Within our Predictive Asset Management (PAM) project at Bavaria, there are new developments that emphasize the potential strength of Predictive Analytics (PA). Last May, a failure incident occurred which led to unexpected stagnation and a loss of production. This involved a stuck bearing. The drive train in which it was lowered was currently out of scope of our pilot project and was not actively monitored.
However, this incident offered KIM Plus Delta the opportunity to revisit the archived data and retrospectively learn what the specific fingerprint was of this failure incident and whether it could have been predicted. After our data analysis, we have seen a strong fingerprint in the frequency domain, which has already been a month and a half in advance. The figure below shows that an early anomaly takes place around February, which then appears to pass. Around March 10., however, it begins to develop a trend that continues and peaks on the day of failure. That the curve returns to the nominal level after the repair underlines the strength of the correlation found. It is clear that a trend could have been used as signal and alarm. Based on this signalling, a (favourable) timed repair could take place and production loss could be prevented.
An exercise like this is the essence of data-driven methods: The data is analysed, with solid domain knowledge (engineering), to identify potential relationships between the incident and the data. And although these results need to be further validated, it shows the far-reaching potential of Data Science for Predictive Maintenance.
Learn more about the Predictive Maintenance pilot at Bavaria by reading here.