According to the problem of the evidence fusion error occurred frequently by the evidence theory of diesel engine fault with the coverage of high diagnosis, a evidence-fusion method based on attribute hierarchical model is proposed. Firstly, the essential difference factor that reflects the difference of evidence in the target layer is obtained by modifying the attribute weight algorithm of the target layer with cosine similarity, and the entropy weight that reflects the relevance of attribute layers is obtained by modifying the attribute layer weight algorithm with Bayesian network rules. Finally, through the fault level association characteristics and the CWAA operator to correct the evidence fusion rules, the weights of attributes in different levels are effectively fused, which can reduce the diagnostic error caused by system complexity. The bench test of R6105AZLD diesel engine shows that the accuracy and robustness of diagnosis have been greatly improved after introducing the method.