Control loop expert Michael Brown recently received a question from reader, Hansell Williams. The question relates to applying neural network and machine learning concepts to plant automation and control. Due to its topical nature, we decided to publish it along with Michael’s detailed answer.
Hansell’s question
As you would be aware, neural networks (NN) and machine learning (ML) are both topical subjects in the field of automation. Having been fascinated by NNs since the ‘90s, I wondered what your take would be on applying them to plant control. Does it even have a place? Regarding the modelling side of things: I recently read a paper on how NNs could be used to discover underlying models of systems, which could have some value.
In the early 2000s De Beers had an innovative research project that used a NN to substitute for the non-availability of an expensive instrument like a weightometer. The NN could be trained on scada history collected over years, which would make it a pretty reasonable predictor of a wide range of situations – good or bad. Adding a few of these around the plant could be an interesting way of putting to good use the millions of time-tagged data points accumulated over years.
Is it possible that historical process data could be fed into the NN to predict PVs or macro outcomes to degrees not possible through classical methods?
Michael’s reply
To address your query: My speciality is in the field of base layer feedback control and I know very little about APC systems like NNs, and have no practical experience of them. However, I know that on certain occasions they can, and do, play an important part in improving process control – sometimes dramatically.
It is quite possible that your suggestion of using process data history to predict outcomes could be of use in certain situations, similar to weather predictions that use masses of past data. However, on the individual loop basis, one must remember that the dynamics of most processes do change with time. I have heard people suggesting that the ‘half-life’ of a well-tuned loop is about 18 months. This sort of information can only be given by people who do optimisation and really understand what they are doing, and who work in the same plant for years. I have only worked regularly in one particular plant, a petrochemical refinery that I visit for one or two weeks every year. In this plant I have found that the dynamics in certain loops change dramatically from one year to the next. I am not sure if this is due to different operating conditions, as the processes are complex, but even the process specialists themselves generally cannot explain the causes. This is one of the many reasons that model based tuning does not work satisfactorily, particularly on complex dynamics.
There is a question that is often posed to me which is relevant to this discussion: “Why are we still using PID control in the modern world? It is a technology that is a century old!”
One reply to this was given at a talk presented many years ago by the renowned Greg Shinskey, at which I was fortunate enough to be present. He was asked the question by a leading academic expert on fuzzy logic. His reply was simple: “It is the only technology in control that can deal with unmeasured upsets without resorting to complex programming for that particular process.”
I always tell delegates on my courses to remember that feedback control is an innate part of life for all sentient living creatures. Virtually all of them use feedback control to move around and perform tasks. However, when human beings started trying to install feedback control systems in machines and processes in the early 20th century, they found it extremely difficult, and it took some of the world’s leading mathematicians to define how it works. This is the reason that few people really understand the technology, as the theory is mathematically intensive and unfortunately there has been a dearth in teaching the ‘practicalities’ of feedback control, which can make it much easier to understand. As mentioned in my articles, over 85% of control loops worldwide are not operating efficiently in automatic. However, once properly set-up, feedback control loops can work unbelievably well.
At this stage in our technology evolution, it is hard to imagine anything that can replace it, mainly as it deals so well with immeasurable upsets. It is for this reason that nearly all APC systems require ‘base layer’ feedback control systems below them, in order to ensure that the processes do in fact get to the setpoints as dictated by the advanced systems.
In conclusion, I think your proposal might well have some merit, and for all I know, may already be used in some APC systems. However, it would not replace, and would still have to use, the feedback base layer below it.
All reader’s with control loop related questions are invited to submit them to Michael Brown, Michael Brown Control Engineering cc, +27 82 440 7790, [email protected], www.controlloop.co.za
Michael Brown is a specialist in control loop optimisation with many years of experience in process control instrumentation. His main activities are consulting, and teaching practical control loop analysis and optimisation. He gives training courses which can be held in clients’ plants, where students can have the added benefit of practising on live loops. His work takes him to plants all over South Africa and also to other countries. He can be contacted at Michael Brown Control Engineering cc, +27 82 440 7790, [email protected], www.controlloop.co.za
| Email: | [email protected] |
| www: | www.controlloop.co.za |
| Articles: | More information and articles about Michael Brown Control Engineering |
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version