Design and evaluation of electromyography signal processing techniques using resource-constrained devices

Resumen

Surface electromyographic (sEMG) is an acquisition technique based on recording muscles potential over the skin. Processing those signals on a resource-constrained device is a key element in achieving wearable health devices. sEMG based devices have a wide range of applications: early diagnose and treatment of neurodegenerative diseases, tracking of daily activities, rehabilitation, and adaptive training. For most of these applications, it is required to identify which gestures the user is doing. This classification is difficult due to low signal to noise ratio, data complexity and size, and changes between sessions and subjects. These are a major challenge in a resources-constrained device.

The sEMG based devices consist of two parts: the signal acquisition and the processing of the data generated. This master thesis focuses on the data processing with two differentiated parts: An analysis of a public gesture dataset and the design of a classifier.

The quantitative analysis to the public available gesture dataset was made using Parallel Factor Analysis (PARAFAC) decomposition. It was found that reduction from 14 to 4 channels was possible without losing significant information. Better understanding of the sEMG signal was achieved, estimating that the most significant information is located under 300 Hz.

In this work, raw sEMG signals have been acquired from 4 passive electrodes and 8 and gestures have been classified using Support Vector Machine (SVM). The classifier achieved over 85% recognition accuracy similar to the state-of-the-art typical results. A factorial design was carried out to identify the relations between the coding of the SVM, the datatype used, the length of the sample and their influence in the memory footprint of the classifier and execution time. Memory was found to be the bottleneck for an embedded device implementation of the SVM algorithm. The SVM execution time was 130 ms and its memory footprint was 868 KB.