Abstract: A huge mass of on-site monitor data of hydropower stations has accumulated, but its practical use is quite limited due to lack of on-site experts. How to mine these data and combine with expert experiences in fault diagnosis is our focus. This paper describes a vibration fault diagnosis model for hydropower units based on a Bayesian network that integrates subsystem models constructed through formulating the network structure and certain node parameters using expert experiences, discretizing data signals with a self-organizing map (SOM) neural network, and determining the probability distribution of the rest of the nodes via learning parameters of the expectation-maximization (EM) algorithm. This model is verified by examining the effect and rationality of its diagnosis results in design tests.

Key words: hydropower unit, vibration, fault diagnosis, Bayesian network, SOM neural network, EM algorithm