In view of the nonlinear characteristics of transformer vibration signal and the problem of winding mechanical state identification, this paper introduces the state identification method combining multiple fractals and bayes, builds a vibration signal acquisition platform to collect the vibration signal of distribution transformer under normal operation of different load current and fault operation under loose winding. Then, the multifractal characteristics of the vibration signal are analyzed by using the multifractal theory, and the multifractal spectrum parameters which vary significantly with the mechanical state of the transformer windings are extracted as the state characteristic parameters. The result of research proves that the vibration signal of transformer has strong multifractal characteristics, the parameters of multifractal spectrum αfmax, αmin do not change obviously when the load current fluctuates, but change evident when the winding becomes loose. The multifractal-bayesian algorithm can accurately identify the normal state when the transformer load current changes and the fault state when the windings become loose, and the accuracy is above 98%. The research conclusion can provide a new idea and new algorithm for the transformer windings fault diagnosis based on the vibration signal in the case of variable load.