We proposed and demonstrated a new machine learning algorithm for wavelength modulation spectroscopy to enhance the accuracy of fire detection. The result shows more than 8% of accuracy improvement by analyzing CO/CO2 2f signals.