Rule-based control of walking by using decision trees and practical sensors

We developed a method for designing a rule-based control (RBC) that is convenient for implementation within a practical functional electrical stimulation (FES) system. The RBC uses data from accelerometers and force sensing resistors (FSRs) to estimate ON.OFF states of muscle activity. The rules are obtained in the form of decision trees. Therefore, they are human readable, interpretable and convenient for real time control. The method was demonstrated by generating rules for control of a four channel FES system that is suitable for assisting gait of an individual with hemiplegia. Experimental data were collected during walking of an able-bodied subject. The subject was instrumented with three dual-axis accelerometers (right leg) and four FSRs (both feet). Electromyography (EMG) was recorded from the muscles acting at the knee and ankle joints. A binary signal denoting ON.OFF muscle state was obtained by thresholding of the recorded EMG. The acquired data were divided into training and validation data sets. A classification and regression tree for estimation of the muscle state was determined for each muscle by using the training data and then tested over the validation data. Overall average and maximum timing errors in the estimated muscle activity were 49 ms (≈ 5%) and 157 ms (≈ 15%), respectively.