Specialty steel is an essential foundational material for equipment manufacturing and large-scale engineering projects. For efficient, rapid and accurate classification and identification of specialty steel, this work proposes a cascade laser-induced breakdown spectroscopy (LIBS) classification model combining the iterative wavelet transform (IWT) and deep Boltzmann machine–backpropagation neural network (DBM-BP). This study investigates seven types of special steel. Initially, the spectral data were subjected to background removal using IWT. Subsequently, principal component analysis (PCA) and DBM were respectively employed to reduce the dimensionality of the preprocessed data. Finally, two hybrid classification models, namely, PCA-BP and DBM-BP were constructed. The experimental results revealed that the signal-to-background ratio (SBR) of the Mn element increased by approximately one order of magnitude after IWT background removal. Furthermore, the DBM-BP model achieved a classification accuracy of 98.53 % on the training set and 96.83 % on the test set for the specialty steel samples, validating the superiority of the DBM-BP model in nonlinear feature extraction. This study demonstrates that the IWT-DBM-BP combined model effectively suppresses the continuous background while preserving critical spectral features, thereby enabling high-precision classification. within the seven types of specialty steel samples investigated. This study offers reliable technical support for practical LIBS applications in rapid steel identification and intelligent scrap metal sorting.