A Novel Domain-Adaptive Multiscale CNN-Transformer for Low-Cost MEMS-IMU Denoising

Low-cost Micro-Electro-Mechanical Systems Inertial Measurement Units (MEMS-IMUs) are widely employed in intelligent navigation systems. However, they are susceptible to complex noise, leading to a rapid degradation in navigation accuracy. In this study, a MEMS-IMU denoising method based on a Domain-Adaptive Multi-Scale CNN-Transformer (DAMCT) is proposed. First, the proposed method is trained on simulated MEMS-IMU data (source domain) by leveraging multi-scale convolution for spatial feature extraction and a Transformer architecture for temporal sequence modeling. Second, pseudo-labels are generated through inference on real measurement data (target domain), and the model is further fine-tuned using a weakly supervised strategy to enhance its generalization capability in real-world scenarios. This enables effective denoising of low-cost MEMS-IMU data without relying on noise-free labeled datasets. Finally, DAMCT achieves superior denoising performance compared with three state-of-the-art methods, yielding average reductions in position and velocity Root Mean Square Error (RMSE) of 68.0%, 82.9%, 60.1% and 22.4%, 80.6%, 43.9% along the East, North, and Up directions, respectively. These results provide a solid foundation for the future development of intelligent navigation systems.