Taguchi Design of Experiments Application in Robust sEMG Based Force Estimation

This paper investigates the impact of the parameters that affect the accuracy of the force estimation from sEMG signals, including signal acquisition factors, pre-processing and training ones. It offers a procedure for developing a reliable estimation approach to deal with uncertainties, such as the signal deviation while performing various daily tasks using the hand. For doing this, the Taguchi design of experiments (DOE) approach is used to determine appropriate levels of the factors to decrease the regression error. Factors such as the number of electrodes placed on the forearm and the arm, extracted features, the cropping window length and the training regularization term have been categorized as either controllable or uncontrollable in the DOE table. The experiments are conducted on four subjects who perform six different tasks. The L225 mixed-level orthogonal array is used to specify the levels of factors in each experiment. The orthogonal array drastically reduces the required number of executions compared to a full-factorial analysis. Using the Minitab software, the signal-to-noise ratios (SNR) are calculated to determine the optimum levels and significance of the factors. Results indicate that the number of forearm electrodes and their placements are the most influential factors. Moreover, the SNR delta for including the arm biceps muscle is about 0.66, which considering its placement difficulties, it does not justify its additional expense.