Myocardial pathology segmentation (MyoPS) is essential for the diagnosis, treatment, and prognosis of myocardial infarction (MI). Recent deep learning MyoPS models have shown promising performance on independent identically distributed data. However, due to the domain shift caused by multi-center data, many methods often fail to generalize to unseen data. In this work, we propose a novel domain generalization framework (SMCANet) for MyoPS, which leverages data statistics (mean and standard deviation) modeling (DSM) and feature covariance alignment (FCA) during training to learn domain-invariant information. Specifically, DSM first models the statistical information of the data images and then performs perturbation-based random sampling to enhance the diversity of data representations. FCA reduces feature discrepancies across different distributions while preserving domain content information through the use of covariance matching loss (CML) and cross-covariance loss (CCL). Experiments on cross-domain MyoPS datasets from CARE2024 MyoPS++ challenge demonstrate that our framework can achieve promising generalization performance without altering the network architecture or employing complex data augmentation networks.