Environment classification using sensor time series in a microprocessor-controlled prosthetic knee

Abstract

In microprocessor-controlled prosthetic knees (MPKs), the amount of sensor data that is available as a basis for controlling the prosthetic functionality continues to increase. Interlinking the information manually to decision rules in MPK control software is consequently less and less feasible. Therefore, for this thesis, the exemplary task of classifying transitions in the walking environment was addressed using a contemporary tool set from statistics, signal processing and machine learning. Sensor data of 14 quantities from five experienced MPK users navigating through the environmental transitions over 1000 times was recorded. The recordings were segmented into the gait phases and augmented via transformations motivated by domain knowledge. Three different classification strategies were used, being distance-based, feature-based and related to deep learning respectively: A nearest centroid classifier (NCC) using the Soft-DTW distance measure, a support vector machine (SVM) operating on automatically extracted and selected time series features and a convolutional neural network (CNN). Five models of varying complexity using these strategies were tested via cross validation and a leave-one-group-out scheme that evaluated the classification performance only on data from one test person, whose data was not used for training. Accuracy during cross validation was generally high, with only two misclassifications in over 1000 samples for the best model (CNN) and 96.8% accuracy for the worst (NCC). In the leave-one-group-out test, accuracy dropped by 2 to 15 percentage points. Data stemming from the loading response phase produced the best results. Two data selection techniques revealed heuristically, that the data augmentation yielded particularly valuable quantities, most notably the estimation of knee and foot trajectories from accelerometer data. Estimating the computational revealed that a 99% accurate algorithm could be executed on the embedded prosthetic hardware without modification. Applicability of the presented strategies in commercial MPKs on the one hand depends on the quality of the practically acquirable training data or the employment of alternative adaptation strategies like reinforcement learning. On the other hand, the valuation of the achieved accuracies is determined by the functional implications of the classification and their impact on safety. It is unlikely that a single machine learning system will cover all control tasks in an MPK in the near future, whereas integrating pattern recognition into existing frameworks is a promising direction of further development.