Dynamic Abstraction of Optical Networks with Machine Learning Technologies

The emerging 5G network will bring a huge amount of network traffic with big variations to optical transport networks. Softwaredefined optical networks and network function virtualization contribute to the vision for future programmable, disaggregated, and dynamic optical networks. Future optical networks will be more dynamic in network functions and network services, with high-frequency network reconfigurations. Optical connections will last shorter than that of the static optical networks. It's straightforward that Programmable optical hardware will require a reduced link margin to improve the hardware utilization. To configure network dynamically, real-time network abstractions are required for both current links and available-for-deploy links. The former abstraction guarantees the established links not be interfered by the newly established link while the latter abstraction provides information for intelligent network planning. In this talk, we use machine-learning technologies to process the collected monitoring data in a field-trial testbed to abstract performances of multiple optical channels. Based on the abstract information, a new channel can be established with maximum performance and minimized interference on the current signals. We demonstrated the dynamic network abstraction over a 563.4-km field-trial testbed for 8 dynamic optical channels with 32 Gbaud Nyquist PM-16QAM signals. The work can be further extended to support complex optical networks.