Abstract: Surface electromyography (sEMG) has important theoretical research value and practical application significance in fields such as human-computer interaction (HCI), but existing methods often face challenges of significantly reduced accuracy and high computational complexity when processing signals from multiple subjects. To address these issues, a robust and intelligent sEMG recognition method based on the multi-scale integrated sequence deep belief network (OmniXceptionDBN) is proposed. Firstly, the raw signals is processed using singular spectrum analysis and fast Fourier transform, and then the OmniXceptionDBN algorithm is constructed by combining XceptionTime, OmniScaleCNN, and deep belief networks (DBN) for sEMG recognition and experimental verification. The results indicate that algorithm achieves a classification accuracy of 97.2% for a single individual subjects and 85.9% for multiple subjects without any additional operations. The proposed approach effectively addresses the challenges of accuracy degradation and high computational complexity by traditional methods in cross-subject signal processing, providing an efficient and robust solution for the field of sEMG intelligent recognition.