Robust Lightweight Mode Decomposition Based on Geometric Normalization and Multi-Scale Feature Extraction

Mode Decomposition (MD) is an important tool for characterizing the modal energy composition and mode states. However, existing deep learning-based approaches have primarily focused on prediction accuracy under ideal conditions, while the stability of their performance under varying input conditions has received considerably less attention. To address non-ideal perturbations, including translation, scaling, rotation, blur, and noise, we propose a lightweight and robust MD method based on RepViT. Several robustness-oriented feature extraction modules are introduced to enhance local–global spatial, frequency-domain, multi-scale, and spatial-channel representations. Moment-Based Geometric Normalization and BlurPool are further introduced to improve the stability of the model against geometric variations and downsampling-induced sampling sensitivity. In addition, a robustness evaluation framework is established using a Clean test set and five graded composite perturbation levels, together with individual-perturbation tests, the mean composite–perturbation correlation coefficient, and the relative correlation degradation rate. Experimental results show that the proposed method achieves a Pearson correlation coefficient of R=0.9317 even under the strongest composite perturbation. Across the five composite perturbation levels, the mean correlation coefficient reaches 0.9728, while the relative correlation degradation rate is only 2.66%. These results demonstrate that the proposed method substantially improves the stability of mode purity prediction under diverse perturbation conditions while maintaining a lightweight architecture, providing a practical reference for robust MD and deployment in fiber-optic systems.