IMC-2017 Learning Session - 41:17
by Simon Xu, Petasense, Inc.
A machine learning based anomaly detection program is the ultimate reliability tool for any industrial maintenance team. Basic temperature and vibration RMS alarms are simple to implement, but with the prevalence of wireless internet-connected sensors, machine learning can be the link to a truly cutting edge reliability program. An anomaly detection program can trigger an alarm when it detects multidimensional pattern changes in the data it collects. It can then perform a measurement to identify how much this anomaly deviates from a baseline, which will correlate with the health of the machine. This tool can provide more value than traditional systems, by capturing the same anomalies that are triggered by band alarms, then going further to identify false negatives from abnormal patterns.
As with any machine learning program, the anomaly detection program requires a training period, which can be as short period as 1-2 weeks. During this time the system observes and learns the normal behavior of the equipment. Since the machine may not be perfectly healthy at the outset, it is important to define the machine’s existent health using accumulated data from similar machines. In just a few weeks, the machine’s health can be accurately determined and the machine learning algorithm can start identifying anomalous behavior autonomously. Over time, different machine faults can also be tagged, training the program to recognize similar defects in the future, even among different machines.
While a machine learning program cannot replace an analyst, it has the potential to become the “go-to” analysis tool to identify machinery faults and make work order decisions more efficiently. A machine learning based health score can also be used to decide whether to perform a PM step or not and change the frequency of a PM task.