Predictive Maintenance for Optical Networks in Robust Collaborative Learning

Khouloud Abdelli, JOO YEON CHO

Keywords: predictive maintenance, federated learning, machine learning, anomaly detection, multi-party computation, autoencoder

Abstract: Machine learning (ML) has recently emerged as a powerful tool to enhance the proactive optical network maintenance and thereby, improve network reliability and operational efficiency, and reduce unplanned downtime and maintenance costs. However, it is challenging to develop an accurate and reliable ML based prognostic models due mainly to the unavailability of sufficient amount of training data since the device failure does not occur often in optical networks. Federated learning (FL) is a promising candidate to tackle the aforementioned challenge by enabling the development of a global ML model using datasets owned by many vendors without revealing their business-confidential data. While FL greatly enhances the data privacy, a global model can be strongly affected by a malicious local model. We propose a robust collaborative learning framework for predictive maintenance on cross-vendor in a dishonest setting. Our experiments confirm that a global ML model can be accurately built with sensitive datasets in federated learning even when a subset of vendors behave dishonestly.

One-sentence Summary: We propose a robust collaborative learning framework for predictive maintenance on cross-vendor in a dishonest setting and confirm the performance by experiments.