Abstract: The vibration signal measured in industrial applications is always contaminated by different noises, which leads to mistakes in fault diagnosis. Most of the de-noising algorithms deal with the time signal directly, which also are affected by computation time and memory space. The samples in feature space extracted from the original signal are more important than the time samples in fault diagnosis, which playing an important role in the application of manifold learning. A singular-value-based de-noising method is presented to denoise the feature samples, and then local preserving projection algorithm is used to reduce the feature dimension. Simulation and experiment results indicate that, comparing with de-noising the original time-domain signal, the proposed method can effectively speed up the computation process and decrease the memory space, while keeping the ability of dimension reduction and classification.

Key words: feature space de-noising;locality preserving projections;manifold learning;fault classification;bearing