The purpose of the automatic fire alarm facilities is to detect the fire earlier which occurs under the various facility environment and to provide a judgment result with high reliability. The facilities protect a human life and a fortune from the fire and they become indispensable equipments. On the other hand, however, the lack of the reliability of the facilities, accompanying the occurrence of the false alarm, is pointed out. We propose a new fire judgment system using the time series of smoke density in early stage in order to improve the reliability. We think that the different causes of smoke arise different statistical properties on the time series. We employ the statistical model which uses a 36-dimensional histogram as the feature vector and a statistical distance as the discriminant function. We evaluated experimentally an efficiency of our method for actual 171 unknown time series. The average judgment rate was 84%, and this rate was higher than the rate of the former method (60%).