Implementation of a Particle Filter to Estimate Torque from Electromyography

Abstract This paper proposes the use of a particle filter to estimate the torque produced by the biceps based on the recorded electromyography (EMG). The aim was to estimate a person’s voluntary effort, so that it can be later implemented in an assist-as-need control scheme for an exoskeleton for stroke rehabilitation, where torque estimates from the dynamics may be susceptible to error and noise. The particle filter successfully improved the prediction from a simplified Hammerstein EMG-torque model. The intended application is for a dynamic contraction, however, analysis in this study was considered for quasi-static movement to enable comparison to previous methods. The prediction accuracy depends on model variances, the dependence on previous predictions, and the prediction and resampling rates. When using an unfiltered and noisy sensor model, the average RMS error is 8.80%, which is comparable to previous methods. The particle filter is a promising method for its desired application.