Is predictive maintenance at a dead end?
Predictive maintenance has become a standard phrase in the equipment business and it is taking hold with OEMs. However, maintenance teams are not seeing much use for it. For them, “predictive maintenance is like chasing one’s own tail” – a perspective we’re increasingly hearing from manufacturing leaders for whom the inability to scale and the time and effort it takes to get benefit from predictive maintenance is a source of major disappointment.
If you are also running plant maintenance, then this sentiment likely resonates with you. The writing has always been on the wall but the industry as a whole was too caught up to notice it. Just like there is no general artificial intelligence, there is no general predictive maintenance, at least not yet! It’s not even that the technology is necessarily to blame here. Simply put, the same problems don’t recur, the patterns are never the same, and hence it’s not possible to have enough examples to learn from to have confidence that the next problem can be prevented.
A thumb rule we’ve arrived at in our own experience is that for classification to work you need 10 samples or 10 recurrences for whatever behavior you are looking for. They don’t even have to be identical recurrences and yet gathering 10 such examples is a tall order for most manufacturers that operate a highly controlled operation. Many of the problems that happen in a plant have a single occurrence in a whole year at most. Another inherent problem is that plants change quite often. The use cases for which you get good results today might not be relevant in the future because your line changed, so its behavior became completely different from what had been modeled.
The lesson here is that attaining a predictive level of operations wherein you get timely early warnings about equipment failure can be your apex goal; it cannot be your foundation. The foundation has to be anomaly detection for minimizing excursion risks and to increase the technician’s efficiency and productivity of engineers. By identifying excursions everywhere and all the time, a competent anomaly detection engine can inform operators before even the first occurrence of abnormal behavior in most cases. This highly reduces the time engineers spend troubleshooting, finding answers, and recovering from problems.
The approach offers greater coverage since anomalies are based on self-learned normal behavior and not a particular behavior of interest. Even if predictive maintenance is unattainable at the moment, predictive analytics certainly isn’t. With anomaly detection, you’re constantly predicting what will happen next, and comparing it against what actually happens. If the two values are not the same, that means there is an anomaly. Rather than predicting when to perform maintenance, you’re simply predicting what happens next. In the process, you are illuminating to the right people where attention is needed, why it is needed, and equipping them to learn from the incidents. This is why we have perfected anomaly detection to a point where it cannot fail. Automated anomaly detection instantly frees you from the constraint of not having enough examples and you can apply it at scale in a highly cost-effective manner.
About the author
This article was written by Dr. Nikunj Mehta, the Founder & CEO of Falkonry. He founded Falkonry after realizing that valuable operational data produced in industrial infrastructure goes mostly unutilized in the energy, manufacturing and transportation sectors. Nikunj believes hard business problems can be solved by combining machine learning, user-oriented design & partnerships. Prior to Falkonry, Nikunj led software architecture and customer success for C3 IoT.