Poster Session 4, 2:15 PM - 3:00 PM: Campus Center Auditorium [A24]

Modeling and Machine Learning–Based Compensation of Rotor-Induced Vibration and Magnetic Interference in UAV IMU Systems

Presenter: Neelesh Kumar Jana

Faculty Sponsor: Xingda Chen

School: UMass Amherst

Research Area: Electrical and Computer Engineering

ABSTRACT

The integrity of Inertial Measurement Unit (IMU) data is fundamental to the reliability of autonomous systems. This study focuses on acquiring and processing high-fidelity IMU data to support robust multi-modal sensor fusion architectures for platforms operating in environments where Global Navigation Satellite Systems (GNSS) may be unavailable, unreliable, or intentionally disrupted. Achieving reliable navigation under such conditions is critical for next-generation surveillance systems and for meeting stringent anti-radio jamming and spoofing objectives.

Rotor operation in multirotor unmanned aerial vehicles (UAVs) introduces two primary disturbance mechanisms: mechanical vibration that corrupts accelerometer and gyroscope measurements, and electromagnetic interference generated by brushless DC (BLDC) motors and high-current power electronics that distort magnetometer readings. These disturbances scale with rotor throttle and degrade sensor fusion performance. We systematically characterize IMU degradation across controlled throttle sweeps (0–100%) using a PX4-based multirotor platform. Time-domain variance and frequency-domain spectral analyses quantify vibration-induced harmonic amplification, while magnetometer magnitude stability and heading consistency assess throttle-dependent magnetic distortion.

To address these challenges, we propose a data-driven compensation framework centered on an automated data generation pipeline. A cyber-physical testbed leveraging a precision robotic arm induces repeatable interference patterns under controlled conditions representative of GNSS-denied operation. Complementary on-drone sensing modalities, including high-frame-rate optical cameras, are used to infer environmental interference signatures. Machine learning models capture non-linear noise distributions and are integrated into a commercial Extended Kalman Filter (EKF), dynamically refining covariance and state transition parameters. Results demonstrate structured, predictable degradation, enabling adaptive correction strategies that significantly enhance state estimation robustness in high-disturbance environments.