To identify the common faults of each in-wheel motor used in electric vehicle under different signal sources, a novel fusion framework is proposed on the basis of Bayesian network (BN) and improved Dempster-Shafer theory (DST). Firstly, vibration and noise signals are used to build the BN diagnosis model for the operation safety of each in-wheel motor. Secondly, the posterior probabilities of BN diagnosis models under different operating conditions are fused. Then an improved DST based on entropy weight is proposed to reallocate the conflicting parts of the posterior probability of BN diagnostic models, and a new basic reliability function is obtained. Finally, the effectiveness of the method was verified with in-wheel motor bench test. It was shown that improved DST can effectively solve the problem of inter-evidence conflict and integrate the BN diagnosis posterior probability based on vibration and noise to realize in-wheel motor fault diagnosis.