A Two-Stage Semi-Supervised Learning Framework for Fault Diagnosis of Rotating Machinery

Abstract

In recent years, the deep learning (DL)-based fault diagnosis has flourished and achieved outstanding progress relying on sufficient and accurate labeled training samples. In practice, sample labeling is a complex and error-prone process, which means noisy labels are likely to exist in health condition datasets. This article aims to transform fault diagnosis with noisy labels into a semi-supervised learning (SSL) procedure and proposes a two-stage fault diagnosis framework. First, a backbone network employing the convolutional gated recurrent unit (ConvGRU) is constructed to extract temporal and spatial information simultaneously from multi-sensor signals and classify health conditions. Second, a high learning rate is applied to accomplish the initial learning for avoiding fitting to noisy labels. Finally, three regularization terms and correction labels are introduced in the second stage, where network parameters and correction labels are jointly optimized through back-propagation. The experiments are conducted on not only synthetic noisy label conditions, but also simulated real-world noisy label conditions. The results indicate that our framework significantly outperforms other state-of-the-art frameworks.