AI Replaces Bad Decisions, Not People
Klaus M. Blache | May 1, 2023
By Dr. Klaus M. Blache, Univ. of Tennessee Reliability and Maintainability Center (RMC); Victor Foster, International Flavors & Fragrances; Christopher Lemmon, Univ. of Tennessee Reliability and Maintainability Center (RMC)
As part of the RMC initiatives, we also get involved in “proof-of-concept” solutions for industry. I asked Victor Foster to tell his story on a machine learning/AI that he performed with Quartic.ai (UT-RMC member) and Chris Lemmon (UT-College of Engineering graduate student):
When I first heard about Quartic.ai, San Jose, CA (quartic.ai) software, I was excited to try it. I’ve been to several conferences and heard how AI would “change the world” and “make our jobs obsolete.” Despite my confusion, I was curious to see how this new technology could fit into my work as a reliability engineer at a large food-manufacturing plant. I started to think about all the different ways I interacted with process and asset health data. I then considered each task that involved decisions based on this data.
A demonstration was very impressive. The anomaly in my test trial was a low-oil event in a gearbox. The software had no idea what was going wrong, but it could mathematically find where something was out of order. Once it detected inconsistent readings, it would send warning communications to maintenance personnel.
I initially thought I could just dump data into the software and it would tell me what to do next. However, that was not the case. AI is more of a tool than technology that would eliminate my job. This realization led me to try the software on one of my biggest plant headaches.
Thankfully, I did not have to go it alone. I partnered with Chris Lemmon of UT-RMC. Chris was already experimenting with Quartic.ai. We met biweekly to review data gathered from the process. I would pull data from my data historian in large Excel files to be uploaded. This was a work-around to use the software since it was not installed at the plant. Normally, Quartic.ai would be connected directly to your facility’s data historian for live data analysis.
The trial was conducted on a planetary gearbox attached to a horizontal centrifuge. This gearbox is always spinning, so visual inspection of the oil level or leaks was impossible. These gearboxes are very expensive to rebuild if damaged, and the most common failure mode was loss of oil due to an O-ring failure.
Before using the new AI software, we tried several maintenance strategies to reduce gearbox damage. With 30 units involved, the “run-to-failure” strategy was not sufficient from a cost-sustainability perspective. The new strategy was to track oil samples every six months to see what units were showing metal wear. This helped improve our ability to proactively replace older gearboxes but didn’t do anything to stop the O-ring failures.
It was decided that we could use an infrared sensor on the outside of the gearbox to obtain advanced warning to take a unit down. This idea made perfect sense and was approved for capital. We installed infrared sensors on all the units, thinking this would save the day. We were too optimistic.
Chris thought this problem would be a great trial for the AI software. We identified 20 process variables from the historian, including motor load, vibration, oil flow in the centrifuge main bearings, and process flow rate. The tool quickly configured and organized the data into easy-to-operate buttons. The data was then visually graphed into a timeline. Fortunately, we had an event in which one of the plant centrifuges automatically shut down due to overload of the back drive motor that supplies power to the gearbox.
After telling Chris exactly what date and time the centrifuge went down over the four-month data set, he went to work in the software. The event data was effective at training the software on what to look for in the data by explicitly defining an anomaly. When the maintenance crew inspected the centrifuge, the only issue they found was a low oil level. They filled the gearbox and were able to operate the centrifuge with no problems.
The timeline now looked like Fig. 1. It had a time before the “low-oil event,” the event itself, and the time after the oil was refilled.
This data organization helped the software quickly learn what to look for to predict the next low-oil event. It’s like having a real person watching the process. The software now provides real relationship data on all 20 of the process variables we uploaded. I was not very happy once I understood the findings.
To my surprise, the data corresponded to the motor back-drive load, rather than the new expensive, shiny, infrared sensors. I had spent three months telling everyone how amazing they would be. Instead of making my job obsolete, the software was keeping me from making bad decisions. The result can be seen in Fig. 2.
Using the correct corresponding process variable for the back drive-motor load, it predicted equipment shutdown 12 hr. before the event. If Quartic.ai was installed and given real-time data, this event could have triggered a controlled shutdown of the unit. The alarm sent to the operator 12 hr. before would have stopped any process upset. After the gear box oil level was corrected, the AI tool did not see any process anomalies.
This trial demonstrated what AI could do for a normal engineer working day to day. It has a tremendous ability to look at different data sources at the same time and pinpoint the best solution. EP
Victor Foster is the lead PdM Reliability Engineer at International Flavors & Fragrances, New York, NY (iff.com). He provides support for four plants.
Christopher Lemmon is a Graduate Research Assistant at the Univ. of Tennessee RMC, finishing his Master’s degree in RME.