Incremental Learning Classifier for Myoelectric Pattern Recognition Against the Interferences of Outlier Motion and Time-variability

Gesture classification models utilizing surface electromyography (sEMG) play a crucial role in controlling artificial hands. However, traditional classification models possess limited recognition capability, failing to meet users’ demands for new gestures. Furthermore, traditional classification models suffer significant degradation in recognition performance due to untrained motions and the time-variability of sEMG. To overcome these challenges, we proposed an incremental learning-based Gaussian-LDA hybrid classifier (GLHC) that integrates an ensemble of Gaussian one-class classifiers with multi-class linear discriminant analysis (LDA). In the GLHC model, Gaussian ensemble can reject outlier motions, while LDA can recognize target motions. Furthermore, two self-update mechanisms have been implemented to enhance the model’s recognition capability and counteract sEMG time-variability effects, respectively. Experimental results indicated that GLHC effectively mitigates both types of interference, with an online expansion of its recognition capabilities. Compared with traditional classifiers and class-incremental models from previous studies, GLHC demonstrated enhanced robustness and superior recognition performance.