Model-driven learning for OTFS detection based on an enhanced GNN with prior information

Orthogonal Time Frequency Space (OTFS), characterized by its unique time-frequency orthogonality, has demonstrated significant advantages in overcoming the difficulties caused by time-varying channels in wireless communications. To realize the full capabilities of OTFS, designing an efficient signal detector is crucial. Traditional detectors often fail to perform adequately under complex channel conditions. Therefore, Graph Neural Networks (GNNs), known for their powerful feature representation and generalization capabilities, have been introduced for OTFS signal detection. However, conventional GNN-based detectors do not effectively utilize prior information and edge features. To address this, we propose a Prior Information-Enhanced GNN (PI-EGNN) detector, which improves signal estimation by integrating prior information and fully leveraging edge attributes. Additionally, a linear attention mechanism is introduced to further enhance the overall network performance. Simulation results demonstrate that, compared to state-of-the-art detectors, the proposed PI-EGNN detector shows improved performance.