Abstract: In order to solve the problem of low signal-to-noise ratio of the early faults of hydroelectric sets, Singular Value Decomposition (SVD) and Deep Belief Network (DBN) are combined for fault diagnosis in this study. First, a Hankel matrix was constructed by using a vibration signal containing noise to decompose its singular values; and effective singular values, selected through the singular value difference spectrum method, were used to reconstruct a phase space and achieve noise reduction. Then, singular value decomposition was applied to the signals denoised, and the resulted singular value sequence was used to construct a feature vector. Finally, a DBN classifier model was developed to realize the fault diagnosis of hydroelectric sets and compared with BP neural network and multi-class support vector machine. The results show this method can identify the fault type of hydroelectric sets more reliably and efficiently.

Key words: hydroelectric sets, fault diagnosis, singular value decomposition, deep belief network