Mechanical fault diagnosis based on non-negative matrix factorization with spectral clustering initialization enhancer

Aiming at the convergence problem of non-negative matrix factorization (NMF), this paper introduces the spectral clustering algorithm to produce the structured initial value for non-negative matrix factorization, and a new mechanical fault diagnosis model is studied, which is based on the NMF with spectral clustering initialization. Firstly, the cluster centers of the vibration signal in characteristic space are calculated with the spectral clustering algorithm of Ng-Jordan-Weiss (NJW); then the NMF algorithm is employed to obtain the basis vectors initialized with the cluster centers; finally, the projection coefficients of the vibration signal are calculated in characteristic space, used as the characteristic vectors and sent to the K-nearest neighbor classifier (KNNC) for fault identification. The proposed fault diagnosis model realizes the automation process from fault feature extraction to fault identification; the spectral clustering initialization method not only enhances the convergence effect and decomposition property of NMF, but also improves the accuracy of fault diagnosis. The diagnosis example of a rolling bearing verifies the effectiveness of the proposed model.