That the support vector machine network is applied to recognize the nonlinear fluorescence spectrum of impurities of different concentrations in air is proposed. Because the number of spectrum channel of the original spectrum data is large, it is cleaned up and compressed through wavelet trausform firstly, and then the principal component analysis (PCA) is used to extract the character information twice in series. It not only ensures the character of original nonlinear fluorescence spectrum, but also compresses the data number the nonlinear fluorescence spectrum from 3979 to 514, and extracts 9 principal components, which reduces the number of the input vector and improves the training speed of the network. The simulation results show that the correct recognition rates for both training spectrum samples and unlearned test spectrum samples reach 100%. So, the training and testing speed is fast enough to monitor the atmospherical impurity in air in real time.