Abstract: The sparse filtering-based methods can effectively realize the bearing fault diagnosis via considering the inherent sparsity of fault feature. However, two shortcomings exist in these methods. One is that the input and output dimensions are set empirically to cause the uncertainty in feature extraction. Another one is that the loss of fault information might be caused due to that the prior knowledge is required for strict sifting of specific components. Hence, a feature fusion method based on the fast intrinsic component filtering is proposed in this study. First, the complexity measure is introduced to design an adaptive selection criteria of the sparse filtering dimension. Meanwhile, the optimization target of sparse filtering is used as an index to select a cluster of fusion sources with rich fault information. Second, a manifold learning fusion strategy is established, which consists of three parts: improve the manifold learning method to fuse the selected fusion source, construct a detection strategy for abnormal amplitude and weight the fused components to maximize the fault information. As a result, the selection of sparse filtering dimension, the information loss caused by second sifting, and the singularity of envelope spectrum caused by the abnormal amplitude can be solved by the proposed method.Analysis results verify that the proposed method is more effective to extract the bearing weak fault feature than the current manifold learning and sparse filtering-based methods.

Key words: fault diagnosis, sparse filtering, manifold learning, feature fusion, rolling bearing