Abstract: Aiming at the difficulty of fault classification caused by high dimension and information redundancy of rotating machinery fault data set, the Group Sparsity Discriminant Preserving Graph Embedding (GSRDPGE) algorithm is proposed to reduce data dimension effectively. First, the algorithm improved the inter-class sparse coding and obtained a more discriminant inter-class sparse weight matrix. Then, the influence of outliers in feature sets on sparse coding was removed by the weighting method. Finally, the optimal discriminant projection matrix was calculated with the objective of minimizing the intra-class reconstruction divergence and maximizing the inter-class reconstruction divergence. The proposed method was validated with an Iris simulation data set and double span rotor system fault data set, and compared with several other typical dimension reduction methods. The results show that this method can simultaneously take into account the global and local aspects of the data distribution status, make the differences between fault categories more prominent, and improve the accuracy of fault pattern recognition. The results show that this method can provide a reference for intelligent rotor fault diagnosis.

Key words: rotating machinery, sparse coding, sparse residual, attribute reduction