Continuous classifier training for myoelectrically controlled prostheses

Myoelectrically controlled prostheses use pattern recognition systems to classify input motions. Typically, these systems are initially trained offline using a set of training data. Changing conditions cause an increase in signal variation, leading to higher error rates. For better adaptability, a continuously trained classifier was developed. Data with valid class decisions are used to retrain the classifier with the class decisions used as classification targets. In this implementation the classifier validates decisions by using a retraining buffer to locate consecutive, identical majority vote decisions. Retraining is performed by incorporating new valid feature vectors, selected from the retraining buffer, into the training set, while discarding older vectors. Using the continuously trained classifier, an average improvement of 2.57% was seen over the noncontinuously trained classifier.