Localisation of multiple faults in motorcycles based on the wavelet packet analysis of the produced sounds

Abstract

Service station experts examine the sound patterns of the motorcycles to diagnose the faults. Automatic fault diagnosis is a challenging task and more so is recognition of multiple faults. This study presents a methodology for localisation of multiple faults in motorcycles. The sound signatures of multiple faults are constructed by fusing the individual signatures of faults from engine and exhaust subsystems. Energy distributions in the approximation coefficients of wavelet packets are used as features. Among the classifiers used, artificial neural network is found suitable for detecting the presence of multiple faults. The recognition accuracy is over 78% when trained with individual fault signatures and over 88% when trained with combined fault signatures.