Transient Feature Extraction Based on Time–Frequency Manifold Image Synthesis for Machinery Fault Diagnosis

Abstract

Fault diagnosis of rotating machinery is crucial to the safety management of the equipment. However, the weaker intrinsic features are generally submerged in the strong noise interference and modulated to multiple frequency scales, which will weaken the extraction and identification of the transient features. To enhance the transient features, a method called time–frequency manifold image synthesis (TFMIS) is proposed in this paper. By inheriting and promoting the merits of time–frequency manifold (TFM) in feature enhancement and in-band noise suppression, the proposed method contributes on the natural compression and enhancement of the time–frequency transient features in the view of the image processing. First, the raw time–frequency image (TFI) is compressed by the 2-D discrete wavelet transform with the principal structure remained. These approximation sub-TFIs are later used to achieve a fast TFM learning process. Then, a relationship between the global TFI and the local TFM can be adaptively built by using the histogram concept with the probability distribution property demarcated. Thereupon, the proposed method can enhance a global manifold transient structure with a matching rule built from the local TFM. Consequently, by a series of inverse transformations, a TFMIS scheme is constructed for the transient feature extraction in a self-learning process. Two case studies, including bearing and gear transient feature extraction, confirm the performance of the proposed method in achieving rather high-contrast results for the natural transients, and more precise results for the fault frequency identification in detection of periodic transient signals.