The traditional leading eigenvector dynamics analysis(LEiDA) employs linear dimensionality reduction methods, which overlook the nonlinear information of the brain. Additionally, the use of k-means hard clustering for brain state identification results in rough classifications, potentially neglecting atypical sub-states and features. To address this issue, the phase coherence graph autoencoder(PCGAE) analysis method was proposed. This method utilized the phase coherence graph autoencoder for dimensionality reduction, followed by the combined soft-hard clustering analysis method for feature extraction and differential feature selection. Finally, the support vector machine was used to build a classification model to evaluate classification performance. The results show that the proposed method achieves superior clustering effects in dimensionality reduction and outperforms traditional LEiDA classification results, with an accuracy of 83.7%. These findings suggest that the proposed method provides a better approach for dynamic functional pattern analysis and disease classification.