In an era where cryptographic devices are widely used, research on Side-Channel Analysis (SCA) has attracted significant attention. While Deep Learning-based SCA (DLSCA) demonstrates notable advantages with the rapid development of deep learning, its performance remains constrained by reliance on single leakage model, leading to inefficient attacks and inadequate information utilization–particularly in grey-box scenarios where optimal leakage point identification proves challenging. Traditional ensemble methods attempting multiple leakage points combination are limited to either homogeneous models or manual feature fusion, failing to address the collaborative optimization of heterogeneous leakage features. To overcome these limitations, we propose a side-channel analysis framework based on multiple leakage models ensemble. This approach defines the training process of neural networks under different leakage models as base learners and employs Bagging ensemble learning to combine complementary leakage characteristics from heterogeneous leakage models. Thereby, it collaboratively optimizes multi-dimensional physical leakage features, significantly improving the efficiency and robustness of key recovery. Furthermore, based on information entropy and Jensen's inequality, we rigorously prove that the predictive distribution of the ensemble model more closely approximates the true key distribution, providing theoretical support for multiple models ensemble. Experimental results show that the proposed modular ensemble framework supports flexible combinations of arbitrary leakage models and exhibits excellent generalization capability and stability across different platforms and adversarial environments.