Identifying hand-motion patterns via kernel discriminant analysis based dimension reduction and quadratic classifier

Mechanomyographic (MMG) signal for prosthetic control has been investigated in recent years and encouraging results in hand-motion patterns identification have been achieved. In this paper, only two accelerometer sensors were used to record the MMG signal in the forearm of fourteen able-bodied people. A kernel generalized discriminant analysis and three linear dimension reduction techniques were applied to reduce the feature dimensionality and improve the class seperability, and then the simple and commonly used quadratic classifier was implemented to identify the four hand-motion patterns. The experimental results have shown that the average identification rate reaches to a high accuracy of 95.12±3.83% by utilizing the three features extracted by kernel generalized discriminant analysis, where two wrist-related patterns are easier to identify while hand close is the most difficult one. It is concluded that two-channel MMG signal is sufficient for identifying the recorded four hand-motion patterns, which made MMG signal another prospective alternative in prosthetic hand control applications.