Nonlinear fluorescence with distinguishable molecular spectra is emitted when fs laser pulses are launched in air due to the nonlinear effects between fs laser pulse and gases. Since every molecule has its particular feature in the fluorescence spectra, these fluorescence spectra can be used to analyze the components of gases in the air. However, since the spectra created by different molecule overlap, it is hard to analyze the nonlinear spectra by the conventional spectroscopic analysis methods. A cascaded neural network model is proposed to analyze the nonlinear fluorescence spectra. To improve learning speed of the neural network and the recognition rate, some preprocessing has been done. 100% correct recognition rates are achieved for both training spectrum samples and test spectrum samples. The simulations show that the proposed algorithm is a new effective method for real-time recognizing the gas components without analytical sampling.