Artificial neural network based gearbox fault diagnosis

Abstract

Gearboxes are employed in a wide range of applications such as in the automobile industry, marine propulsion systems, aviation industry etc. Breakdown of the gear train can not only result in significant financial losses but may also prove to be fatal. It is, therefore, necessary to identify a fault in a gearbox before catastrophic failure occurs. The vibration and sound emission signatures (also referred to as acoustic signatures in the thesis) of a gearbox carry abundant information about its condition. Whenever a fault develops in a gearbox, its vibration and sound emission signatures exhibit a change. In order to ascertain the condition of the gearbox, meaningful parameters representing the gearbox health condition need to be extracted from the acquired gearbox signatures. This process is referred to as the feature extraction step. This step usually involves extraction of features in the time domain, frequency domain or the time-frequency domain. Not all features extracted during the feature extraction step have the same fault classification potential. Often feature reduction is required to retain only the parameters with superior fault discrimination capability. However, this feature reduction comes at the cost of additional computational burden. An automated method of gearbox fault diagnosis is desirable wherein the condition of the gearbox can be assessed with minimal intervention of the operator. Artificial intelligence (AI) paves the path towards automated gearbox fault diagnosis (GFD). The present study is based on the application of artificial neural networks (ANNs), in particular, the back propagation neural network (BPNN) and the adaptive neuro-fuzzy inference system (ANFIS) in diagnosing gearbox faults. An ANFIS is a hybrid system of ANN and fuzzy inference system wherein the fuzzy sets andrules are automatically tuned as the network is trained. The features extracted from the acquired gearbox signals are utilized as inputs to train and test an intelligent classifier such as an ANN with the objective of diagnosing the condition of the gearbox. This step is referred to as the pattern recognition step.