Unbalance, misalignment, and cracks are prevalent issues in rotor systems. The coupling of these faults during equipment operation exacerbates the challenge of multi-fault classification. In this paper, the expression of torsional strain energy was derived to capture the dynamic characteristics of rotor systems in different states (including coupling faults) from the perspective of energy, and a multi-fault classification method of rotor systems based on characteristic pattern was proposed. Firstly, the energy trajectories of the rotor system in three-dimensional space were derived by integrating torsional energy with axis orbits, which were then mapped into energy and radial displacement space to obtain the characteristic patterns of the rotor system under different faulty states. Finally, a fine-tuned pre-trained ResNet50 network was employed to automatically extract the shape features from the characteristic patterns, which achieved high-precision classification of the different states of the rotor system. The proposed method has practical engineering significance for the multi-fault classification of rotor systems.