LAUSR

In movement classification through surface electromyography signal processing, the classification method must identify the user’s intention with satisfactory accuracy to promote an adequate biosignal interface. Traditionally, classical methods such as Support Vector Machines, Artificial Neural Networks, and Logistic Regression have been used to this end. Recently, Non-Iterative Methods based on Artificial Neural Networks have been revisited in the form of Random Vector Functional-Link Networks (RVFL) and its most recent derivation, the so-called Extreme Learning Machines (ELM). In this work, we evaluate the performance and potentialities of RVFL and ELM with Moore-Penrose (RVFL and ELM) and Ridge-Regression (R-ELM and R-RVFL) methods to classify 17 different upper-limb movements through surface electromyography (sEMG) signal processing. 341 different sets of tests involving sEMG channels and features were performed for each one of the 20 subjects (ten amputees and ten non-amputees) from NINAPro database. Overall, the NIM methods presented consistent advantages of accuracy rate and time processing when compared with most traditional classifiers. Once the best setup of inputs was defined, the R-ELM presented the best accuracy rate. While results up to 80% were already reported for NINAPro data using Deep Learning techniques which are blatantly costly on a computational perspective, there is no evaluation performed in embedded platforms using this database. Therefore, we conducted an embedded study case of the ELM method applied to a Raspberry Pi platform using: a) a timestamp segmentation and b) a sliding-window approach to emulate an online application of the technique. The first trial reached an average accuracy rate of 90.9% for the non-amputee and 63.1% for the amputee subjects. The second trial reached 77.2% of average accuracy for the non-amputee and 55.3% for the amputee subjects, pairing the results in literature, even with the limitations of an embedded platform.