Learning a Normal World Model for Few-Shot Boundary-Calibrated Abnormality Detection

2026-06-20

AI Summary

This work addresses the challenge of anomaly detection under scarce anomaly labels and the inadequacy of binary labels in capturing degrees of deviation. The authors propose a few-shot anomaly detection method grounded in a normal-world model: it first learns the intrinsic system structure using abundant normal data and then calibrates the boundary of normality with only a few anomalous samples. The core innovation lies in a hypergraph entropy–driven normal-world model that represents multivariate time-series windows as context-conditioned hypergraphs. By integrating temporal prediction surprise, hypergraph consistency surprise, and latent manifold deviation, the method establishes an anomaly scoring mechanism centered on “normal-world energy.” Evaluated on the NASA C-MAPSS dataset, the approach achieves an AUROC of 0.9983 on the FD004 subset under zero- and few-shot settings. Ablation studies confirm its ability to identify healthy states, track degradation trajectories, and detect disruptions in variable couplings.

Abstract

Abnormality detection in complex systems faces two practical barriers: abnormal labels are scarce, and binary labels do not quantify how far an event has departed from normal behavior. We study a normal-world modeling formulation for this setting. Instead of learning a large and incomplete space of abnormal classes, the model learns the normal world from abundant normal events and uses a few abnormal examples only to calibrate the boundary of normality. We instantiate this idea as a Hypergraph Entropic Normal-World Model. The model represents multivariate sensor windows as context-conditioned hypergraphs, where hyperedges capture high-order relations among groups of variables. It then defines abnormality by an entropy-aware normal-world energy that combines temporal prediction surprise, hypergraph consistency surprise, and latent normal-manifold departure. On the NASA C-MAPSS turbofan degradation benchmark, the proposed full energy achieves strong zero-shot and few-shot performance across all four subsets and reaches AUROC 0.9983 on FD004, the most complex setting with multiple operating conditions and fault modes. Beyond standard detection metrics, we introduce mechanistic validation tests to probe whether the energy encodes normal-world structure rather than a superficial input-output mapping. The learned energy accepts unseen healthy engines, increases along degradation trajectories, and sharply penalizes context-mismatched cross-variable coupling breaks. These results suggest that normal-world energy can serve as an anomaly score, a graded risk measure, and a testable representation of normal system behavior under severe abnormal-label scarcity.

Problem

• abnormality detection
• few-shot learning
• normal-world modeling
• anomaly scoring
• label scarcity

Innovation

• normal-world modeling
• few-shot anomaly detection
• hypergraph representation
• entropy-aware energy
• boundary calibration