Enhancement of decomposed spectral coherence using sparse nonnegative matrix factorization

The signals from faulty rotating machineries are characterized by the series of transient or by the cyclostationarity, denoting a process whose statistical properties vary periodically with time. One of the primary purposes for condition monitoring system is to detect the cyclostationarity in the measured signal as early as possible, which is perhaps mostly achieved by an analysis of the envelope spectrum [1]. Cyclic spectral analysis yielding statistical descriptors like the spectral correlation (SC) or the spectral coherence (SCoh, a Integration of the spectral coherence over the domain of spectral frequency is one popular way for reaching the envelope spectrum which is an indispensable tool for the fault diagnosis of rotating machineries. Envelope spectrum can be enhanced by introducing a decomposition of spectral coherence with the aid of nonnegative matrix factorization frequently exploited for the data clustering. Based on this regime, the present study aims to deal with further improvement of the envelope spectrum by taking two major considerations. First, it is to impose a sparsity constraint to the minimization problem treated in the standard NMF, eventually allowing a sparse representation of the envelope spectrum. By taking advantage of a randomness of NMF solution, the second is to establish how to correctly choose the number of clusters, a prerequisite for starting the NMF algorithm. Finally, the suggested method is verified throughout a synthetic data and experimental measurement from propeller cavitation. (c) 2021 Elsevier Ltd. All rights reserved.