Abstract
An important demand in the machinery environment is the frequent handling of variant faults occurring during active performance of motors and gearbox and through its vibrational signals. Accordingly, effective intelligent approaches detecting and treating causes of low productivity, reliability and higher maintenance cost have been introduced. A huge attention in recent years recommended deep learning and reports higher capabilities in this context. However, one limitation of using the deep learning is that its performance (in terms of evaluation matrices and/or over-fitting) is highly connected to the availability of large number of training samples as it learns by examples. This paper tackles this problem, proposing novel generative adversarial networks based augmentation method to balance the dataset and hence increase generalizability of the machine learning model. The generator of the generative adversarial networks method provides more samples with main objective of class balancing. Its discriminator recognizes the real and fake samples and assigns classes to the inputs as well. Accompanying deep framework based on the pervious augmentation method and a two subsequent unsupervised sparse autoencoders is also proposed to develop accurate intelligent identification model for motor and gearbox multi-faults with 10 classes. Precision, recall, and F1 measures proved the validity of the proposed augmentation when evaluated on the dataset of gear and motor signals and enhanced the class balancing. In addition, it enhances results of deep learning and reports identification accuracy of 93.7%.