Abstract: Feature selection is an important part during the process of qualitative and quantitative analysis of infrared spectrum. In order to solve the disadvantage of traditional methods, such as multi-parameters, slow convergence, poor accuracy, prone to premature, etc., a novel feature selection algorithm was proposed, which combined the basic bat algorithm and Lvy flights search strategy. Meanwhile, due to the original version of bat algorithm was only suitbale for continuous problems, a binary version of bat algorithm was proposed. Three infrared spectrum datasets were used to check the performance of proposed method while the comparisons with traditional genetic algorithm, simulate anneal algorithm and uninformative variable elimination methods were also implemented. The experiment results show that, the proposed method can quickly find the global best combination of sub-intervals and improve the accuracy and stability of feature selection. More importantly, the selected wavenumbers have exactly physical meanings. Meanwhile, the generalized performance of the model established based on the selected wavenumbers was better than the whole spectral range. The tests on three different phases(solid, liquid and gas) and different spectral range indicated that, the proposed algorithm has a widely practical scope and value.