Deep Learning System with Data Augmentation for Electric Machinery Fault Diagnosis from Vibration Signals

A challenging aspect of applying Artificial Intelligence (AI) for electric machinery fault diagnosis systems in practical engineering scenarios is to collect adequate cases with sufficient faulty conditions for training and validation. In particular, imbalanced data sets, such as motor vibration signals having much more data toward no-fault or common fault conditions than the rare ones, usually lead to biased predictions or misdiagnosis. In this paper, a statistical feature extraction model, the proposed Synthesis Data Augmentation Method and the Deep Neural Network (DNN) are investigated and implemented to enhance the classification accuracy from vibration signal measurements. The experimental results have shown an increase in prediction accuracy from 70.48% to 92.38%.