Blind Source Separation Based Classification Scheme for Myoelectric Prosthesis Hand

For over three decades, researchers have been working on using surface electromyography (sEMG) as a means for amputees to use remaining muscles to control prosthetic limbs (Baker, Scheme, Englehart, Hutcinson, & Greger, 2010; Hamdi, Dweiri, Al-Abdallat, & Haneya, 2010; Kiguchi, Tanaka, & Fukuda, 2004). Most research in this domain has focused on using the muscles of the upper arms and shoulders to control the gross orientation and grasp of a low-degree-of-freedom prosthetic device for manipulating objects (Jacobsen & Jerard, 1974). Each measured upper arm muscle is typically mapped directly to one degree of freedom of the prosthetic. For example, tricep contraction could be used for rotation while bicep flexion might close or open the prosthetic. More recently, researchers have begun to look at the potential of using the forearm muscles in hand amputees to control a multi-fingered prosthetic hand. While we know of no fully functional hand prosthetic, this is clearly a promising new area of EMG research. One of the challenges for creating hand prosthetics is that there is not a trivial mapping of individual muscles to finger movements. Instead, many of the same muscles are used for several different fingers (Schieber, 1995). To identify hand gestures and actions that are a result of multiple active muscles, relative muscle activity from the different muscles in the forearm has to be identified. For this purpose, the sEMG needs to be recorded using multiple electrodes. However due to the close proximity of the different active muscles, each of these electrodes record muscle activity from multiple muscles, referred to as cross talk. In case of the hand there are number of muscles in the close proximity and often crossing over each other and cross talk is a major cause of low reliability in identifying the action. This is further exaggerated when the muscle activity is weak such as during maintained isometric gestures due to the relative low signal to the background noise. Spectral and temporal overlap makes the use of conventional filtering quite useless. Differences between people make accurate modeling and generalization of the sEMG not possible. To overcome this, number of researchers have used generalization pattern recognition tools and some of these works have got good results (Cheron, Draye, Bourgeios, & Libert, 1996; Koike & Kawato, 1996; Meyer, 2000). However these require manual intervention and it is obvious that such techniques are not suitable for reliable operations that can be automated. Numbers of researchers have attempted to use sEMG for controlling the prosthetic devices (Doerschuk, Gustafon, & Willsky, 1983;Zardoshti-Kermani, Wheeler, Badie, & Hashemi, 1995). While the current techniques to classify sEMG are suitable for identifying gross