A novel assessable data augmentation method for mechanical fault diagnosis under noisy labels

Abstract

Data augmentation technology has achieved great success to expand the training set for several years. As a representative technology, generative adversarial network and its variants are widely applied in many data augmentation tasks. But the quality of training samples is rarely considered. In this paper, a novel assessable data augmentation named ADA is proposed for mechanical fault diagnosis under noisy labels. First, a sample quality assessment procedure including assessment model construction, approximate calculation based on influence function and screening decision is presented. Thereby, the optimized training set can be obtained. Then, the WGAN-gp model can be established based on the optimized training set and the data augmentation can be accomplished. Finally, a classifier can be trained with the expanded training set and achieve the task of fault diagnosis. The results of two experiments show that the proposed ADA method can effectively improve the fault diagnosis accuracy for various classifiers.