112 / 2021-07-30 20:32:23

An Enhanced Intelligent Fault Diagnosis Method to Combat Label Noise

Special Session > SS4 - Weak Signal Processing and Incipient Fault Detection

Despite the excellent performance achieved by deep learning-based fault diagnosis approaches, however, the

fault diagnosis under noisy labels is remaining a problem that need to be solved. Due to its complex network structure, the deep learning model can easily fit the noisy labels, which degrades the fault classification performance of deep model. Aiming at the aforementioned issue, this paper proposes a novel noisy label

learning method. Based on Gaussian mixture model, Coteaching technique and semi-supervised learning strategy, the noisy labels are filtered out, refined and reassigned subsequently based on the estimated clean possibility and model prediction. In order to combat the negative effects caused by noisy labels, a more robust training strategy and training goal are presented. The extensive experimental results show the superiority of the proposed method compared with traditional approaches.