Development of an anomaly detection method with a novel hidden semi-Markov model incorporating unlearned states

This paper outlines a novel pattern recognition approach incorporating consideration of unexpected anomaly patterns in time-series data. In this approach, probability density functions of unlearned states are incorporated in a hidden semi-Markov model allowing consideration for the temporal dependence of data, as the general classifier used misidentifies abnormal data not belonging to classes predefined in training. In the experiments performed, the proposed method was applied to the classification of artificial time-series data and motion recognition problems for simulated care tasks. The characteristics seen in motion recognition were extracted from physical data obtained from a care worker via a motion capture system. The proposed method produced higher levels of motion recognition than previous approaches, and the results demonstrated the effectiveness of the technique.