Deep learning-driven optical fiber sensor for 2D vector flow velocity detection

Flow velocity, direction, and temperature are key parameters for characterizing underwater flow fields. However, traditional optical fiber sensors suffer from low integration, poor sensitivity, and severe cross-sensitivities. Inspired by the lateral line receptors of fish, we propose a biomimetic cilium-based multimode fiber speckle sensor with multi-head decoding for multi-parameter sensing. The sensor employs a cylindrical cilium structure, embedded in a PDMS, to convert external fluid disturbances into modal coupling and phase shifts within the multimode fiber, resulting in local speckle displacement and micro-texture variations for optical encoding of flow direction and velocity. Temperature variations induce near-common-mode phase drifts that manifest as smooth global changes in the speckle field, producing identifiable temperature features. A shared backbone multi-head deep network jointly decodes the three parameters to mitigate cross-sensitivity and task interference. In the velocity-direction dataset, covering a tested flow-velocity range of 0.167–0.5 m/s, the proposed method achieves a flow-direction accuracy of 93.62% and a flow-velocity MAE of 0.0137 m/s. Across the tested temperature range of 5-40°C, it achieves a temperature MAE of 0.5412°C. The results validate a compact laboratory scale fiber sensor architecture for controlled multi-parameter optical sensing and demonstrate its potential for future underwater monitoring applications.