A Machine Learning Approach for Dynamic Optical Channel Add/Drop Strategies that Minimize EDFA Power Excursions

We demonstrate a machine learning approach to characterize channel dependence of power excursions in multi-span EDFA networks. This technique can determine accurate recommendations for channel add/drop with minimal excursions and is applicable to different network designs. Introduction The advent of dynamic optical networks demands constant agility and configurability in response to traffic and fault handling. As networks grow in both capabilities and complexity, the concept of cognitive networks – systems that can autonomously monitor, optimize, and adapt – is particularly promising to improve the networks’ management and resilience. As a crucial component of modern optical transport networks, Erbium Doped Fiber Amplifier (EDFA) has the ability to achieve economic regeneration of dense wavelengthdivision multiplexing (DWDM) channels and extend the reach of optical communication beyond the confines of cities and continents. However, EDFA systems face an unsolved challenge of channel-dependent power excursions. Modern EDFA systems employ automatic gain control (AGC) to maintain the output power levels of the amplifier within a tolerant regime. In cascaded EDFA networks, upstream AGC ensures appropriate optical power levels for downstream amplifiers and receivers. However, AGC maintains the global mean gain, while each channel sees a wavelength dependent gain. If a channel with high gain is added, AGC responds to an increase in mean gain by reducing the gain on all channels. This leads to the high-gain channel effectively stealing power from lower-gain channels. Conversely, adding a low-gain channel feeds power to higher-gain channels. In both cases, the power excursion increases the disparity among channel powers; this discrepancy may be further exacerbated by downstream EDFA spans. We thus define undesired power excursions as the ones that increase the standard deviations (STD) of the output power levels. Proposed solutions and limitations The characteristics of the excursion depend on the types of EDFAs, the gain-control mechanisms, and the number of EDFA spans and light paths (LP); therefore, it is difficult to derive an analytical description that applies to all systems. Consequently, past proposed solutions focused on fully characterizing a specific EDFA system and reducing excursions by optimizing input power levels, balancing input channels, or adjusting the pumping level of the amplifier. These techniques, while effective on the specific systems analyzed, are not necessarily transferrable to different networks. They also rely on the deterministic model of the gain profile, which is difficult to acquire for livenetwork equipment that cannot be disrupted. Preventative approaches such as optimized wavelength assignment algorithms have been shown to reduce the excursions, but at the tradeoff of spectral efficiency. Case-based reasoning (CBR) is also applied to make heuristic guesses on EDFA tuning, but it requires a large number of past records to be effective. We present an efficient, low-overhead machine learning (ML) engine to characterize the channel dependence of power excursions in multi-span EDFA links. Historical snapshots of the network are collected and mathematically generalized. Once the ML model is trained, it is able to predict the best practices of channel add/drop to alleviate undesired excursions. The approach is non-disruptive and applies to EDFA networks of different designs. Fig. 1 illustrates the functionalities of the ML engine. Fig. 1: Functionalities of the ML engine for minimizing undesired power excursions. Experiment design We construct the multi-span AGC-enabled EDFA network shown in Fig. 2. The WDM sources transmit 24 DWDM channels from ITUT grid Ch. 21 to Ch. 44 with 100 GHz spacing, which are combined via a wavelength-selective ML Engine EDFA Cascade OPM WDM Sources