End of the Assembly Line Gearbox Fault Inspection Using Artificial Neural Network and Support Vector Machines

Abstract

Gear fault diagnosis is important not only during the routine maintenance of machinery, but also during the inspection of newly manufactured gearboxes at the end of the assembly line. This paper discusses the application of an artificial neural network (ANN) and a support vector machine (SVM) for identifying faults in the gearbox, using the psychoacoustic and conventional statistical features extracted from acoustics and vibration signals. It is observed that at the end of the assembly line, the gearbox is tested by mounting it on a test bench and driving it by an electric motor. Based on the sound emitted while running on the test bench, the operator decides on the acceptance of the gearbox for further assembly on a vehicle or machine. This method of acceptance or rejection of the gearbox involves subjectivity and it is not reliable. Hence, it is important to have a reliable and objective fault detection and diagnosis method. To eliminate subjectivity, psychoacoustic features, which are derived from the science of listening in human beings, are proposed to be used as features, along with ANN and SVMs as classifiers. To ascertain the ability of the psychoacoustic features to classify faults, laboratory experiments are carried on a test setup by simulating faults like a gear shaft misalignment, a profile error of a gear tooth, a crack at the root of the tooth, and a broken tooth. ANN and SVM are trained with the psychoacoustic features extracted from the acoustic signal and other statistical features from the acoustics and vibration signals. The trained SVM and ANN are tested for fault classification for these features and their accuracy is compared. Fault classification accuracy is found to be 95.65% for ANN and 93.44% for SVM with psychoacoustic features and is found to be better than pure statistical features obtained from the vibration and acoustic signals. With the optimised ANN and SVM architecture, SVM is found to be performing better than ANN. It is concluded that the psychoacoustic features, along with the ANN and SVM method, could be adopted at the end of assembly line inspection to make the inspection process more objective.