Kurt Pichler, Thomas Buchegger, and Matthias Huschenbett
Fault Detection, Time-frequency analysis, Switching Model
This paper presents a novel data-driven approach to diagnosing faults in reciprocating compressor valves. Based on easily available vibration data, a transformation to a high-dimensional vector space is performed. By defining a metric in this vector space, the distance between the actual compressor state and a reference compressor state is calculated. Excessive distances indicate a faulty valve. Large distances may be due not only to faulty valves but also to non-steady-state load conditions. Hence, steady-state operating modes must be recognized automatically. To enable evaluation of multiple levels of steady-state operating mode, the reference compressor state is determined by a switching model. The switching logic is based on Voronoi diagrams of the load control variables of the compressor. The experimental results confirm that the method performs well if the compressor is operated at steady-state load levels similar or identical to training instances.
Important Links:
Go Back
