Isolation and Identification of Compound Faults in Rotating Machinery via Adaptive Deep Filtering Technique

Compound defects commonly occur on rotating machinery under fatigue and heavy loads, and their fault signatures are coupled and easily buried in strong unwanted vibrations and background noise. The isolation and identification of the compound fault signatures are still a challenge especially when the transient impulses induced from the compound defects share common resonant frequency. In this paper, a data-driven, adaptive deep filtering technique which mainly contains filtering and isolating procedures is proposed for compound faults diagnosis. During the filtering procedure, an empirical wavelet transform (EWT) based correlated kurtogram is presented to adaptively obtain the proper spectral segments for filtering the vibration measurements, such that to enhance the signal-to-noise ratio (SNR) of compound faults in the filtered signals. Subsequently, during the isolation procedure, windowed correlated kurtosis (WCK) which outputs pure periodic pulses indicating the occurrence moments of interested fault impulses is presented to isolate each interested fault mode and to determine the defects number. The performance of the proposed technique is tested on simulated signals and validated via analyzing experimental measurements from high-speed locomotive bearings which suffer multiple damages. The results validate that the proposed method outperforms dyadic wavelet transform and spectral kurtogram in isolating and identifying compound faults in rotating machinery.