Liquid-transformer temporal self-supervised network for few-shot class-incremental fault diagnosis of servo mechanisms

Abstract

Electric servo mechanisms are critical actuators in aerospace equipment, where emerging fault types continuously appear under varying operating conditions. Few-shot class-incremental diagnosis is therefore constrained by two coupled challenges: highly nonstationary vibration responses and extremely limited labelled samples, both of which weaken temporal representation learning and aggravate catastrophic forgetting in conventional models. To address these issues, a Liquid-Transformer Temporal Self-Supervised Network (LTTSNet) is proposed for continual recognition of both base and emerging fault classes. Its backbone integrates a Liquid Neural Network (LNN) with a Lightweight Transformer (LTransformer), in which the LNN captures local transient dynamics through learnable time constants, whereas the LTransformer models long-range cross-cycle dependencies. In the base stage, a simple framework for contrastive learning of visual representations is employed to learn invariant representations from scarce unlabelled signals by contrasting augmented views. Pseudo-labels are generated via nearest-neighbour clustering under a cosine-similarity threshold and are jointly trained with labelled samples, thereby introducing latent new-class information before incremental updates. In the incremental stage, attention-weighted prototype estimation and one-step gradient prototype distillation are jointly employed to refine new-class prototypes. The backbone is kept frozen, and only the classifier head is updated, enabling rapid adaptation to new classes while preserving old-class discrimination. Experiments on a laboratory electric servo mechanism fault dataset and the Case Western Reserve University bearing dataset demonstrate that LTTSNet delivers significantly improves overall accuracy, new-class recognition, and forgetting suppression under cross-condition few-shot settings with both single and compound faults.