The plant is controlled as usual by a process control system which takes inputs from in-line analyzers (in this case we have sulfur analyzer). During the process, after many hours of work, the in-line analyzers tend to become unstable giving bad results; this is due to several mechanisms: loss of linearity, stream flow blocks, dirty stream flow ...
This situation makes the process control unit practically "blind" and therefore unable to keep the process in-control.
The only thing to do is then to stop the process in order to check and reset the in-line analyzers.
This practice yields to loss of hours of time, product quality and money.
Comments
In order to handle this problem, a consolidated practice among process engineers, is that of comparing the results of in-line analyzers with lab analysis by taking 4/5 stream samples during a whole process (a whole process average duration is 3 days). In this way experienced engineers can "try to guess" the behaviour of in-line analyzers. Obviously this practice can't avoid plant stop but in some rare cases can help to anticipate the analyzer's response drift.
Furthermore money needed for the lab work must be taken into account.
Solution
Our solution consists of two software components which implement several ad-hoc algorithms developed by DepQuest and mainly based on non-linear dynamics, chaos theory, signal processing and machine learning techniques.
The first component (soft sensor) is able to predict several hours in advance the analyzer malfunctioning by recognizing whether the behaviour changes are due to a real change in stream parameters or to analyzer failure.
The second component is capable of learning process behaviour in terms of in-line analyzer outputs and then simulate the presence of the analyzers themselves, when they are excluded from the control loop for maintenance, by forecasting real time measures to the process control unit.
These two components together allow to achieve the following results:
no more plant stop (because of the presence of the data forecaster component)
no need of laboratory cross validation
unknown and new analyzers and process failures can be recognized and classified (machine learning)
preventive maintenance can be planned in advance
The working logical schema are represented in the following animation:
Animation 1 - process data flow
(press the play buttons to view data-flow in the different situations)
Our Results
Let's take a look to a real process analyzer (in this case study sulfur analyzer ) output:
Fig.1 - sulfur analyzer output
The above plot represent "percentage vs. time (in days)".
We can see that after 1.75 day c.ca the analyzer's response becames first unstable and then practically unusable. In normal conditions this event is detected only when it happens, it means too late.
Below you can see the result of our Soft Sensor computation referred, for simplicity, to the H2S time series:
Animation 2 - SoftSensor in action
As you can see in Animation.2, the Soft Sensor can detect the "analyzer's failure" c.ca 5 hours in advance!
This important trigger is then used to activate the "data forecasting" unit which simulate analyzer's data and feeds the process control unit with these data. In this way the process can be kept running.
In-line analysers can therefore be excluded from the plant for maintenance without stopping the process.
Moreover, the system is able to "classify" and "recognize" the type of transition followed by the system when it changes its behaviour from "normal" to "failure". With this last tool, process engineers have the possibility to classify the failure causes and to discover new ones.
Conclusions
There are no conclusions, because this is a work always in progress!
As we apply this techniques to other plants (distillation columns, blending ...) we discover other powerfull algorithms and our solution library grows up.
In the next future we'll publish other documents related to:
