Acceptance and Classification of Gears Based on Sound Signals Spectrum Analysis

Abstract

This paper work presents our proposed systematic approach in solving the industrial problem on identifying the good and bad gears by analyzing the spectrum of sound signals produced by the gears; In this proposed approach we treat this problem as a pure machine learning and classification problem having two classes (good gears and bad gears), whereby we have analyzed the spectrum of several sound signal samples from good and bad gears then extracted five audio features from their spectrum (Short time energy, zero crossing rate, Spectral entropy, pitch and block energy entropy) which after investigation, we found a significant difference between the two classes. We formed a 5D features vector; we used 10 features vectors from good gears and 10 from bad gears as our training samples for finding the discriminating point. After features extraction, we apply Support Vector Machine (SVM) learning classification method to classify the new features vector extracted from the unknown sound from the gear.