2018 年 54 巻 1 号 p. 9-15
This paper proposes a novel sequential pattern recognition method enabling calculation of a posteriori probability for learned and unlearned classes. To discriminate undefined classes via model parameter estimations using given learning samples, probability density functions of unlearned classes are incorporated in a hidden Markov model. This method can be applied to various pattern recognition problems such as the motion classification with electromyogram (EMG) signals and the automatic discrimination for diagnosis support. In the experiments, artificial time series data generated from hidden Markov models and EMG patterns measured from forearm muscles were classified to validate classification performance of the proposed method for learned and unlearned classes. The motion classification using EMG signals were conducted with five subjects and eight forearm motions including four learned motions and four unlearned motions, were performed. Compared with previous methods, the proposed method provided a higher classification performance on learned (artificial data: 100%; EMG patterns: 91.5%) and unlearned (artificial data: 99.5%; EMG patterns: 89.6%) classes. From the results, the effectiveness of the proposed method was demonstrated.