How Smart Condition Monitoring Can Help Operations Teams Keep Up with Industry Changes

Summary

The main challenge facing manufacturing today is talent shortage. And yet the demands on manufacturing to improve productivity and quality remain quite consistent. The bulk of the productivity challenge is caused by new abnormal conditions arising somewhere in plants at any given time. How are we to cope with so many surprises in operations and maintenance practices? Can we even do that when we haven’t been able to prepare for AI to the level that experts claim to be necessary?
 
Condition monitoring is the path to go from scheduled maintenance to condition-based maintenance. However, classic condition monitoring methods are not working. Such methods need specialized training, tools and, most importantly, a lot of time that the industry simply does not have. Even if some of these resources are mobilized, the most a model-based approach can do is track known failure modes that have happened in the past.

The challenge

To understand how time and talent can be potential bottlenecks to productivity gains, let’s take a typical integrated steel mill as an example–it can generate upwards of 20,000 signals from multiple sources. By spending a lot of man hours of an already stretched workforce you might be able to monitor a few hundred of these signals for certain behaviors of interest with custom models. This means 98% or more signals go unmonitored. Just imagine the number of missed opportunities for getting early warnings about excursions and other faults.

Your original goal was to monitor every single signal, for every possible failure, at all times! But for the kind of scale we just described, it’s very difficult to anticipate every signal variance that might indicate a problem. A more intelligent approach is needed and that is what leads us to smart condition monitoring.

The reality

Manufacturers are convinced that smart condition monitoring is real and it can improve their operations. However, in the course of helping manufacturing organizations apply smart condition monitoring solutions, we have seen several of them facing the above challenges. What is needed to overcome such challenges is a  self-supervised anomaly detection approach, which does not require you to spend time defining which signals are the right ones to monitor, which periods of time correspond to normal behavior, or which groups of signals are primary contributors to a particular failure. A capable AI system should be able to learn all of that by itself, thereby reducing upfront effort, time spent in training on specialized tools, and the need for subject matter expertise at every stage. Such a system should only need minutes to days of data instead of years of accumulated historical data.
 
An intelligent approach like the one we described also overcomes the challenge of spurious benign conditions and new operating modes. It can also automatically incorporate new behavior into the baseline, including any ignored anomalies, removing significant effort from ongoing maintenance of a smart condition monitoring solution.

The result

With such a smart condition monitoring approach you don’t have to cough up prior occurrences of problems, nor do you have to perform labeling of the data. There’s no need to prep the data or document all the failure modes in advance. Plus, you get value out of the 99% of your automation data which you have not been able to exploit so far.