In order to realize the industrial on-site rapid detection and identification for special steel, a mobile laser-induced breakdown spectroscopy prototype based on optical fiber delivering laser energy is adopted in this experiment to collect the spectral data of 14 special sheets of steel. The spectra of special steels were rapidly classified via dimensionality reduction in which pre-selected spectral lines were traversed, combined with a support vector machine (SVM).In the experiment, original spectral data, normalized spectral data and normalized spectral data after traversed were used as the input vectors of the SVM classification model, and the recognition accuracy of the model for special steels under different input vectors was compared. The results show that on the basis that more than 51 spectral lines were selected as input variables, the recognition accuracy of normalized spectral data as input variables for steels reaches 95.71%. It is significantly higher than 11.43%, whose accuracy was used raw spectral data as the input vector. Further, the MATLAB program was used to traverse the spectral line combination to choose the optimal input features. When 6 specific spectral lines were selected, the accuracy of special steels recognition reached 100%, and the modeling speed was also improved accordingly. It can be seen that when a large number of common feature data are pre-selected, automatic feature selection by machine has obvious advantages over the spectral line of manual selection. The SVM algorithm based on this dimension reduction method has a good industrial application prospect in LIBS rapid classification technology.