Filling generative adversarial network: a novel intelligent machinery diagnostic method towards extremely limited data

Machinery health monitoring and diagnostics in the presence of extremely limited data is challenging in industrial scenarios, which may result in inaccurate health assessments and misguide maintenance decisions. Though generative adversarial networks (GANs) provide an effective solution for data augmentation, the current GANs-based methods are still incapable to resolve the compromised diagnostic performance issue caused by low-quality generated samples, unstable training, mode collapse, and suboptimal sample expansion strategy when handling extremely limited data. To address these issues for intelligent machinery diagnostics, this study proposes a novel data-filling approach based on filling generative adversarial network (FGAN). The presented FGAN model includes a data prediction module, a similarity discrimination module, and a data filling module, which can predict and fill in the subsequent data values based on a small number of original signals, thereby effectively generating high-quality samples and ensuring the data integrity even in the signal under-sampling scenarios. Moreover, a novel regularization term is proposed to integrate into the loss function of FGAN to avoid the overfitting issue of model training. Finally, experiment validations of two challenging machinery fault datasets considering the extremely limited data have been conducted to prove the efficiency and advantages of our FGAN method for intelligent diagnostics under extremely limited data. Detailed experimental results have verified that our FGAN method yields outstanding diagnostic accuracy and outperforms several advanced methods even in extremely limited data scenarios.