Machine learning techniques for the control of FES-assisted locomotion after spinal cord injury

Functional electrical stimulation (FES) has been used as a substitution for the missing neural excitation of the paralyzed muscles after spinal cord injury. FES-assisted walking with preset stimulation patterns is usually controlled by a therapist or the subject using manual controls mounted on the handgrips of a walking aid. Automation of the switching control can be done by designing a rule-based control algorithm which will replace the decision making process the person uses to control the stimulation manually. These rules are usually designed by intuitive 'hand-crafting' and by applying them on a set of sensory feedback signals. This process has to be repeated for each subject due to highly specific disabilities resulting from physically similar injuries.;In this thesis a method is proposed, developed, and applied for automatic generation of control rules, which may provide a much faster evaluation of new subjects than the "hand-crafting" method. The rules are extracted from a set of sensory feedback signals and stimulation control signals recorded during FES-assisted walking controlled by a skilled therapist or the subject. The rule-generation method is evaluated using two different machine learning techniques, Adaptive Logic Networks (ALNs) and Inductive Learning (IL). Very fast training and high generalization of both techniques justified the design of the integrated control system (ICS). The ICS, currently based on ALNs, provides an efficient tool to acquire sensory and control signals, to process these signals, to train the ALNs in mapping the control function, to test the trained ALNs and to use them for control signal generation in real-time control of the FES-assisted walking of subjects with incomplete spinal cord injury. The IL technique was also evaluated in rule-generation for control of walking of subjects with complete spinal injury and its potential for cloning the subject's skill in switching the stimulation was demonstrated. In addition, ALNs were evaluated for continuous control of single joint flexion-extension, based on signals recorded from natural sources, such as nerves and muscles of cat's hind limb. Through experimental work it has been demonstrated that both techniques are able to generate control rules quickly and to generalize, not only over daily subsequent walking sessions but also over the sessions occurring several days after the training This provides a good basis for design of robust control systems for FES-assisted walking.