AI-Embedded Optical Modules With Millisecond-Granularity Power Analysis for Autonomous Metro Transport Network and Field Trial

The rapid expansion of 5G-A/6 G and AI-driven data centers demands robust optical networks with real-time fault detection capabilities. To address this need, we propose an intelligent optical module for edge deployment featuring millisecond-granularity power sampling and AI-driven analytics for high-precision monitoring of optical failures. The system integrates a hardware-software co-design enabling 10 ms resolution with a dual-attentive optic-diagnose network (DAOD-Net), which combines conditional GAN-based data augmentation and a TimeTransformer architecture for enhanced fault classification. Real-time inference for multiple failure modes is achieved, including device failure, fiber break, pigtail detachment, connector loosening, and fiber bend. Field trials in the deployed network demonstrate an overall fault identification accuracy of 92.6% with the consistent >95% precision across all fault categories, outperforming traditional neural network methods by 14.7% in accuracy. This solution is promising for reducing optical fault diagnostic time from hours to seconds, advancing the state-of-the-art in network automation capabilities while providing a scalable framework for proactive maintenance in next-generation 5G-A/6 G and hyperscale data center networks.