CO was chosen as an early fire detection factor through analyzing all kinds of characters in the process of fires, and an experiment system was established based on Fourier transform infrared spectrometer. Through this system, lots of early fire experiments were carried out, and the authors got the CO concentrations of all kinds of materials. Using the concentration of CO, an autoregressive integrated model was established by time series analysis, then the process characters phi1 and phi2 were extracted from them. Through analyzing the phase graph of the process characters, it was found that the real fires and the nuisance fires were distributed in different regions. Plenty of experiments indicate that this detection method can discriminate between real fire sources and nuisance sources quickly when fires occur.