To improve the efficiency of condition based maintenance and the capability of abnormal condition detection and warning ability, a new method which based on dynamical threshold for abnormal operation condition detection was proposed. Baselines of the parameters are depicted through the establishment of dynamical prediction model. On such basis, the Bayesian network is utilized for calculating the operating state of transformer, which uses the probability to decide its potential condition. Finally, the proposed method is verified using the field acquired data. Numerical simulation results show that the proposed method in this paper can detect the abnormal state of the power transformer effectively and which can also provide a warning of those possible abnormal conditions.