The infrared spectrum of mixed gas got in the battlefield and complex environment results in overlapping and stagger of the primary and secondary peaks, so its feature extraction of qualitative recognition is particularly important. The infrared spectral data collected from a variety of chemical warfare agents and organic gases are high-dimensional data. Centralizing before reducing dimension was used for feature extraction to capture the essence of more information it contained. Since the infrared spectrum of the mixed gas was non-linear and non-Gaussian signal, this method regarded non-Gaussian as independence measure to separate each component as independent component. In order to meet real-time requirements, its iterative process was optimized based on the traditional fast independent component analysis(FastICA) algorithm and extreme learning machine(ELM) model was applied to quantitative analysis. Experiment results show that the iterations of optimized method reduces compared with the traditional method and mean square error of quantitative analysis is E=2.392 6×10-4 and regression coefficient is R=0.999. And the optimized method improves the isolated efficiency of separating pure substances spectra from mixture substances without affecting the separate accuracy.