Adaptive filtering for microelectromechanical inertial sensors using empirical mode decomposition, Hausdorff distance and fractional Gaussian noise modeling

We present a denoising technique for inertial sensors based on Empirical Mode Decomposition (EMD). The method uses the fractional Gaussian noise (fGn) model to accurately quantify errors, addressing the complex nature of sensor noise encompassing both Gaussian white and colored components. A modified Hausdorff Distance technique is integrated with the fGn energy model to categorize the resulting Intrinsic Mode Functions (IMFs) into dominant noise, dominant information, and mixed components. The relevant signal IMFs are selectively chosen, while the mixed IMFs undergo hard thresholding before recombination for optimized signal reconstruction. Comparative assessments with conventional wavelet transform, Savitzky Golay, and Moving Average filtering methodologies validate our approach, including analysis via Allan Variance (AVAR), realistic dynamic scenario simulations, and an inertial navigation algorithm implementation. Leveraging data from Microelectromechanical System sensors, our noise removal technique outperforms the other methods in all these analyses. We observe nearly complete suppression of white noise and an approximate 5% reduction in colored noise. Additionally, a 90% decrease in mean squared error and a 10% extension in GNSS-denied navigation duration are noted relative to the original data.