Emotion recognition is a critical area in brain-computer interfaces, with electroencephalography (EEG) shown to be effective for emotional analysis. In domain generalization, cross subject emotion recognition encounters significant generalization challenges, including excessive feature collapse and insufficient capture of EEG features. To tackle these issues, we propose an Anti-Collapse Cross-Domain Consistency Network (AC-CDCN), which leverages Maximum Mean Discrepancy (MMD) to reduce distribution discrepancies between source domains, facilitating the capture of domain-invariant features, and innovatively introduces an Anti-Feature Collapse Strategy (AFCS), which incorporates an Anti-Collapse Domain Discriminator (ACDD) and the code rate loss function, effectively preventing excessive feature collapse. Furthermore, we propose a Flexible Feature Rebalance Module (FlexiReMod), a plug-and-play component that enhances generalization and dynamic feature capture through feature fusion and attention mechanisms. Experimental results indicate AC-CDCN achieved 87.14% (±5.60) and 71.77% (±12.92) accuracy on SEED and SEED-IV datasets, underscoring its significant generalization advantage.