Publication

Hyperautonomy Artificial Intelligence Lab

크로스 어텐션 기반 라우팅 캡슐 네트워크를 활용한 모터 고장 진단 방법

본문

Conference
한국PHM학회 2025년도 정기학술대
Author
임주현, 김성종, 윤헌준, 윤병동
Date
2025-06-25
Presentation Type
구두

Abstract

Capsule networks offer promising avenues for integrating multi-modal signals in fault diagnosis tasks. This study introduces a novel cross-attention mechanism within capsule networks to enhance feature interaction between vibration and  current signal modalities for motor fault classification. Specifically, we propose a dual-branch architecture where capsules are initialized from vibration and current signals separately. Cross-attention is applied to dynamically align capsules between modalities, facilitating effective information exchange. We validate our approach on a dataset comprising constant-speed motor operation scenarios with five fault modes, demonstrating superior classification performance compared to baseline models. Our findings highlight the efficacy of cross-attention routing in enhancing the interpretability and accuracy of multi-modal capsule network for industrial fault diagnosis applications.