Wavelet Scattering Cyclostationarity Representation for Machine Intelligent Fault Diagnosis

Machine intelligent fault diagnosis(IFD) model consists of a feature extractor and a pattern classifier. As machine fault information is conveyed by the cyclostationarity of the monitored signals, the feature extractor should capture the monitored signal's cyclostationarity to ensure feeding the classifier with diagnostic-relevant features. Conventional cyclostationarity representations, e.g., cyclic spectral correlation(CSC), cyclic modulation spectrum(CMS), and multi-scale envelope spectrogram(MuSEnS), are unstable to signal deformation caused by working condition variability. As a result, the IFD model taking conventional cyclostationarity representation as the feature extractor will deteriorate when working condition variation occurs. This current work formulates wavelet scattering as a novel cyclostationarity representation and verifies its stability to working condition variability theoretically and experimentally to remedy this gap.