MachineMetrics blends unsupervised ML for predictive maintenance


Predictive maintenance might not be an AI application that seizes the public’s imagination, but is one that is delivering real cost savings and business benefits. It is also a fertile proving ground for algorithms that deal with dirty data, where the challenge is to extract meaningful trends from a lot of background noise.

MachineMetrics, based in Northampton, Massachusetts and founded in 2014 with $2.1 million raised so far, is one start-up targeting AI based predictive analytics for machine tools. It has combined a traditional rule-based approach capable of recognizing known conditions that might be indicative of problems or failures with unsupervised machine learning to identify relevant anomalous behavior. The latter has been shown capable of detecting impending faults in advance not picked up by the rules-based system.

The rules engine can be configured to monitor a stream of data and, when triggered, prompt some action which could just be a text notification or email, or alternatively an automated response. A basic rule might detect an asset that has been down for more than 10 minutes, while a more advanced rule might be triggered when say a spindle load exceeds a threshold several times in the previous eight hours.

Such rules can be refined in the light of experience but cannot anticipate all problems. Unsupervised learning can fill in these gaps but has proved challenging to deploy because there is a lot of noise hiding meaningful trends in given metrics, such as a level of vibration. MachineMetrics has employed various techniques to sort out the wheat from the chaff, starting by measuring the level of white noise, or entropy, within time based data series captured from a machine. This can help identify an anomaly when the level of entropy changes significantly.

Then among other techniques is the subtler one called first-order autocorrelation, which is a measure of how predictable, or random, a data series is. The point here is that when a machine starts to go wrong this may either make the data less random or more so.

It might become less random if a machine develops erroneous but consistent behavior as when say a bearing is worn, which would inject an element of predictability into the data. On the other hand, an intermittent fault could make the data more random. Unsupervised machine learning can detect such anomalies but is much more effective if first each data variable has been properly characterized so that relevant changes can be identified readily.

One client exploiting such features is Fastenal, a US industrial supply company based in Winona, Minnesota. It supplies fasteners, tools, and supplies to manufacturers and turned to MachineMetrics in the hope of making its own manufacturing more efficient or “leaner”. The analysis gained has indirectly improved efficiency by pinpointing where utilization could be improved. The company had assumed utilization was better than it was and was surprised to discover a given machine was only up 39% of the time. This prompted further investigation and improved utilization.

The data is gathered via an Industrial IoT edge device communicating with tools via either secure Ethernet, WiFi or cellular communication, interfacing directly to PLCs (programmable logic controllers) and controls. Encrypted data is streamed to the MachineMetrics cloud, where it is structured and aggregated to enable visualizations and analytics.