Recently I came across an article by Peter Reynolds (Analyst, ARC Advisory Group), where he asks, “Will Machine Learning eat Historian?”. His conclusion was; “Yes, for breakfast”. As usual, Peter’s comments are insightful and well educated. Today I would like to take the opportunity to add a couple of my own thoughts on the matter.
The 1% Critical Assets
Too often when speaking about analytics within the Process Industry, suppliers will focus on predicting asset failures. By working intensively with Industry Owners and Operators, we have learned that the majority of the traditional data modeling exercises only pay off (positive ROI) for the 1% critical assets because they are engineering intensive projects and typically require a skilled data scientist who is not necessarily familiar with the process. Aditionally most of these solutions are quite expensive from a licensing perspective. Of course that is something that would be easier to disrupt by a new player than the technology aspect.
Some questions do arise with these limitations: “what with the other 99% assets where our modeling strategy does not make sense economically?” and “what about general abnormal situations and operational issues (ie. heat exchanger fouling, filter pollution, catalyst degeneration,…)?”
These questions are not about major plant failures and downtimes that result in large immediate costs. But the sum of all provides significant improvements and brings your plant one step closer to achieving operational excellence. With TrendMiner we learned that for these activities you need an alternative approach, which we call self-service analytics; data science for the Subject Matter Expert (SME) whithout the need for a data scientist or traditional data modeling technologies.
Unsupervised Learning
Now on to the main topic of the article; Will Machine Learning eat Historian? Machine learning, or more specific unsupervised learning is not new and could become very powerful going forward. More specifically for the process industry though, there are a couple of downsides of unsupervised learning. First of all it requires large datasets, which doesn’t necessarily work when comparing one full process pattern against a limited set of previous results. And secondly it is not that good with ‘dirty’ data and will thus require cleaning in order not to provide too much false positives.
That being said there is still a lot of improvement possible with (semi)supervised learning for cases such as hypothesis testing. The problem with complete unsupervised learning is that it takes the SME’s experience out of the equation. It is the SME who is able to turn human intelligence into machine intelligence by using self-service analytics. Even with smaller sets of data, cleaned by the SME, an obvious pattern with a certain multivariate signature could be used for monitoring and Early Event Detection.
Yes, I too believe that unsupervised machine learning will become more and more important. But I would add that we are a long way from replacing the human intelligence and experience of the subject matter experts. In my view Machine Learning will primarily help people make better decisions.
Discovery Analytics
The Gartner Analytics Maturity model (see image below) provides us with 4 levels of analytics; Descriptive, Diagnostic, Predictive and Prescriptive. I’m not going to explain these in detail in this post, but if you are a Gartner customer, you can download the full paper here.

What we have experienced while developing TrendMiner, is that there is one question missing in this process: has this happened before? At TrendMiner we call this discovery analytics. In order to truly diagnose a current situation you have to be able to search the past, based on multi-variate pattern recognition. This is where the historian solutions currently lack in capabilities but replacing all historians with any other tool is definitely not the answer. TrendMiner can help with its robust plug-and-play discovery, diagnostic and predictive self-service analytics on top of these data sources. Don’t we all know; past behaviour is the best predictor for future behaviour.
In the end Peter left us with four major aspects to look for in modern analytics solutions for the process industry. I will address how TrendMinder takes these into account in a separate post.
So will Machine Learning eat Historian? Only history will tell 
Bert Baeck – Founder and CEO of TrendMiner.
Disclaimer: TrendMiner is an ARC Advisory Group customer. TrendMiner has attended the ARC Advisory Forums as a sponsor and mr. Peter Reynolds was the analyst who wrote the ARC View paper on TrendMiner, which you can download here.