An Adaptive Classification Methodology for Myoelectrically Controlled Prostheses

Abstract

Myoelectric signals (MES) have proven to be effective inputs to control systems of powered prosthetic devices. A number of output motions can be derived from the MES monitored from multiple control sites by employing pattern recognition techniques; however, MES measurement conditions will change over time, causing increased signal variation from the training data, making initial training data inadequate classification exemplars. To create a dynamically adaptable system, a classifier that undergoes continuous online training was developed. This classifier validates decisions and uses valid feature vectors for retraining, with classification decisions as classifier targets. Validation utilizes a retraining buffer to find 64 consecutive and identical majority vote decisions. The use of a large buffer ensures a higher confidence that the class decisions are correct. Every 8^th feature vector from the buffer is incorporated into the training set, discarding older feature vectors to maintain a constant number of training exemplars. Retraining the classifier with this new training set allows the classifier to adapt to changes in the MES. This study compared the continuously trained linear discriminant analysis classifier with a noncontinuously trained classifier, using data collected from six subjects. An average improvement of 2.57% was seen with the continuously trained classifier.