Self-service analytics for the subject matter expert

Technology brain

Will the rise of machine learning render historians obsolete? Or could self-service analytics leverage past data to the benefit of all assets?

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. This is because they are engineering intensive projects that typically require a skilled data scientist who is not necessarily familiar with the process. Additionally, most of these solutions are quite expensive from a licensing perspective.

Questions arise from these limitations, such as: “What with the other 99% assets where our modeling strategy does not make sense economically?” and “What about general abnormal situations and operational issues like heat exchanger fouling, filter pollution, catalyst degeneration…?”

These questions are not about major plant failures and downtime 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. This is data science for the Subject Matter Expert (SME) without 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 specifically unsupervised learning, is not new and could become very powerful going forward. However, within the process industry there are a couple of downsides of unsupervised learning. First of all it requires large data sets, 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, if they are 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.

Analytics Maturity

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 instead adds robust plug-and-play discovery, diagnostic and predictive self-service analytics on top of these data sources, leveraging historical data for a data-based prediction of the future.

So will machine learning eat historian? Only history will tell.


More information?

Learn 5 ways to improve production with self-service process analytics. Discover how a self-service approach allows your process engineers and subject matter experts to analyze, monitor and predict process performance without depending on data scientists.