Abstract: To treat the feature selection for mechanical fault diagnosis, a novel method is proposed to find the low-dimensional non-negation sparse principal component representations from the feature set of the measured signals. In facilitating the interpretation of the extracted principal components, combine non-negative and sparse constraints with the L1-norm variance, the non-negative sparse components can be selected. The cumulative percentage of variance explained is used to select the optimal sparsity of the principal components, and the number of principal components is decided by the demand of sparsity in fault diagnosis. The experimental results from simulation data and the ball bearing vibration analysis show that the proposed method is more effective for machine fault diagnosis than principal component analysis method. Analysis of the optimal sparsity parameters and the sensitive features, suggests that the proposed method not only self-adaptively obtains the degree of sparsity, but also effectively determines the sensitivity of the original features.

Key words: Fault diagnosis, Feature extraction, Principle component analysis, Sparse representation