Very Large Scale Distributed Training Data and Federated Learning: A Case Study in DAS

Abstract

FOTAS is an optoelectronic system used to measure acoustic signal values along the length of a fiber optic cable. By creating an acoustic profile along the length of the fiber optic cable, the system uses deep learning to detect and classify events in the physical environment. It provides high detection accuracy over long distances, it is reliable against electromagnetic interference, and it is also reliable during harsh working conditions. In order to train deep learning models known to the FOTAS system, the data must be stored in a central location. However, the migration of data is difficult because of its large size. After installing sensors in various regions and collecting data, it is very expensive to migrate this data to a server for model training. In addition to the difficulties that may emerge from data migration, problems related to data security may also occur. Federated learning enables model training on a given device by using data from server and client devices, which eliminates the need for a problematic data migration process. Client devices process data locally and share model weights with the server device. Therefore, there is no need to transfer large amounts of data. In this study, we apply the federated learning method on the FOTAS system and compare the results with the results from using classical methods. By selecting two client devices and one central device, we transferred model weights between remote devices. By choosing asynchronous device features, we observed how the federated learning system responds to asynchronous conditions. Based on the results of our study, we can conclude that federated learning can be used in commercial systems.