Abstract: The objective of Domain Generalization (DG) is to generalize knowledge acquired from multiple source domains and apply it to unseen target domains. Current methods commonly rely on modeling the statistical correlation between input data and labels to learn representations independent of the domain. However, these approaches overlook the inherent causal mechanisms, leading to a failure in generalization when the statistical correlation changes with the distribution of the target domain. In response, this paper proposes a Domain Generalization algorithm based on Causal Representation Learning (CRLDG). This algorithm first introduces a causal interference mechanism to isolate the causal factors between the input data and labels in the source domain, and then decomposes these causal factors to obtain representations independent of each dimension. To enhance the aggregation of causal representations within the same category, a central alignment method is introduced. In the experiments, the CRLDG method was evaluated using ResNet-18 and ResNet-50 as the backbone networks on the PACS dataset, achieving average classification accuracies of 84.75% and 88.29%, respectively. Furthermore, when ResNet-18 was used as the network backbone of the Office-Home dataset, the CRLDG method achieved an average classification accuracy of 66.79%. In short, the CRLDG method outperformed all comparative methods, and the effectiveness of CRLDG on PACS and Office-Home datasets was validated through this series of experiments.