Analysis of process historical process data, alarm logs, and operator actions does not reveal why the damage occurred in the first place.

The Consequence

The equipment is repaired or even replaced, and the plant restarts. Hopefully, the situation does not repeat itself.

The Cause

There are instances where failures occur for no reason. In the absence of any data to the contrary, it is easy to conclude that the failure occurred as a result of bad luck and that nothing could have been done to prevent the catastrophe.

When developing algorithms that control the plant, the engineer has to consider every conceivable scenario and take appropriate action. For example, if condition A occurs AND condition B occurs, then action C must be initiated. These rule-based solutions tend to be binary in nature and may run into many thousands of combinations and permutations.

The challenge with rule-based thinking is this: in the above scenario, what if the plant is tending in the direction of condition A, but at the same time it is tending away from the direction of condition B. Is action C negated? These changes fall outside the realm of discreet logic and hence go undetected and raise no suspicion until it’s too late.

As a further example, imagine driving a vehicle. When you press the brakes, you naturally expect the vehicle to start decelerating immediately. However, how would we react if the vehicle accelerates even by a tiny amount when you press the brake? It would be clear to you that something serious has gone wrong, but a system designed to monitor the health of the vehicle may not pick up the error unless the engineers had specifically anticipated the issue and had developed a rule to deal with the anomaly.

The Solution

There may be warnings of imminent failure as far as 32 hours in advance of the failure event. These warning signs are very subtle to detect and will not result in process alarms being triggered immediately.

In order to detect anomalies of this nature, statistical models can be developed based on historical data that calculate the relationship of each process variable in the plant with every other process variable.

Examination of historical data to find a period when the plant was running optimally, and this data is then used to train the statistical model to understand what ‘good’ looks like.

The statistical model then runs parallel to real plant operations and compares real-world values against the statistical ideal. Minute changes in the relationships between plant variables are immediately apparent and are examined to understand their root causes. These anomalies are early indicators that equipment may fail in the near future unless corrective action is taken beforehand.