Abstract:

：Facing on the crucial problem that the recognition function of current early fault identification methods for rotating machinery declines easily in condition of sparse training samples, a novel early fault identification method based on dimensionality reduction with graph-implanted probability-based semi-supervised discriminant analysis (GIPSSDA) is proposed in this paper. In the case of sparse training samples, GIPSSDA is proposed to reduce the high-dimensional time-frequency domain early fault feature sets of training and testing samples to the low-dimensional eigenvectors with better category segregation, so that the early fault identification accuracy of the terminal learning machine called Optimized Evidence-Theoretic k-Nearest Neighbor Classifier (OET-KNNC) is improved. With the incorporation of the semi-supervised graph-implanted technique, GIPSSDA can exploit both discriminative information and locality geometry of testing samples to search for the optimal projection subspace for classification, which allows GIPSSDA to bring about good classification effect even if the training sample set is small. Experimental results of early fault identification on deep groove ball bearings show the effectiveness and advantage of the proposed method.

Key words: dimensionality reduction, early fault identification, graph-implanted probability-based semi-supervised discriminant analysis, manifold learning, rotating machinery