Abstract
Recently there has been a growth in the world-wide demand for an intelligent fire alarm system. Such a system is required to have a high reliability, ensuring detection of a fire as early as possible and eliminating false alarms. To achieve this, the system may need to use several kinds of sensors to detect fire components such as smoke, temperature and gas and also to make a fire/non-fire judgment by forecasting changes in the phenomena.
Fires, other than explosions, may be divided into two types depending on how they progress: the flaming fire and the smouldering fire. Since a flaming fire progresses quickly to a dangerous level, the fire judgment needs to be done as quickly as possible. On the other hand, a smouldering fire progresses rather slowly, and the fire judgment can be done less quickly than for a flaming fire. However, to detect a smouldering fire in the early stage, higher accuracy of fire judgment is required. After ignition occurs, a smouldering fire progresses in the same manner as a flaming fire. Non-fire phenomena, such as cigarette smoke and temperature increase due to an airconditioner, usually progress more slowly than a smouldering fire but sometimes include pulse-like changes. Since the behavior of such non-fire phenomena may often cause a false alarm, the system should be able to cope with them.
For judgment of a fire or non-fire in the early stage, we have developed an algorithm that can recognize the difference among timedomain patterns of the sensor outputs against various kinds of fire and non-fire phenomena. The algorithm was realized in the following way. First, the outputs of the sensors are filtered to remove those slowly changing or pulse-like signals that may cause a false alarm and the electrical noise generated in the sensors or the transmission channels. From the filtered signals, an approximate function is made to forecast the change of the phenomena. If the approximate function has a curve that will reach a prescribed level of danger, the period of time which needs to elapse from the present point of time until the curve reaches that level is calculated. If the danger forecast time is shorter than a certain time, the present phenomena are judged as a fire. If it is longer, the present phenomena are judged as a non-fire. This judgment is performed continuously against the sensor output data.
We made an experimental fire alarm system using the above algorithm and obtained the following results. For smouldering fires that eventually progressed to a flaming fire, the system was able to issue a fire alarm a maximum of six minutes before ignition occurred. For flaming fires, the system was able to issue a fire alarm within a maximum of four minutes after ignition occurred.
