Condition monitoring of a reciprocating compressor using wavelet transformation and support vector machines

Abstract

Condition monitoring techniques were applied to areciprocating compressor in order to determine if faults werepresent in a system. Through the use of vibration basedsensors, fault monitoring of the crank-side discharge valvesprings was accomplished. Data was collected through arange of injected fault conditions and analyzed through theuse of discrete wavelet transformations. The waveletcoefficients produced were transformed into a sixdimensionalfeature space though the use of first and secondorder statistics. By using a support vector machine classifier,the nominal and faulted condition data was used to train afault monitoring classifier. This classifier was verifiedthrough the use of additional test data, and resulted inclassification rates of 90% and above. This result is based onthe trial of a multitude of different wavelets and supportvector kernels in order to achieve the optimal performancefor the dataset.