All-Optical Wearable Gesture Sensing With a Learnable Photonic Linear Filter Bank for Near-Sensor Classification

Le résumé fourni par la source

Wearable gesture recognition systems typically rely on multi-channel electrical sensing followed by analog-to-digital conversion, which can incur substantial data movement, latency, and power overhead. Here, we present a photonic feature extraction neural network (PFENN), which provides an electrical-transduction-free optical sensing interface and a photonic preprocessing front end. Gestures are encoded directly in the optical domain via calibrated macro-bending loss in standard single-mode fibers attached to finger joints. The resulting multiple optical intensity streams are processed by a programmable silicon photonic chip that implements a learnable photonic linear filter bank based on a Mach–Zehnder interferometer (MZI) mesh. After this optical linear preprocessing step, a single photodiode and a compact electronic classifier complete recognition. This work demonstrates a training-to-deployment workflow by mapping software-trained weights onto on-chip phase shifters using measured device responses. On a 13-gesture, single-subject dataset (200 samples per class), the system achieves 96.63% accuracy. The system operates without any active electro-optic modulation at the sensing point; on-chip MZMs serve as static, bias-controlled weights during inference. These results support a practical route toward low-overhead near-sensor wearable intelligence by combining passive fiber sensing with learnable photonic linear preprocessing.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.