An identification method for fire smoke based on a neural network is discussed. The neural network′s input used three couples of extinction coefficient ratio of three laser beams with wavelengths of 670 nm, 1060 nm and 1550 nm respectively. The network′s output used fire smoke and non fire smoke factors. Experiments of multi wavelength lasers attenuation by typical fire smoke and non fire smoke factors were conducted, resulting in twenty six network sample pattern definition tables acquired from selected data. Three hundred and ninety one times of simulation training resulted in an output error less than 0.0001. The verification experiments prove that this method is effective in distinguishing between fire smoke and non fire smoke factors. This identification method has practical application for solving fire smoke recognition and other similar non structural problems.