In view of the difficulty in analyzing the strong nonlinear coupling effect between four-wave mixing and stimulated Raman scattering in single-mode optical fibers, this paper introduces a novel multi-scale physically constrained network (MSPC-Net), which effectively integrates fundamental physical mechanisms with advanced neural network techniques. The proposed model incorporates the frequency domain residual derived from the nonlinear Schrödinger equation directly into the network optimization procedure as a differentiable physical constraint term. This strategic inclusiveness ensures that the learning process is consistent with the fundamental physical principles governing light propagation in optical fibers. Furthermore, the model architecture adopts a multi-scale dilated convolution module specifically designed to capture and fuse features across different granularities, including fine local spectral details, intermediate-range broadening effects, and long-range attenuation trends. This multi-scale approach can realize the simultaneous and high-precision inversion of both separated spectral components and critical physical parameters.

Experimental evaluations are conducted using single-mode quartz fibers with lengths of 250 meters and 500 meters, respectively. The results demonstrate that the Stokes spectra reconstructed by MSPC-Net achieve remarkably low root mean square errors, only 0.014 and 0.0173 for the two fiber lengths respectively. This performance represents a reduction of more than 68% compared with that of traditional convolutional neural networks. Additionally, the average absolute errors of frequency offset prediction are as low as 0.03 nmr and 0.04 nm, with an accuracy improvement of approximately 90% compared with those of existing state-of-the-art methods. Under noisy conditions with a signal-to-noise ratio of 6 dB, the model maintains an exceptional detection accuracy of up to 95.3% for identifying four wave mixing (FWM) sub-peak information, while keeping the pseudo-peak rate below 4.7%.

Owing to the embedded physical constraints and lightweight structural design, the proposed model shows just a 9.8% increase in root mean square error even under challenging noise conditions with a signal-to-noise ratio of 15 dB. Moreover, MSPC-Net demonstrates satisfactory real-time processing capabilities, making it suitable for deployment on embedded devices. This practical efficiency makes the model a promising solution for optimizing high-power optical communication systems and advancing distributed optical fiber sensing applications. By successfully combining strict physical laws with multi-scale feature extraction, this research presents an effective approach to resolving the analytical difficulties associated with complex nonlinear effects in long-distance optical fibers, while significantly improving both the theoretical consistency and noise robustness of the prediction outcomes.