Structurally guided nonnegative matrix factorization for compound fault feature decoupling of rotating machinery

Spectral coherence (SCoh), as an effective tool for revealing potential cyclostationary components in signals, has been widely applied in rotating machinery fault diagnosis. However, most existing SCoh-based methods rely on searching for optimal demodulation bands to construct improved envelope spectra, which struggle to suppress in-band noise and fail to accurately separate multiple fault features residing in the same frequency band. To solve such issues, a structurally guided nonnegative matrix factorization (SGNMF) method is proposed to improve the decoupling capability and feature extraction accuracy for compound fault. First, an optimized candidate fault frequency identification strategy is developed by integrating harmonic consistency and local energy enhancement to accurately extract key cyclic frequencies. Then, the fault information abundance at target fault frequencies across SCoh frequency slices is evaluated to construct a physically interpretable basis matrix, which reflects the spectral responses of potential fault sources. Finally, the coefficient matrix is initialized using singular value decomposition and iteratively optimized under a fixed basis matrix by minimizing the reconstruction error, which enables the accurate decoupling of compound fault features in the cyclic frequency dimension and the effective suppression of in-band interference. The analysis results of the simulated and experimental signals demonstrate that SGNMF can effectively extract and separate fault features under complex conditions such as overlapping frequency bands and noise interference, showing superior compound fault decoupling performance.