Machine learning framework for timely soft-failure detection and localization in elastic optical networks

Abstract

This work proposes a soft-failure evolution and localization framework to detect and localize the root cause of future hard-failure incidents in a timely manner enabling repair actions to effectively take place with reduced operational expenses (OpEx). To model soft-failure evolution, the capabilities of an encoder–decoder learning framework are leveraged to forecast the progression of soft-failures over an extended time period. This enables timely detection of the event of a costly hard-failure to proactively schedule the necessary repair actions. Repair actions are subsequently guided by the soft-failure localization algorithm, triggered once the hard-failure event is predicted. Specifically, the root cause of a future hard-failure is localized through a correlation algorithm that ranks all the soft-failures suspected of causing the predicted hard-failure. It is shown that the proposed framework is capable of triggering a repair action several days prior to the expected day of a hard-failure, contrary to myopic soft-failure detection schemes that are based on rule-based fixed quality-of-transmission margins, ultimately leading to either premature repair actions or repair actions that are taken too late. This fact, along with the proposed soft-failure localization approach, shown to effectively rank the suspected soft-failures, leads to reduced OpEx associated with both the reduced frequency of repair actions that are triggered and the effort required by technicians to localize and repair the root cause of an expected hard-failure in a timely manner.