Applied Horizontal Federated Learning with Deep Neural Networks on Non-IID Data in Manufacturing

Abstract

This master’s thesis focuses on the application of Horizontal Federated Learning for industrial use cases. The study aims to determine to what extent Federated Learning can be considered a viable alternative to traditional centralized machine learning, particularly in addressing the challenges posed by non-IID (non independent and identically distributed) data. Manufacturing companies are increasingly using machine data to optimize production processes. However, concerns regarding data security arise, as this data often contains sensitive process knowledge. Federated Learning offers a solution by keeping training data on the local devices of participants, transmitting only model parameters to a
central server for aggregation. Two practical use cases, the workpiece processing recognition and the ski edge detection, were selected to evaluate the performance of Federated Learning compared to centralized machine learning. For these scenarios, different Federated Learning algorithms (FedAvg, FedProx, FedOpt, Scaffold, and Ditto) were examined for their suitability. The implementation employed the NVIDIA FLARE Federated Learning framework. Additionally, an NVFLARE experimentation framework was developed during the course of the thesis to simplify the hyperparameter tuning process. Comprehensive hyperparameter tuning and model evaluation were performed for both use cases using the developed
NVFLARE experimentation framework. The results demonstrate that Federated Learning is a promising alternative to centralized machine learning, particularly in terms of maintaining data security and model flexibility. Experimental investigations showed that Federated Learning produces robust global models that ensure balanced performance across participants. Additionally, it enables the development of personalized models for individual participants, which deliver
better predictions than a generic global model.

Date of Award	2024
Original language	English (American)
Supervisor	Michael Affenzeller (Supervisor) & Verena Angela Pietsch (Supervisor)