Bayesian Approach to Infer Types of Faults on Electrical Machines from Acoustic Signal

Considering the classification of failures in electrical machines, the present paper aims to use supervised machine learning techniques in order to classify faults in electrical machines, using attributes from audio signals. In order to analyze data and recognize patterns, the considered supervised learning methods are: Bayesian Network, together with the BayesRule algorithm, Support Vector Machine and k-Nearest Neighbor. The performances and the results provided from these algorithms are then compared. The main contributions of this paper are the acquisition process of audio signals and the elaboration of Bayesian networks topologies and classifiers structures using the acquired signals, since the algorithms provide the generalization of the classification model by revealing the network structure. Also, the utilization of audio signals as input attributes to the classifiers is infrequent in the literature. The results show that the Support Vector Machine and k-Nearest Neighbor present a high accuracy. On the other hand, the Bayesian approach is advantageous due to the possibility of showing, through graph representations, the generalized structure to represent the trend of faults in real cases on industry applications