Multi-Episode Prototypical Networks Improve Few-Shot Sensor Fault Diagnosis

Summary

Few-shot sensor monitoring is useful when only a small number of labeled fault examples are available, but standard prototypical networks can form unstable class prototypes from a very small support set. The paper introduces Multi-Episode Prototypical Networks (MEPN), which compute prototypes from multiple disjoint support episodes and average them into the final class representative. The method leaves the encoder architecture unchanged and is intended to reduce prototype variance. On the DeFACTO sensor dataset, the authors test five-way fault classification with synthetic bias, drift, spike, and noise faults injected into real industrial measurements. Across 100 independent runs, MEPN in the per-episode one-shot setting used 1 labeled example per class in each of 10 aggregated support episodes; the abstract reports a performance value, but the supplied text contains an unresolved placeholder instead of the percentage. With an equal total support budget of 10 samples, MEPN and a single ProtoNet using 10 shots per class were statistically indistinguishable. The authors interpret this result as evidence that the gain comes from accumulating prototypes across episodes, rather than from better learning under a larger fixed support set.