AUV Few-Shot Class-Incremental Fault Diagnosis Method with Distillation Projection Constraint Embedded in Neural Collapse Representation

Objectives To address few-shot overfitting and catastrophic forgetting in AUV fault diagnosis under scarce samples of new faults and continually emerging fault categories, this paper proposes a few-shot class-incremental fault diagnosis method for AUV multivariate time-series data with distilled projection constraints embedded in neural collapse representations. Methods The simplex equiangular tight frame from neural collapse phenomenon is adopted to fix the classifier prototypes, which are frozen throughout training to alleviate decision-boundary conflicts between base and novel faults. A residual skip-connected transitional projection layer is constructed, with an inductive bias toward additive local corrections of frozen backbone features. In addition, logit-based knowledge distillation oriented to base classes is introduced to explicitly constrain updates of the projection layer, thereby suppressing forgetting of base-class knowledge during incremental learning and improving the joint recognition accuracy of base and novel faults. Results Experiments are conducted on the Haizhe AUV tank-test dataset and the Qianlong-II ocean voyage dataset, covering controllable laboratory tank conditions and complex real-sea operations. Results show that the proposed method achieves multi-session micro-average accuracies of 86.14% and 77.65% on the two datasets, respectively, demonstrating favorable anti-forgetting capability and few-shot adaptability under a weak-replay setting. Conclusions This work is among the first to investigate AUV multi-sensor time-series fault diagnosis under few-shot constraints and class-incremental dynamics, and provides a solution embedding distilled projection constraints into neural collapse representations, offering technical support for continual updating of intelligent fault diagnosis models during long-term AUV operations.