Artificial neural network based iterative learning control for stroke rehabilitation

An artificial neural network (ANN) is combined with gradient descent to form a model-free iterative learning control (ILC) approach than can be applied to a wide range of nonlinear discrete-time systems. The ANN is recursively trained on the entire set of past data collected from the system and uses a passivity condition to determine when the ANN can be used to compute the next ILC update, or if an identification test is needed. Convergence properties are established alongside design selections that ensure the passivity condition is fulfilled. By minimising the reliance on identification tests, this methodology is substantially faster than existing model-free ILC algorithms. It is tested on a key stroke rehabilitation problem using functional electrical stimulation (FES) for hand/wrist tracking. Experimental results using the new ILC approach with eight participants show that three hand/wrist references can be tracked using an average of 56% fewer experimental inputs compared with the most accurate previous approach. As the first approach to combine ILC and machine learning in upper limb rehabilitation, the results demonstrate how their combination addresses their individual deficiencies.