Virtual–Real Fusion-Based Transfer Learning With Limited Data for Gearbox Fault Diagnosis

Abstract

An intelligent fault diagnosis of the gearbox is essential to guarantee the reliability of rotating machinery. However, a huge number of historical data are needed in the modeling procedure. In a real industry scenario, it is generally difficult to obtain enough data because fault samples are limited or unavailable. The fault diagnosis model developed based on a limited dataset often suffers from overfitting and poor generalization. In this article, a virtual–real fusion-based model is proposed to implement gearbox fault diagnosis with limited data. First, a digital physical model is proposed to generate virtual signals for data augmentation. Second, an adversarial local domain adaptation network (ALDAN) is proposed to transfer knowledge learned from simulated samples to real samples in industrial applications. Third, a dynamic convolutional network is employed as feature generator to learn features from the simulated signals for fault diagnosis of the real signals. A gearbox fault diagnosis case is used to validate the effectiveness of ALDAN. The results illustrate that ALDAN can learn the transferable features from virtual and real signals. It can significantly improve the fault diagnosis performance with limited samples.