Abstract
An anomaly detection method based on multi-dimensional time-series sensing data has been developed on the purpose of enabling condition based maintenance. The proposed method generates normal state models using the learning data selected by the plant operation information and detects anomaly based on the distance between the model and the data. Local sub-space classifier is applied for normal state model and adequate threshold is calculated using learning data. The proposed method was evaluated using 4 datasets of time-series sensing data obtained from real equipments. It was confirmed that anomaly signs several days before equipment faults was detected properly while false detection hardly occurred.
