<p>Magnetotelluric (MT) is an important geophysical exploration technique widely used to investigate deep crustal structures and mantle electrical conductivity. However, due to the weak amplitude of natural electromagnetic signals, MT data are highly susceptible to interference from external sources such as industrial power systems and transportation equipment, particularly in low-frequency bands. The superposition of noise and useful information severely degrades data quality and inversion accuracy. Traditional signal-processing approaches rely on empirical parameters and idealized assumptions, making them ineffective for complex and time-varying noise. Although deep learning has demonstrated strong results for noise removal, obtaining genuinely noise-free field MT recordings for supervised training is generally infeasible. Consequently, many studies employ synthetic data generated via forward modelling to obtain sizeable training sets, though such synthetic data may not fully capture all complexities of real-world measurements. To overcome these limitations, this paper proposes a Physics-Constrained Self-Supervised Neural Network for Multi-Channel Magnetotelluric Noise Separation (P-MTNS), enabling automatic noise separation without labeled or clean training data. By incorporating physical constraints, the method exploits intrinsic correlations among multi-channel electromagnetic observations. A loss function constrained by phase consistency and time-domain coherence is introduced to guide the network in learning physically meaningful signal representations. Experiments on synthetic and field datasets demonstrate that the proposed method effectively separates diverse noise sources and substantially improves inter-channel coherence: the band-averaged coherence for the <i>E</i>_<i>x</i>-<i>H</i>_<i>y</i> pair increases from 0.8330 to 0.9978. These results indicate that P-MTNS provides a more reliable data foundation for subsequent MT inversion.</p>