Existing fault diagnosis models are usually trained on limited data and are only applicable to independent and identically distributed data based on the closed-set assumption. When confronted with different operating conditions, locations, and machine data, epistemic uncertainty exists due to inconsistent data distribution. This causes the occurrence of overconfident misclassification for out of distribution samples, which greatly affects the reliability of fault diagnosis. A trustworthy fault diagnosis method is proposed in this paper. Firstly, the uncertainty carving method is established by using the measure of uncertainty distribution. Based on the support vector machine, the distances from uncertain variables to hyperplanes are derived which obey linear uncertainty distribution. Then, a new uncertainty support vector machine (USVM) model is constructed and the corresponding equivalent solution method is given, which effectively reduces the influence of epistemic uncertainty on the fault diagnosis. Finally, the proposed model is used in the bearing dataset to validate the effectiveness.