Metric-based meta-learning model for few-shot fault diagnosis under multiple limited data conditions

The real-world large industry has gradually become a data-rich environment with the development of information and sensor technology, making the technology of data-driven fault diagnosis acquire a thriving development and application. The success of these advanced methods depends on the assumption that enough labeled samples for each fault type are available. However, in some practical situations, it is extremely difficult to collect enough data, e.g., when the sudden catastrophic failure happens, only a few samples can be acquired before the system shuts down. This phenomenon leads to the few-shot fault diagnosis aiming at distinguishing the failure attribution accurately under very limited data conditions. In this paper, we propose a new approach, called Feature Space Metric-based Meta-learning Model (FSM3), to overcome the challenge of the few-shot fault diagnosis under multiple limited data conditions. Our method is a mixture of general supervised learning and episodic metric meta-learning, which will exploit both the attribute information from individual samples and the similarity information from sample groups. The experiment results demonstrate that our method outperforms a series of baseline methods on the 1-shot and 5-shot learning tasks of bearing and gearbox fault diagnosis across various limited data conditions. The time complexity and implementation difficulty have been analyzed to show that our method has relatively high feasibility. The feature embedding is visualized by t-SNE to investigate the effectiveness of our proposed model.