Adaptive resonance theory-based hand movement classification for myoelectric control system

Electromyography (EMG) also referred to as the Myoelectric, is a biomedical signal acquired from skeletal muscles. Skeletal muscles are attached to the bone responsible for the movements of the human body. In case of prosthetic hand, an EMG based control system known as Myoelectric Control System (MCS) has been widely attracted research in the field. Despite there has been a great development in prosthetic hand industry during the last decade, it is considerably needed to investigate an effective control algorithm for affordable prosthetic hand. This thesis investigates a pattern recognition approach for MCS that classifies hand movements accurately and computationally efficient to actuate different functions of a prosthetic hand. Five distinct hand movements are classified with an Adaptive Resonance Theory (ART) based neural network implemented, as it uses a combination of features extracted from four EMG signals. In order to prove the contribution of the proposed MCS approach, two different evaluation processes have been done. First evaluation considers the investigation of feature extraction method; where the proposed multi-feature consisting of Mean Absolute Value (MAV), Zero Crossing (ZC), Waveform Length (WL), Slope Sign Change (SSC), Root Mean Square (RMS), and Mean Frequency (MNF) has been compared to 2 well-known high accuracy and simple multi-feature methods. The second evaluation is included comparing ART-based methods versus Linear Discriminant Ananlysis (LDA) and k-Nearest neighbor (KNN) as two accurate and simple implementing classifiers. The study outcome reveals that the proposed multi-feature has better extraction performance in terms of class separability and accuracy; while the performance for the proposed multi-feature (82.51%) is at least 6% better than the other 2 methods. Classification results obtained by using the proposed multi-feature have shown better performance of ART-based methods; considering average accuracy of 89.09% for the ART method, 83.98% for the KNN and 82.52% for the LDA. Further investigation has been done on a computation time evaluation between proposed ART-based methods, LDA and KNN. Regarding training time (75.69ms), classification time (49.57 ms) and elapsed time (3.77s), evaluation showed significantly less computation time of ART-based methods than LDA : training time (153.65ms), classification time (344.2 ms) and elapsed time (7.92 s) and KNN:training time (165.42 ms), classification time (230.91 ms) and elapsed time (6.58 s). At last, an accurate and computationally efficient hand movements’ classification approach for Myoelectric Control System (MCS) has achieve