Deep Machine Learning: Predicting Assets and The Process
Geoffrey Hinton at the University of Toronto in the mid-2000s developed the latest, most effective Machine Learning (ML) techniques called Deep Learning. Deep Learning involves a set of algorithms that can decompose patterns into successively simpler layers which identify lower and lower levels of common patterns. This technique is computational very intense, but can produce more accurate results with less labeled events needed for training. In industry, common results can be found for each device in pools of similar equipment.
At this point in this blog, we can appreciate that using ML techniques can recognize issues that were undetectable five years ago. Enough data, smart algorithms, and big enough computers and the problem can be identified. But now, the question becomes how to get that work done. And, please note that we still have not got to the Prescriptive part which does more than identify the problem. Rather, prescriptive means clear advice, with a precursor of accurate prediction, which indicates “how” to solve the problem and causes an action to change the future. Many products claim to do this, but few actually can.
Back to ML. I assert that data, ML algorithms, and computers are not enough. The algorithms are free. Google publishes them, so does MIT; Microsoft and SAP make them available. Flexible and “elastic” cloud platforms such as Amazon, IBM, Microsoft, and others make the computing platforms available. Such an approach demands new people with new data science skills that you may not currently have on staff. This is “textbook” Machine Learning where each new ML deployment means a “special” project, with enormous effort and cost, that takes a long time, and cannot work in real-time. That approach does not scale. There are not enough data scientists available, the approach costs far too much, and it delivers results too late.
What else is needed for an efficient, timely solution that will scale, not require new skills, and provide timely results? First, an understanding that the data must be collected automatically in real-time from all relevant systems, by a productized ML application that does not need human help to verify, validate, cleanse and prepare data in the appropriate forms for analysis. A good solution will fit into the work processes already in place in the industry without demanding new people/skills. The software must be able to perform the analysis on behalf of the staff already at work; with only a little guidance and without them learning new skills. It must be able to detect data patterns in real-time and within seconds alert the appropriate staff about the identified problem, including all information that helps the receiver understand the precise issue, the detected change in behavior, the potential outcome, exactly when it will happen, and (best) what to do about it. That is a tough demand and it is tantamount to an automation and control system for maintenance and reliability.
And on that note, it reinforces a particularly relevant notion. If the project people coming into your plant do not know the nature or real-time data, do not understand plant historians, cannot read a P&ID (piping and instrumentation diagram), and (for equipment failure detection) do not know the data available in an asset management system, then “Houston, we have a problem.”
This article originally appeared Industrial IoT/Industrie 4.0 Viewpoints: INSIGHTS ON THE DIGITAL TRANSFORMATION OF INDUSTRYAugust 7, 2016 on the ARC Advisory Group blog:
About Mike Brooks: Mike Brooks is the former COO and President of Mtell. Mike is now an APM Advisory Consultant at AspenTech since its acquisition of Mtell in 2016. Mike comes with 25 years of leadership and management in Exxon, Chevron, and start-up companies in process operations, planning and scheduling, control systems, and IT systems. Mike holds a B.Sc. in Chemical Engineering from the University of Bradford, UK.