Time–frequency manifold for nonlinear feature extraction in machinery fault diagnosis

Abstract

Time–frequency feature is beneficial to representation of non-stationary signals for effective machinery fault diagnosis. The time–frequency distribution (TFD) is a major tool to reveal the synthetic time–frequency pattern. However, the TFD will also face noise corruption and dimensionality reduction issues in engineering applications. This paper proposes a novel nonlinear time–frequency feature based on a time–frequency manifold (TFM) technique. The new TFM feature is generated by mainly addressing manifold learning on the TFDs in a reconstructed phase space. It combines the non-stationary information and the nonlinear information of analyzed signals, and hence exhibits valuable properties. Specifically, the new feature is a quantitative low-dimensional representation, and reveals the intrinsic time–frequency pattern related to machinery health, which can effectively overcome the effects of noise and condition variance issues in sampling signals. The effectiveness and the merits of the proposed TFM feature are confirmed by case study on gear wear diagnosis, bearing defect identification and defect severity evaluation. Results show the value and potential of the new feature in machinery fault pattern representation and classification.