Multimode fiber (MMF) specklegram sensing shows considerable potential for structural health monitoring (SHM). However, its application in multi-parameter scenarios is still limited by pronounced crosstalk and the high computa-tional cost of conventional convolutional networks. In this study, we propose a lightweight dual-task deep learning model, Light-MS-SparseNet, which simultaneously decodes the bending degree and bending position from a single specklegram. The architecture combines multi-scale and dilated convolutions to expand receptive-field coverage, adopts depthwise separable convolutions to substantially reduce the number of parameters and computational com-plexity, and introduces a variance-driven Light-Sparse-Activation module together with a context-gating mechanism to suppress redundant activations and enhance global consistency. In the experimental design, a self-built dataset covering five bending degrees (no bending, 1.25 cm, 2.5 cm, 3.0 cm, and 4.3 cm) and five discrete positions (0-4 m) was constructed for model training and validation. The experimental results show that Light-MS-SparseNet achieves 100% accuracy in bending-degree classification and 99.85% accuracy in bending-position classification. With only approximately 91.2K parameters, the joint metric reaches 99.9762%, outperforming ResNet and other deep-learning-based baselines in terms of both accuracy and efficiency. These results indicate that Light-MS-SparseNet can effectively mitigate multi-parameters interference while maintaining feasibility for edge deployment, providing a reusable and scalable paradigm for distributed structural health monitoring.