Wavelet-based acoustic analysis for determining health condition of motorized two-wheelers

The garage mechanics assess the health condition of a vehicle based on sounds of vehicles. This expert knowledge needs to be stored in repositories and employed for automatic fault diagnosis. In this work, we propose a method to determine the health condition of motorcycles, based on discrete wavelet transform (DWT). The 1D central contour moments and invariant contour moments, of approximation coefficients of DWT form the feature vector. Dynamic time warping (DTW) classifier along with Euclidean distance measure is used for health classification. The results show over 98% classification accuracy for some subbands and some feature sets.

Vehicles generate different sound patterns under different working and health conditions. Garage mechanics diagnose the vehicle faults by monitoring the sound signals generated by them, based on the expertise acquired over the years. This work is an attempt to classify the motorized two-wheelers into healthy and faulty based on the acoustic signature. The acquired sound signals from motorized two-wheelers are decomposed into 14 subbands using Daubechies DB4 wavelets. The computed central contour moments and their invariants, for each subband, are used as the feature vectors. The averaged feature vectors of healthy motorcycles are used as reference feature vectors. Similarly, the reference feature vectors of faulty motorcycles are also obtained. The dynamic time warping (DTW) algorithm is used to compare the test and the reference feature vectors. Finally the Euclidean distance measure is used to distinguish between healthy and faulty DTW distances of the test sample to decide the health condition of the motorcycles. The proposed methodology is tested with central contour moments and invariant contour moments independently and also in combination. The testing is carried out for various combinations of subbands and for disjoint sets of training and testing samples. The classification accuracy varies from subband to subband and features. Over 98 % classification accuracy has resulted for some of the subbands and some of the feature combinations. The proposed work finds application in automatic fault detection and location in motorized two-wheelers. It also finds applications in automatic fault diagnosis of machines in industries, patient diagnosis in medical field, speaker recognition, musical genre classification, and the like.Figure optionsDownload as PowerPoint slideResearch highlights
► The work proposed finds applications in preliminary fault detection of two-wheelers.
► It employs the wavelet analysis, contour moments, followed by DTW classifier.
► Exhaustive testing is carried out with all the subbands and features.
► The results are encouraging and leave future scope for further analysis. It may be extended for fault source location in two-wheelers.