A machine learning approach to gait cycle classification for Rheo 3 low-end torque failure

Abstract

This study aims to assess the feasibility of employing machine learning algorithms to classify low-end torque failure in Rheo 3 prosthetic knees to find a reliable way of diagnosing this fault in real-world scenarios, considering that in most cases, this malfunction typically goes unnoticed by the wearer. Using a data set comprising of gait cycles, this research explored and compared two distinct classification algorithms: numerical feature extraction combined with various traditional machine learning algorithms and Continuous Wavelet Transform combined with a Convolutional Neural Network. The findings demonstrated that both approaches exhibit promising results, and accurately classify unseen patients with reasonable accuracy, thereby preventing potentially hazardous faults from remaining unnoticed. This ability to handle novel situations suggests their applicability in real-world situations where the algorithms are expected to encounter diverse and previously unseen data. A key differential outcome of this research was the superior performance of the CWT-CNN model concerning knees that were close to the decision boundary. This finer threshold for classification suggests a better sensitivity and precision in the CWT-CNN model. Future research could explore further wavelet analysis techniques to increase further the classification accuracies, the feasibility of using the CWT-CNN approach for preemptive fault detection, and further refining the traditional machine learning model approach with handcrafted feature extraction methods.