Telemetry and Data Collection for Artificial Intelligence in Optical Systems

Abstract

Increasing internet traffic, along with rising complexity of optical communication systems have motivated the development of novel experimental, mathematical and computational tools. Optical networks are being used in more applications such as 5G wireless systems in which internet traffic continuously varies in both time and location. Slow reconfiguration in optical networks leads to network inefficiency and under-utilization of resources. This problem created interest in realizing a reliable method that can perform real-time switching and wavelength provisioning, in order to adapt to the fluctuating traffic patterns. Such systems having rapid-switching capabilities are susceptible to power excursions that diminish the Quality of Transmission (QoT). Widely used reconfigurable optical add-drop multiplexer (ROADM) based transparent optical networks, which do not need optical-electrical-optical (O-E-O) conversion, can potentially benefit from signal quality monitoring in the optical domain (i.e. optical performance monitoring). Additionally, for optical switching the QoT must be evaluated prior to establishing the new wavelength route in a given transparent wavelength division multiplexed (WDM) network. The complexity of optical communication systems has resulted in researchers looking for new methods to determine QoT. Artificial Intelligence (AI) and its sub-field of Machine Learning (ML) together with telemetry to collect more data have attracted interest for use in optical networks to improve QoT prediction. In this thesis, a non-disruptive optical probe based real-time telemetry solution capable of measuring QoT and performing optical performance monitoring (OPM) was built. Our device, the non-disruptive optical probe monitor (ND-OPM) uses short optical pulses to probe a network’s performance to obtain data for QoT metrics. Unlike other methods, probing can be done before wavelength provisioning/reconfiguration, supporting wavelength routing and assignment decisions prior to switching. Using this technique, the wavelength dependent gain of an Erbium-doped fiber amplifier (EDFA) was estimated non-disruptively, along with prediction of the channel power and optical signal to noise ratio (OSNR) of a provisioned channel. Additionally, a remote and automated SDN-based process for running experiments and collecting large datasets required to characterize optical systems deployed in the COSMOS Testbed was studied. The experimental results include measurements of amplifier gain and power spectra for randomized channel loading configurations. This data can be used to develop ML algorithms for QoT estimation. Finally, we have reviewed some of the AI/ML techniques that have been applied to the field of optical communications.