Surface electromyography driven hand motion recognition for long-term rehabilitation use

Student thesis: Doctoral Thesis


The control of prosthetic hands and other upper-limb assistive device for rehabilitation relieson the premise that users’ hand motion intention is accurately recognised. Among all the feasible modalities, surface electromyography (sEMG) based hand motion recognition has been most widely adopted for its intuitiveness and effectiveness. However, the reported promising recognition accuracy is mostly confined to intra-day scenarios, which ignores the performance degradation of inter-day application for long-term use. To address the challenging inter-day hand motion recognition for long-term use, current sEMG drivensolutions are further developed with an improved performance verified by experiments in this thesis. The contributions are recognised in terms of improved pattern recognition based classification, additional sEMG feature extraction and selection, novel multimodal fusionc based hand motion recognition, and new long-term sEMG benchmark building.
First, both conventional pattern recognition and deep learning approaches are developed to accommodate the long-term use with inadequate and adequate training data respectively. Based on the feasibility of a force driven subclass division in our preliminary work, subclass division based linear discriminant analysis (LDA) frameworks using solely sEMG signals are proposed. Both explicit and implicit subclass division strategies are explored including the K-nearest neighbour based LDA (KNN-LDA) and subclass discriminant analysis (SDA)with a verified improvement of long-term hand motion recognition accuracy for inadequate training data. A convolutional neural network (CNN) architecture is adopted using raw sEMG as the input without preprocessing, whose significant improvement of long-term recognition accuracy has been seen with adequate and pooled training data across multiple days and subjects. Then the feasibility of merging handcrafted features and non-handcrafted features is proved in combination with a diversity of classification algorithms for the longterm hand motion recognition. And a novel multi-threshold based handcrafted feature vector is proposed and achieves an improved recognition accuracy. The feature selection is conducted with the bacterial memetic algorithm for achieving different targets including a compromised yet comparable recognition result at a largely reduced computational cost with selected subsets of existing features, and an improved recognition accuracy with selected features from enriched sub-segments of multiple lengths. To further remedy the lack of deep muscle activity sensing in myoelectric sensing, the ultrasonic sensing is investigated as a complementary modality and integrated with the myoelectric sensing, which contributes to an improved accuracy of hand motion recognition. Finally, the lack of long-term constraints and low-density representations in existing public databases is addressed by building a new dataset comprising the long-term sEMG signals of 13 hand motions captured from 10 subjects in consecutive 10 days under a standardised protocol as a public benchmark for the research community.
Date of AwardJun 2019
Original languageEnglish
Awarding Institution
  • University of Portsmouth
SupervisorHonghai Liu (Supervisor)

Cite this