A Data-driven Denoising Approach to Compensate for Nonlinear Errors in MEMS IMU Signals

Microelectromechanical systems (MEMS) inertial measurement unit (IMUs) provide three-dimensional angular velocity and acceleration sequence signals, commonly used in attitude estimation and navigation applications. However, raw MEMS IMU signals often exhibit nonlinear errors that accumulate over time, compromising overall system accuracy. This paper proposes a data-driven denoising method designed to estimate and compensate for the nonlinear errors inherent in MEMS IMU signals. The approach leverages a temporal convolutional network (TCN)-based model to learn error features from ground-truth data on attitude and velocity. The trained model utilizes MEMS IMU sequence signals to estimate IMU errors during testinging, thereby reducing integration errors in orientation, velocity, and position. Validation is performed using a public integrated navigation dataset containing various grades of MEMS IMUs. Experimental results demonstrate that the denoised consumer-grade MEMS IMU signals achieve industrial-level accuracy in attitude, velocity, and position estimation. This confirms the method's effectiveness in compensating for errors in MEMS IMU signal sequences.