Traditional fire detection technologies usually measure the smoke particles or the temperature increase resulted from fire. However in the early stage of fire, few particles and low heat are generated. Current fire algorithms is based on comparing the fire variables with a given threshold value, the transient sampled values are often affected by some stochastic disturbances. Consequently current methods are hardly alarm fire fleetly and reliably and often give false or failing alarm. A new fire detecting technology was presented based on early fire process signature and fuzzy clustering algorithm. The process eigenvector is made up of CO concentration in detected environment as well as its increasing rate and acceleration. The eigenvectors are divided into two categories of real fire and non-fire, the two cluster centers are obtained by using fuzzy clustering analysis. According to threshold membership principle, the real fire sources can be distinguished from non-fire sources successfully. The result of experiments has shown that the presented technology is feasible for early fire detecting with lower rate of false and failing alarm, and give fire alarm much early than any other traditional method.