Tunnel collapse risk assessment is a multi-attribute decision problem due to many influencing factors. It is difficult for the assessment method of a single information source to fully consider all risk factors, leading to bias in the prediction results. To assess the tunneling collapse risk and provide a more accurate risk-controlling strategy, this research proposes a new multi-source information fusion approach that combines cloud model (CM), support vector machine (SVM), and evidence-based reasoning (ER). Multiple sources of information were analyzed to obtain different collapse risk assessment models (where classification probability values for visual inspection data are obtained from SVM, and probability values for monitoring data are obtained from the cloud model). The quality of each model is evaluated by reliability and importance weights. The ER theory is then applied to fuse the results of each assessment model to give an overall collapse probability risk assessment. Compared with the D-S theory, the ER rule has more advantages in dealing with high-conflict information. When the risk assessment results of different single information sources are inconsistent, the fusion by the ER rules considers the importance weight and credibility of the assessment results, which is more suitable for the high-conflict information fusion. The novel approach has been successfully applied in the case of Yutangxi tunnel of Pu-Yan Highway (Fujian, China). The results indicate that the proposed multi-source information fusion method has an evaluation accuracy of 87.5%, while the single-source information method has an accuracy of less than 70%. Furthermore, the fusion model has excellent performance even if the risk result of different models has high conflict.