Abstract: The utilization of heterogeneous data is essential for spacecraft damage diagnosis, yet traditional fusion methods suffer from high-confidence misclassification due to strong noise and inter-source conflicts. To address these issues, an anti-conflict fusion framework based on evidential deep learning (EDL) is proposed. A two-stream vision transformer (ViT) backbone is employed to extract damage textures from infrared images and transient energy features from vibration time-frequency diagrams, whose outputs are mapped to a Dirichlet distribution for uncertainty quantification. Furthermore, a nonlinear belief gating mechanism is designed to dynamically adjust probability assignments based on uncertainty, adaptively suppressing interference from unreliable sources. Experimental results demonstrate that the proposed method achieves an overall diagnosis accuracy of 98.26% on the heterogeneous dataset, surpassing the performance of single-source methods. Specifically, the recognition rate for micro-cracks, which tend to be missed with a single vibration modality, is improved from 82.1% to 92.9%, effectively resolving the synthesis paradox in heterogeneous data fusion.