The advantage of intelligent fault diagnosis (IFD) based on industrial big data lies in the powerful feature extraction ability of machine learning models. However, it has become extremely difficult to apply machine learning-based fault diagnosis models to the actual industry due to the problem of labeled data insufficiency and class imbalance. Ensemble learning, which leverages the aggregation of multiple base classifiers to effectively utilize data, is regarded as a promising approach to address this issue. In this study, we propose an ensemble learning framework that integrates multiple stacked capsule autoencoders (SCAEs) for accurate fault diagnosis. The proposed ensemble framework introduces a novel method for evaluating intrinsic templates based on a symmetric graph Laplacian with the aim of selecting capsules that can effectively reduce information redundancy. Finally, a new decision fusion method is proposed to achieve the decoupling of composite fault labels by DS evidence. The proposed method is validated to achieve fault classification accuracy of up to 100% and 91% on datasets with sufficient and insufficient samples. In addition, the accuracy is higher than 94% on four imbalanced datasets. The experimental results demonstrate that the proposed method exhibits enhanced resilience against dataset defects, thereby offering more adaptable and reliable fault diagnosis services in real-world industry.