A Low-Cost Semi-Active Orthosis and Neural Network Based Prediction System for Alleviating Gait Degradation

Exoskeletons are one method by which a patient's ability to walk can be maintained or even rehabilitated in the face of pathologic gait disorders, however in many cases more advanced, active exoskeletons suffer from being prohibitively expensive for the average user. The muscular imbalances caused by Cerebral Palsy, Strokes, and such conditions can cause excessive strain upon musculature and bone over time, and considerably increase the risk of osteoporosis, pathological gait issues, and stunted growth. As a result, sufferers of such conditions may begin expressing pathologic gait disorders or lose the ability to walk at all should these issues continue. The cost of exoskeleton-based treatment however imposes great strain on a medical system to aid in funding these devices for affected individuals or limits the time such individuals have to access these devices, such as within set "rehabilitation sessions". As neither of these cases are ideal, there is an opportunity to explore methodologies that seek to reduce the financial impact of these devices. This could be achieved by creating a Low-Cost Orthosis that may aid in reducing gait degradation in between such rehabilitation sessions whilst remaining as inexpensive as is feasible without seriously reducing the device's effectiveness.

To this end, this Thesis seeks to propose and test the effectiveness of a Low-Cost Exoskeleton Methodology. It will in so doing examine the structure, actuation methods, and control systems of other Exoskeleton Implementations for potentially innovative solutions to maintain effectiveness even within considerable limitations. This Low-Cost Methodology consists of analysing the cost of exoskeleton components from a perspective of not only Financial cost, but also of Power Consumption, Weight, and the Effectiveness of key exoskeleton components such as the actuator, sensor, and control system.

This study then puts this methodology into practice by outlining the creation of a simple exoskeleton design that follows the principles of the Low-Cost Methodology. This exoskeleton makes use of a rotary electric motor driven series elastic actuator and is controlled by an embedded microcontroller implemented Neural Network capable of real-time predictions of the wearer's knee angle into the future with a level of accuracy equivalent to that of models on far more powerful hardware. The exoskeleton itself will be capable of reliably following and pre-empting user motion that the model had been trained to recognise. The impact of different Neural Network types, such as Recurrent and Convolutional Neural Networks, and factors such as input size, model size, training data quality and quantity, and depth of prediction were considered and reviewed in the creation of the prediction system.

The Resulting prediction system that was developed was capable of running on inexpensive, Low-Cost, Low-Impact Microcontrollers. This system was capable of predicting the wearer's knee angle movements 150ms into the future, with a Mean Average Error (MAE) from Reality of 2.5%. It had a maximum potential prediction depth of one second into the future.