Fuzzy Torque Estimation During Knee Extension with LSTM Neural Network and sEMG Signals

Advanced controls such as fuzzy control and learning algorithms such as long short-term memory (LSTM) and recurrent neural networks (RNNs) are suitable for rehabilitation tasks with the Assist as needed strategy. Fuzzy control has the advantage of implementing complex systems without an exact mathematical model. RNN is able to estimate from noisy signals such as biological signals. The main difficulty in developing AAN systems is the torque estimation provided by the user during rehabilitation. The state of the art proposed to use surface electromyographic sEMG signals to perform this estimation. However, obtaining a numerical value of the exact torque is difficult because they are noisy signals. The fuzzy control advantage is it does not require exact values so we propose an LSTM to estimate the torque as a linguistic variable: fuzzified torque. We assume that the estimation as a linguistic variable performs better than the torque estimation as a value in N m. The expert system acquires real-time data from a position sensor, an accelerometer, and sEMG signals from the muscles. We made two torque estimates for the conventional and the proposed process and compared them. In the conventional process, we estimate the torque as a numerical value, which is then fuzzified. In the proposed process, we estimate torque as a linguistic variable. The proposed process showed metrics between 0.6796 and 0.9798 in the coefficient of determination, which was higher than the obtained in the conventional process between 0.5331 and 0.9861.