Kurtogram manifold learning and its application to rolling bearing weak signal detection

The kurtogram method has been proposed for weak signal enhancement and been validated very powerful. The essence of the kurtogram method is de-noising by optimal band-pass filtering, however, the in-band noise are left unprocessed. As a result, the fault induced transient impulses, which spread within a wide frequency band, would be still contaminated by noise. Aiming at the flaws encountered by the kurtogram method, this paper proposes a novel kurtogram manifold learning method for signal de-noising in whole time-frequency space. In this method, the sub-signals split by kurtogram are fused to build a high dimensional transient impulse feature space, manifold learning is conducted to mine the transient impulse feature space. On this basis, the in-band noise can be filtered out, thereby the transient impulse features will be uncovered. The experimental validation results exhibit the proposed method outperforms kurtogram method and is effective for rolling bearing weak signal detection.