Improved Representations for Continual Learning of Novel Motor Health Conditions through Few-Shot Prototypical Networks

Intelligent machine condition monitoring (CM) for automatic fault diagnosis relies on data-driven algorithms to characterize machine health for predictive maintenance activities on smart factory floors. Since data collection can be expensive, CM data sets may not cover all the possible fault conditions, necessitating that CM algorithms continually learn new conditions. State-of-the-art CM research has focused on detecting unknown conditions rather than integrating unknown conditions into future predictions. Therefore, CM-ready Continual Learning (CL) solutions should learn to classify new conditions and use improved representations that minimize the need for future fine-tuning. Meta-learning approaches like Few-Shot Prototypical Networks (FSPN) regularize base-task learning to find these more generalizable representations. Experiments on a motor data set demonstrate that FSPN with only 5 or 10 examples of the novel fault consistently outperforms static, fine-tuning, and Elastic Weight Consolidation (EWC) approaches for CL, increasing the overall accuracy by up to 19 points (53% to 72%). Compared to recent FSPN work for image classification, these results show that FSPN may be advantageous for CM due to the limited class diversity of CM data sets. Future work should extend the FSPN architecture to include open set recognition and quantitatively analyze varying numbers of base-task classes.