This study deals with a novel signal processing approach for automated valve condition classification of a reciprocating compressor with seeded faults. The classification system consists of a front end time-frequency analysis platform for the vibration signal measured, fault feature vectors for making the formidable amount of time-frequency data manageable, and a probabilistic neural network for automatic classification without the intervention of human experts. Rather than representing each time-frequency data set with one single feature vector comprising three indices, namely time, frequency, and amplitude, the time-frequency plane is further partitioned into an appropriate number of sub-regions to enhance the characteristics representation of the time-frequency data. This study shows that a flawless classification can be realized by using the proposed approach with appropriate selections of index modification method and number of time-frequency subregions without resorting to the removal of similar fault cases.

Included in

Engineering Commons