It is difficult to extract the fault features of gears under complex operation conditions. The traditional fault diagnosis and recognition accuracy are easily affected by the manual feature extraction, and the information obtained by a single sensor is not comprehensive. To solve the above problems, a gear fault diagnosis method based on deep belief networks (DBN) and information fusion was proposed in this paper. Firstly, the vibration signals collected by each sensor were fused by multi-sensor information fusion technology at the data layer, and then DBN was used for adaptive feature extraction to achieve fault classification. In order to avoid the problem of model recognition accuracy degradation caused by artificial selection of DBN structural parameters, an improved shuffled frog leaping algorithm (ISFLA) was proposed to optimize DBN structural parameters. Experiments show that the information fusion and optimization methods proposed in this paper have higher fault recognition accuracy than BP neural network, DBN, and single sensor fault diagnosis.