“Virtual and Real Coexisting”: Zither Composition Practice and Aesthetic Reconstruction Through AI and VR Integration

This study presents a multi-modal signal processing framework for automatic guzheng composition and VR-based performance simulation. The main melody is extracted using a similarity matrix, while LSTM-DQN networks, combined with quantized reward mechanisms based on musical theory and playing techniques, generate harmonic sequences. Multi-channel sensory data—including audio, tactile feedback, hand gesture, and line-of-sight tracking—are processed and integrated to construct an interactive VR performance environment, forming a closed-loop feedback system analogous to multi-channel signal transmission and control in engineering. Experimental results demonstrate that the LSTM-DQN model achieves an average note prediction accuracy of 67.6% (σ=2.34%) and maintains high rhythmic fidelity, while user evaluations indicate satisfactory aesthetic quality (average score 3.04/5). The framework enables rapid transformation from compositional ideas to virtual performance outputs, highlighting the effectiveness of multisensor signal integration, temporal sequence learning, and real-time feedback in complex systems. This approach provides an engineering-oriented perspective for integrating AI-driven generative models with interactive feedback, supporting future applications in intelligent multi-channel signal processing and virtual performance control systems.