Black box modeling of Paralyzed Arm for FES-based Reaching Movement Control in SCIPatients

One of main the issues in achieving to a successful FES control is using an as much as possible accurate model of the under electrical stimulation system so that it can adequately indicate the system behavior. Classical computational models that are commonly used for this purpose have a reductionism nature; so they cannot consider the interaction existed in biological systems. Considering these restrictions, recently behavioral black box models are mostly used. These models focus on input/output dynamic, which is certainly the necessary modeling information for control design; thus the system is dealt with as a whole, which has hidden the interactions between components inside. Such a model has notbeen presented for elbow angle movement so far. Therefore in this study, we have been to present and verify a black box model of elbow joint movement in the transverse plane, forreaching movement control in people with C5/C6 SCI using dynamic neural networks, including time-delayed feedforward and recurrent networks. Extreme flexibility of time-delayed feedforward architectures was obtainedin a 2 layer structure including 5 hidden neurons and using 1.25s of history of input with performance indexes of 89.89% & 4.85% for cross correlation coefficient and normalized mean square error respectively. The best recurrent network with NARX architecture and equal history of input & output was also occurred in a 2 layer structure having 12 neurons in the hidden layer and using 0.1s of history, with performance indexes of 89.89% & 4.85% for cross correlation coefficient and normalized mean square error respectively. Comparison between best results of training using feedforward and recurrent networks, clearly illustrates both qualitative and quantitative excellency of the latter one in identification of the under-study system.