Simultaneous Multi-Point Force Detection via Specklegram Analysis and Deep Learning

This study presents a deep learning-based framework for simultaneous multi-position force recognition using optical specklegrams generated from a single multimode fiber (MMF). The proposed method employs an autoencoder architecture with skip connections to extract force-dependent features from high-dimensional speckle patterns. Unlike prior approaches limited to pointwise classification or narrow force ranges, this system performs continuous regression of force magnitudes at three distinct positions along a single MMF, spanning a range of 0.26–1.26N with a fine resolution of 0.01 N. The model achieves exceptional accuracy with a regression R² value of 0.9989 and over 99% of predictions falling within ±0.02N across all positions. The model’s performance remains consistent across temporally spaced datasets, confirming its robustness to environmental variation and generalizability beyond training conditions. This architecture enables effective handling of the nonlinear and high-dimensional nature of specklegram data while preserving localized spatial information. The proposed framework requires minimal hardware making it compact, scalable, and cost-effective. These results establish significant advancement in specklegram-based fiber sensing, with strong potential for deployment in applications such as robotic spines, smart prosthetics, structural health monitoring of bridges and pipelines, and biomedical force tracking, where distributed, real-time sensing is critical.