• By Chris Lee
  • IIoT Insight

By Chris Lee

When you are swapping parts in a compressor for the third time trying to get production restarted, you do not care what was for lunch last Saturday.

With the increased workloads brought on by a competitive market, plant engineers do not have much bandwidth to spare. They are frequently running from crisis to crisis. Their ability to think about historical problems is limited. Unless that history has bearing on an immediate issue, why spend hours, or even minutes, sifting through old data about a problem that was resolved months ago?

Yet this is how industrial analytics projects are frequently carried out today. Assemble a team of highly trained data scientists to look at volumes of historical data, logs, and contextual information (like maintenance records). Then spend months pouring over that data to create and validate models that can predict the behavior of interest—as it happened in the past. It is no great surprise, then, that one of the most difficult parts of these projects is getting timely input from the engineering subject-matter experts (SMEs). Yes. It may help in the future, they say, but how is this helping me now? I’m busy. I want to help, but I just can’t, they say. Short of deprioritizing production, what can be done?

Data analytics ought to focus on the problems the SMEs are dealing with today, instead of on problems that happened in the past. It is like a really smart new engineer who is learning every day by talking with experts about what they are finding. The key to their success is how quickly they learn and put what they learn into action. This is not easy to do with traditional analytics and machine-learning techniques today, but it will soon be. The proliferation of Industrial Internet of Things (IIoT) sensors is putting the power of “now” in the hands of people who need it. What might this look like?

Start with the data you have

A plant’s process or production engineer is starting up a new variant of an existing process. The equipment is not new, so there is a history of sensor data from running the old process. However, there is no data about what the new process looks like on this equipment. That is, the individual sensor behaviors, much less the interplay of the various subsystems on the sensors’ output, are uncharacterized. Nor is there a history of plant issues peculiar to this process.

Instead of trying to build a physics-based digital twin of the new process, or extrapolating rules from old behaviors, the plant engineer just flags the start of a new use case and begins collecting data. This lets him or her get feedback using insights from current operations.

In a short time, the system establishes what the new, system-level, multivariate “normal” looks like based on a few short, qualitative questions to the engineer each day. As the system finds novel behaviors, it alerts the engineer, explains the insight from the sensor data and past behaviors, and then asks for confirmation about what these events were: Were they different kinds of normal behavior? Were they known, adverse issues? Were they novel conditions? Based on these answers, the system automatically updates its alerts and provides more finely tuned notifications over time.

Improve performance while gaining valuable insights

Through this approach, plant engineers see things in real time that they might otherwise have missed. For example, subtle instabilities in the equipment related to certain parameter combinations trigger alerts, allowing a faster response to those instabilities. Or correlations are found between final product quality and alerts of unknown system behaviors, identifying conditions that lead to poor quality. This is done without taking plant engineers out of their day-to-day work. Every minute spent working on analytics saves multiples in lost production.

The traditional way of evaluating analytics is built on historical data. This is fine for those who are not in the trenches keeping a plant operating. But those on the front lines need to stay focused on the now. Only by applying analytics to today’s operations and validating them on current issues will plant managers gain confidence that the solutions apply to the real problems costing the business today. We need to start from now rather than meandering to it.

Reader Feedback

We want to hear from you! Please send us your comments and questions about this topic to InTechmagazine@isa.org.

Like This Article?

Subscribe Now!

About The Authors

Chris Lee is a customer success manager at Falkonry. Based in Sunnyvale, Calif., Lee guides customers from use-case concept to successful proof of value to production implementation of the company’s machine-learning-driven software that performs time series pattern detection and classification for Industry 4.0 applications.