Rethink Training: Synthetic Data Powers Efficient Fiber-Optics Sensor Demodulation Neural Networks

Accurate demodulation of fiber-optics sensors (FOSs) is essential for real-world environmental monitoring. Recent advances in AI have shifted the demodulation techniques from traditional signal processing, which has limited generalization capabilities, to machine-learning (ML) approaches that circumvent the need for considering physical constraints. However, these ML methods typically require extensive, diverse datasets to train models from scratch, a requirement that is impractical in many scenarios. To overcome this limitation, this article introduces a brand-new FOS demodulation strategy that achieves robust performance using only synthetic data. Specifically, we initially collect real data pairs-wavelengths and transmitted intensities-using the wave decomposition capabilities of arrayed waveguide gratings (AWGs). This data is fed into a Wasserstein generative adversarial network (WGAN), which precisely controls the generation of specific data types and quantities. The demodulation network is then trained exclusively on synthetic data and evaluated against real datasets. Our extensive experiments on Fabry-Perot FOS pressure demodulation, demonstrate the method's effectiveness and superiority, achieving excellent wavelength interrogation. Moreover, the proposed method exhibits excellent generalizability across various scenarios, offering a cost-effective solution for FOS demodulation.