Advanced Electromyogram Signal Processing with an Emphasis on Simplified, Near-Optimal Whitening
Estimates of the time-varying standard deviation of the surface EMG signal (EMGσ) are extensively used in the field of EMG-torque estimation. The use of a whitening filter can substantially improve the accuracy of EMGσ estimation by removing the signal correlation and increasing the statistical bandwidth. However, a subject-specific whitening filter which is calibrated to each subject, is quite complex and inconvenient. To solve this problem, we first calibrated a 60th-order “Universal” FIR whitening filter by using the ensemble mean of the inverse of the square root of the power spectral density (PSD) of the noise-free EMG signal. Pre-existing data from elbow contraction of 64 subjects, providing 512 recording trials were used. The test error on an EMG-torque task based on the “Universal” FIR whitening filter had a mean error of 4.80% maximum voluntary contraction (MVC) with a standard deviation of 2.03% MVC. Meanwhile the subject-specific whitening filter had performance of 4.84±1.98% MVC (both have a whitening band limit at 600 Hz). These two methods had no statistical difference. Furthermore, a 2nd-order IIR whitening filter was designed based on the magnitude response of the “Universal” FIR whitening filter, via the differential evolution algorithm. The performance of this IIR whitening filter was very similar to the FIR filter, with a performance of 4.81±2.12% MVC. A statistical test showed that these two methods had no significant difference either. Additionally, a complete theory of EMG in additive measured noise contraction modeling is described. Results show that subtracting the variance of whitened noise by computing the root difference of the square (RDS) is the correct way to remove noise from the EMG signal.