A hybrid physics-aware and self-supervised generative framework for heterogeneous DFOS traffic monitoring

Deep learning has been widely adopted in distributed fiber optic traffic monitoring. However, the performance of most models is constrained by the scarcity of high-quality annotated vibration data. This issue is further exacerbated by the high sensitivity of fiber signals to environmental coupling conditions, which severely limits cross-scenario generalization. To address this bottleneck, a hybrid generative data augmentation framework is proposed. Unlike conventional augmentations that disregard sensor noise disparities, or unconstrained generative models (e.g., GANs) that frequently induce topological breakages in slender signal structures and compromise annotation boundaries, our approach explicitly enforces spatial and physical fidelity. First, a physics-aware fusion mechanism based on a decoupled vibration-background library is introduced. By embedding vehicle instances into heterogeneous noise profiles via gradient-domain blending, this strategy explicitly bridges domain gaps in signal-to-noise ratios. Then, a self-supervised data generator employing a masked autoencoder architecture is developed. Crucially, even under high masking ratios, this strategy infers global signal topology from sparse structural anchors. This uniquely preserves the long-range spatiotemporal continuity of vibration trajectories and strictly locks geometric boundaries, thereby ensuring absolute semantic consistency for zero-cost label reuse. Finally, empirical results demonstrate the framework’s strong universality. Specifically, the proposed strategy elevates the average detection precision from a baseline of 41.6% to 75.6%, yielding a net improvement of 34.0%, with the most significant gain reaching 48.8% for high-capacity architectures. Furthermore, models trained exclusively on synthesized counterparts (synthetic-only generalization) surpass the real-data baseline by 21.8%, thereby establishing robust Sim-to-Real transferability without subsequent labor-intensive recalibration.