AHSWN Home · Issue Contents · Forthcoming Papers

M2A-Twin: An MLLM-driven Active Perception Framework for Digital Twin Consistency Diagnosis in IoT Sensor Networks
Zhixian Xu, Chengwei Ma, Ping He, Xiaowei Zhu, Weijian Guan, Qiang Zhang, Zhendong Guan, Qidang Liang, Dongdong Li, Jianqiang Mai, Si Shi and Zhen Tian

Digital Twins are increasingly applied in the lifecycle management of critical infrastructures such as power grids, yet their reliability is often threatened by “digital-physical inconsistency.” To address this challenge, we propose M2A-Twin, a Bayesian decision framework that integrates Multimodal Large Language Models (MLLMs) with active perception under wireless sensor network (WSN) constraints. Unlike passive synchronization approaches, M2A-Twin operationalizes a closed-loop diagnosis cycle: when sensor data indicate high uncertainty, an MLLM-driven policy selects the next sensing action to maximize expected in formation gain while minimizing energy and communication costs.

Simulation experiments demonstrate that M2A-Twin improves diagnostic accuracy (0.90) and uncertainty reduction (0.867 nats), highlighting its potential for reliable and cost-aware Digital Twin maintenance.

Keywords: Digital Twin, Active Perception, Multimodal Large Language Models (MLLM), Bayesian Decision Loop, Information Gain, Sensor Networks, Smart Grids, Consistency Diagnosis