Manifold-informed Meta-Learning for Out-of-Distribution Detection and Classification in Small-Scale Industrial Datasets

Neural networks (NNs) have demonstrated remarkable success in various manufacturing applications, but often require large labeled datasets and fail to recognize inputs from novel or unseen classes. This limitation becomes especially critical in small-sample manufacturing scenarios, where only a few labeled samples are available per task. In this work, we propose a novel meta-learning framework that simultaneously identifies out-of-distribution (OOD) samples and classifies in-distribution (ID) samples using limited labeled data. Our method leverages class-specific generators capable of modeling fine-grained distributions of ID data. Based on these generators, we synthesize two complementary types of OOD samples: those located just beyond the class manifold, and those located near class decision boundaries on the manifold. These artificial samples guide the proposed meta-learning model in constructing tight class boundaries, significantly enhancing its ability to detect near-OOD samples while preserving strong classification performance on ID data. Empirical results on multiple benchmark manufacturing datasets show that our approach outperforms existing meta-learning and OOD detection baselines, demonstrating improved generalization and robustness in limited data scenarios. The source code will be publicly available upon paper publication.