Local-Dictionary Sparsity Discriminant Preserving Projections for Rotating Machinery Fault Diagnosis Based on Pre-Selected Multi-Domain Features

In this paper, a novel fault diagnosis framework for rotating machinery is put forward, centering on a firstly proposed dimensionality reduction algorithm named local-dictionary sparsity discriminant preserving projections (LDSDPP). To involve abundant fault-related information for model construction, multi-domain features are extracted directly from vibration signals as along as their sub-band signals decomposed via time-frequency domain analysis. To reduce computational complexity and improve modelling accuracy, features are pre-processed and those highly related to faults are selected based on filtering models. For feature reduction, sparsity and local structures of data points are introduced in LDSDPP by constructing graphs using sparse representation based on a local dictionary containing only neighbors for each sample. Moreover, the global information of data samples is also integrated into LDSDPP for better discriminant power. Experiments based on two datasets have demonstrated the effectiveness of the proposed framework and the superiority of LDSDPP to other comparable algorithms in classification performance.