Currently we maintain “stuff” with no insight into what happens next. We let “stuff” break or periodically check it irrespective of need. Our ML application directs users to only maintain “stuff” when failure is imminent. This eliminates emergency call-outs and PMs, and so changes the business model (it’s disruptive). I’ll highlight the dilemma of “reliability”; is it worse to miss an impending failure, or raise a false alarm? There are three parts to an ML application. The learning phase; finding a pattern within the dataset. The labelling phase; associating a pattern with a name. And the matching phase (is the pattern known). I’ll talk about each with respect to predictive maintenance.
Required audience experience: No specific knowledge required.
Objective of the talk: To touch upon the disruptive aspect of predictive maintenance, and the dilemmas it creates. To convey the multi-disipline, multi-part nature of ML, and how we can break-down the problem and signal (dataset). To convey how learning is best approached collaboratively whether in a society or ML.
Keywords: ANN, ML, maintenance, prediction, heuristics, reinforcement, disruption
You can view Allister’s slides via the link below:
Allister Mannion: ML in Predictive Maintenance
You can watch Allister’s presentation below: