Artificial Intelligence is a big wave rising and rising in the networks, and the more we progress in building smart models thanks to the large amount of data from our global scale, the more it is clear how much critical it is for our customer to keep their data private. This challenge is addressed by a new branch of the Machine Learning research called Federated Learning. It was born in the context of mobile phones, where AI models are improved globally without sharing the user’s personal data, and it is now moving to the industry to enable predictive models to grow smarter for everybody without customer data leaving their local context.
One of the contributions that R&D department of Ericsson Telecomunicazioni is providing to TERMINET project is related to the implementation of the use case on optical modules predictive maintenance, able to exploit this concept.
It is important to recall that the worldwide mobile network only in the “last mile” is wireless, from antennas to mobile phones or industrial mobile terminals. Most of the network traffic runs over laser beams in glass fibers connecting antennas, radio units, optical fronthaul units, baseband units, routers and cloud nodes in the data centers. The key components to drive light into the fibers are the pluggable optical modules, which are small and relatively inexpensive laser transceivers. They are deployed in tens of millions every year in every segment of the network and they are very reliable. However, due to the huge numbers, it is possible that some laser divers degrade their performances faster than the expected lifetime, or that they are defective at deployment, in a way that it is not discovered immediately. The component degradation or the exposure of some defects of birth is made worst by environment conditions, which are especially critical in the case of the outdoor radio units. It is therefore important to identify faulty modules in advance to prevent expensive field operations in case of module failure and operate the swap within the planned maintenance schedule.
In this context the TERMINET Federated Learning Framework enables to learn fail cases from all the networks, without the need of sharing sensitive data on specific network performance. Ericsson contributes with a patented process to generate prediction of the behavior of the modules in harsh environmental conditions. The process is integrated in an edge application which interacts with TERMINET edge-based federated learning client to locally process optical modules data and generate predictions. The optical modules in the demonstrator (aka SFP modules, as Small Form Pluggable) are plugged into an Optical Fronthaul unit of the Ericsson family FH6000 flagship product with the involvement of the Optical Solutions and Fronthaul Ericsson R&D unit in Italy.  Â