The oscillation of frequency in power grid is studied in this paper. The possibility association of frequencies measured at different locations are modeled by a Bayesian network with the logical structure learned using Bayesian structure learning and real measurements in the U.S. power grid. Frequency data analysis and the detection of incorrect frequency measurements (caused by equipment error or malicious attack) are performed over the logical Bayesian network structure. Such application of Bayesian network is a powerful mathematical tool in computational intelligence. Without the physical power network topology information, a two-branch search-and-score structure learning algorithm with L -1 regulation is proposed to learn the logical structure, achieving around 97% correct prediction rate for future frequency and 92% detection rate for false frequency data with 2% false alarm rate. The tool of epidemic propagation over this logical network is also exploited to analyze the propagation of frequency changes. Using the Kolmogorov-Smirnov test, such logical structure is demonstrated to be well approximated by the Small World network model. And the propagation of frequency changes is demonstrated to be described by the Susceptible-Infectious-Susceptible (SIS) model quite well. The Bayesian structure obtained from the real measurement is statistically validated using a 5-fold training data and the Pearson system.