Sleep disorders and sleep disturbances are gaining increasing attention, and accurate sleep staging is key to assessing sleep quality. Existing studies utilize spatio-temporal features of multi-channel physiological signals such as EEG and spatial topological information of brain regions for sleep staging, but they suffer from insufficient mining of spatio-temporal correlations, neglect of the nonlinear geometric structure of signals, and feature redundancy, leading to low accuracy in sleep stage classification. To address these problems, this paper proposes a multi-view spatio-temporal convolutional integration network based on manifold learning. Firstly, three types of feature representations of multi-physiological signals are constructed via short-time Fourier transform, discrete wavelet transform, and raw signal processing. Secondly, an attention module that fuses Euclidean space and manifold geometry is designed to represent and enhance multi-modal physiological signals on the Riemannian symmetric positive definite manifold, mining their potential spatio-temporal representations. Finally, combining these three types of sleep features from adjacent time segments, long short-term memory (LSTM) is used to model temporal dependencies to accomplish sleep stage classification. Experimental results on public sleep stage classification datasets show that the proposed method outperforms existing mainstream methods in accuracy, F1-score, and Kappa coefficient, and ablation studies validate the effectiveness of the manifold attention mechanism and multi-view feature combination. This work provides a new method for nonlinear fusion of multi-modal physiological signals.