Neuro-fuzzy System Combined with Clustering Algorithm for Continuous Ankle Angle Estimation from Tibial and Peroneal Electroneurograms in Functional Electrical Stimulation

Electroneurograms (ENGs) from muscle spindles are used to extract joint angle and serves as feedback for functional electrical stimulation. Both structural and black box models have been developed, however, a major obstacle hampering online applications was the lengthy training time of model parameters. The main goal of this study is to develop a practical model with fast learning speed for online applications. In seven anesthetized rabbits, passive ramp-and-hold and sinusoidal-like ankle joint movements with or without electrical stimulation of muscle were performed. Tibial and peroneal ENGs were recorded using two cuff electrodes. A neural-fuzzy system, adaptive-network-based fuzzy inference system (ANFIS) combined with subtractive clustering, was constructed to model the relationship between ankle angle and the ENGs. Input variable selection and input space partition, and a hybrid learning algorithm with the least squares estimate and backpropagation gradient algorithm in a series-parallel configuration was adopted. Simulation results demonstrated that the clustering-based ANFIS could effectively shorten the training time. The performance of the algorithm was satisfactory with both types of ankle movement, although the prediction of sinusoid-like movement was better than that of ramp-and-hold movement, which meant that the ANFIS is more feasible for swing motion, such as human walking. Only a slight decrease of performance during muscle stimulation indicates that the proposed algorithm of angle estimation was workable under stimulus interference. In conclusion, the proposed method is fast enough for online FES application and the performance is acceptable even under electrical stimulation of muscles.