Well Checked Systems International, LLC
17:00 - 17:30
Wednesday, 18 September 2019
T3.6 Machine learning and IIoT enabling unsupervised anomaly detection on compressors
Large machinery in a compressor station costs millions, and significant downtime costs even more. However, preventing a catastrophic engine failure is often as simple as a mechanic hearing a change in normal operation. Mechanics, though, do not have the time to listen to each compressor daily. Additionally, the presence of multiple compressor packages at a single location masks distinctive sounds as does the decibel levels of those machines.
Well Checked Systems International, in research with Johns Hopkins University's Machine Learning department, is working with XTO Energy to change this pattern. This research uses machine learning combined with a mechanic's expertise to create a preventative maintenance notification system. The foundational hypothesis is that sounds emitting from a compressor package are the result of vibration on or within the compressor. Because different problems cause different patterns of vibration, the sounds will be distinct from each other but similar within their category. Analytics of this audio are accomplished using a matrix of microphones to collect resonance waves for anomaly detection and distinction analysis through an IIoT framework.
The goal was to create a system which can detect anomalies on a device, then classify and alert based on these noises. The IIoT solution equipped each compressor package with a ruggedized computer which collects and analyzes audio. The system is currently installed as class 1 div 2 and can be installed as class 1 div 1. Initially, sample problem sounds were caused by a mechanic on-site manipulating the compressor package. As development continued, any other sounds that organically occurred were collected and used as part of training. Using a global feedback system, the intelligence and new sounds from each system benefits the entire network. Initial tests achieved recognition of normal and abnormal sound and classification at 94% accuracy. These tests were controlled for type of compressor and microphone location.
In conclusion, it is possible to detect audio anomalies, classify them, and notify an expert that there is a possible issue. The implications of these findings are significant: enabling predictive maintenance, exception-based monitoring, and maximization of the labor. This unsupervised anomaly detection can be done without advanced instrumentation or systems-wide integration and is model agnostic.