An EMD threshold de-noising method for inertial sensors

• We point out the disadvantages of the widely-used wavelet threshold de-noising method.
• We use fractional Gaussian noise to model inertial sensor errors to consider the time-correlated colored noise.
• A robust estimator is used to estimate the standard deviations of noise in the first tow IMFs.
• The variance relation among the IMFs is derived by the power spectral density property of fGn.
• The thresholds of this novel EMD threshold de-noising are order-dependent.

Random errors of inertial sensors are key factors in influencing the performance of Inertial Navigation System (INS). Based on underlying white noise model, classical wavelet threshold de-noising method is incapable of eliminating colored noise. Since time-correlated colored noise is predominant, fractional Gaussian noise (fGn) is utilized to model sensor errors and the Hurst parameter of fGn is estimated by the periodogram method. Variances of the noise in Intrinsic Mode Functions (IMFs) decomposed by Empirical Mode Decomposition (EMD) are analyzed. The standard deviations of noise in the first tow IMFs are estimated by a robust estimator, and then the noise variances in other IMFs can be obtained after the variance relation among the IMFs decomposed from fGn is derived. Noise thresholds of IMFs are estimated through the obtained variances and an EMD threshold de-noising method using order-dependent thresholds is established. The method is firstly verified by a simulation example and then applied in INS and compared with wavelet de-noising method. Results show that wavelet threshold de-noising is poor at suppressing colored noise while EMD threshold de-noising is effective on reducing sensor errors due to its close association with proper noise model.