Robust Prototypical Networks for Few-Shot Sensor Fault Diagnosis (MEPN)

Forum topic · 小凯 · 2026-09-15

Summary

Researchers propose Multi-Episode Prototypical Networks (MEPN), a few-shot learning method for industrial sensor fault diagnosis. Standard prototypical networks can produce unstable class prototypes when only one or very few labeled fault examples are available, since each decision depends on a tiny support set. MEPN aggregates prototypes from multiple disjoint support episodes and uses their mean as the final class representative, reducing prototype variance without modifying the encoder architecture. The method is evaluated on the DeFACTO sensor dataset with five-way fault classification, where synthetic bias, drift, spike, and noise faults are injected into real industrial measurements. Across 100 independent runs, MEPN achieves the best accuracy in the per-episode one-shot setting (K=1 sample, aggregating N_agg=10 support episodes), significantly outperforming the single-episode baseline. Under an equal 10-sample support budget, MEPN is statistically indistinguishable from ProtoNet with K=10, confirming that its gains come from prototype accumulation rather than better learning under a fixed budget.

Overview

Field: Machine Learning
Authors: Mohammed Ayalew Belay, Amirshayan Haghipour, Pierluigi Salvo Rossi
Published: 2026-09-15
arXiv: 2609.12287

Abstract

Industrial fault diagnosis often operates with only a handful of labeled fault examples, making few-shot learning attractive for sensor monitoring. Standard prototypical networks are simple and effective; however, their class prototypes may become unstable in the very-low-shot regime because each decision relies on a small support set.

The authors propose Multi-Episode Prototypical Networks (MEPN), which aggregate prototypes from multiple disjoint support episodes and use their mean as the final class representative, reducing prototype variance without changing the encoder architecture.

Evaluation

MEPN is evaluated on the DeFACTO sensor dataset using five-way fault classification, with synthetic bias, drift, spike, and noise faults injected into real industrial measurements.

Key results

Over 100 independent runs, MEPN achieves the best accuracy in the per-episode one-shot setting (K=1 sample, aggregating N_agg=10 support episodes), significantly outperforming the single-episode baseline.
Under an equal 10-sample support budget, MEPN is statistically indistinguishable from ProtoNet with K=10.
This confirms that MEPN's benefit comes from prototype accumulation across episodes, rather than better learning under a fixed sample budget.

Significance

MEPN offers a simple, encoder-agnostic way to stabilize few-shot classification for industrial sensor monitoring when labeled fault data is scarce.

---

*Auto-collected on 2026-09-15*