Embodied Intelligence-Driven Active Perception and Detection Strategy Learning for Display Defects in Electric Metering Devices

Abstract

Electric metering display inspection is a critical task in power equipment maintenance, yet practical field images often suffer from reflection, blur, occlusion, low illumination, and subtle segment-level defects. Existing passive inspection methods usually rely on fixed-view image recognition and cannot determine whether additional observations are necessary when visual evidence is unreliable. To address this problem, this paper proposes UAP-DSL, an uncertainty-guided embodied active perception and defect strategy learning framework. The method integrates quality-aware state encoding, uncertaintydriven action selection, cross-view feature fusion, and visual–logical defect analysis to jointly optimize detection accuracy and sensing cost. Experiments on MVTec AD, VisA, BTAD, and MVTec LOCO AD show that UAP-DSL consistently outperforms PaDiM, PatchCore, FastFlow, and EfficientAD in I-AUROC, P-AUROC, PRO, and F1-score while requiring fewer sensing steps. The results demonstrate that active perception and logical consistency modeling can improve robust defect inspection under complex visual conditions.