This work has been conducted in the context of pattern-recognition-based control strategies for electromyographic prostheses. It focuses on the conceptual design, implementation and validation of learning techniques based on the k-nearest neighbour (kNN) scheme for gesture recognition. After theoretical considerations and the identification of the topic within the contexts of prosthetic control, biomedical signals — specifically electromyography (EMG) –, and machine learning, particular state-of-research concepts are pointed out. With regard to nearest-neighbour-classification this also concerns methods for dataset reduction in order to cope with the problem of high computational demands in the prediction phase of instance-based learning. Requirements for a kNN-based learning scheme suitable in the field of surface-EMG-controlled prosthetics are specified, comprising high accuracy and success rates, as well as incrementality, and proportionality for not explicitly learned gestures (intermediate levels of intensity). Furthermore, the requirement for an applicability of the proposed methods on embedded systems is stated for an extended approach, considering real-time behaviour and determinism. On the one hand, these requirements are evaluated by theoretical examination. On the other hand, the methods proposed are practically implemented. Datasets are captured by means of a state-of-the-art eight-channel-EMG armband positioned on the forearm to test the implementation. Based on this data, the influence of kNN’s main parameter k on block-wise cross-validation accuracy is analyzed while furthermore varying weighting factors and distance metrics. In addition, the effect of windowing concepts is investigated. Moreover, the effect of varying proportionality schemes is investigated, regarding both cross-validation accuracy and furthermore success rate in pilot experiments. These are conducted as online target achievement tests, moreover incorporating the evaluation of thresholding schemes for kNN classification. Additionally, an assessment of different dataset reduction techniques’ adequacy for embedded control applications is made by applying kNN on the reduced prototype set and analyzing the cross-validation accuracy as well as the timing behaviour when using captured EMG data. Among these methods, the Decision Surface Mapping algorithm (DSM) proves itself as most suitable. Furthermore, a randomized, double-blind user study is conducted in order to compare the implemented methods, namely kNN with a specific set of parameters and kNN after applying prototype generation via DSM, with the state-of-research algorithms Ridge Regression as well as Ridge Regression with Random Fourier Features. The results from these experiments show a statistically significant improvement in favour of the kNN-based algorithms in comparison to the ridge-regression-based techniques. Notably, the approach of kNN applied on the DSM-reduced set achieves higher success rates than the original technique in some cases. Although the difference between kNN and DSM-kNN has no statistical significance, it is remarkable in consideration of only using seven prototype samples in the reduced set in total, thus yielding a reduction rate of over 99% while preserving accuracy. With k set to 1 – which turned out to be an excellent choice – the running time complexity of both kNN (in every prediction step) as well as DSM-kNN (in the training phase) can be considered as linear with respect to the number of original samples, speaking in favor of an embedded applicability.
Presented on 17.06.2020 by Tim Sziburis