The prediction of fatigue life for mechanical structures under multiaxial loadings is considered an important task in engineering design, but the conventional models are mostly limited to special loading conditions and materials. In this work, a deep learning and physics-based method is proposed for predicting the multiaxial fatigue life. The relationships between the damage parameters and fatigue life are quantified by establishing a deep belief network (DBN) model. The hyperparameters of the DBN model are adaptively adjusted using a particle swarm optimization (PSO) algorithm. Importantly, the input features of the PSO-DBN model are determined based on the physical model. The proposed method is validated through fatigue tests conducted under various loading paths. The results demonstrate the high effectiveness and applicability of the proposed method for accurately predicting multiaxial fatigue life.