Emotion Recognition Based on Fusion of Topological Features and Trajectory Images Derived from EEG Phase Space Reconstruction

Electroencephalogram (EEG) signals, as a direct measure of the brain’s cortical electrophysiological activity, can objectively capture emotion-induced neural changes. Phase space reconstruction is an effective method for processing nonlinear time series. It maps time series to a high-dimensional phase space, thereby better preserving subtle dynamic information in the signal. This paper proposes a method for emotion recognition in EEG signals based on phase space reconstruction. First, the macro-topological features of the trajectories are constructed via phase space reconstruction. The time delay and embedding dimension are then optimized using the minimum cross-prediction error and the G-P method, followed by dimensionality reduction to a two-dimensional plane via local linear embedding. Building on this foundation, and in response to the limitations of manually designed features, we further propose a deep learning-based method for extracting multiscale dynamic features from trajectory images. The designed GN-MVXXS framework, which utilizes a granularity-adaptive module to adaptively switch the receptive field and a noise-filtering module to suppress isolated noise points, thereby effectively uncovers microscopic evolutionary features at the image level. Finally, to leverage the complementary strengths of macro- and micro-level information, we propose a fusion method based on dynamic attention. This approach aligns the dual representational dimensions through global average pooling and nonlinear dimension expansion, and utilizes a dynamic attention mechanism to adaptively assign feature weights, enabling the model to collaboratively enhance both overall dynamic patterns and local details based on sample characteristics. The experimental results show that the model achieved an accuracy of 96.11% in the three-class classification task on the SEED, 86.33% in the four-class classification task on the HIED, and 83.67% in classification across normal-hearing and hearing-impaired individuals, significantly outperforming single-feature models and traditional fusion methods.