A Theoretically Grounded Multichannel Fusion Framework Integrating Dual-Metric Trust and Uncertainty Rectification for Intelligent Diagnosis
Xinming Li, Zhikang Gao, Jinrui Zhang, Kehui Zhu, Jiawei Gu, Wenhan Lyu, Yanxue Wang · IEEE/ASME Transactions on Mechatronics · 2025
Fault diagnosis systems in real-world applications must increasingly integrate heterogeneous sensor data under uncertainty, noise, and dynamic operational conditions. However, many existing fusion methods lack solid theoretical foundations, demonstrate limited robustness, and provide minimal interpretability, thereby limiting their applicability in safety-critical mechatronic environments. To address these challenges, this study introduces generalization error bound-guided dynamic signal integration (GDSI), a robust and interpretable diagnostic framework. Built upon generalization error theory, GDSI incorporates a covariance-constrained objective to explicitly link fusion weights with generalization performance. A dual-metric trust mechanism, based on intrinsic confidence and cross-source consistency, enables adaptive and interpretable multichannel fusion. In addition, an uncertainty-aware rectification strategy dynamically modulates fusion behavior in response to signal reliability changes, improving robustness under noisy and nonstationary conditions. Experiments on multiple datasets demonstrate that GDSI consistently enhances diagnostic accuracy, generalization ability, and robustness compared to existing fusion methods. By integrating theoretical rigor with practical adaptability, GDSI offers a deployable and generalizable solution for intelligent condition monitoring in complex mechatronic systems.