Federated Learning for Cyber-Physical Systems in Smart Environmental Monitoring

Abstract

Localization in complex environments poses significant challenges due to factors like signal attenuation from obstacles and environmental variations. This research introduces a deep learning-based framework integrating a convolutional denoising autoencoder (CDAE) and a convolutional neural network (CNN) for source localization. The CDAE mitigates signal attenuation effects caused by walls, while the CNN accurately localizes the source based on spatial radiation patterns. To enhance generalizability, the framework employs a federated learning (FL) approach, enabling collaboration across models trained in distinct environments without compromising data privacy. A novel aggregation strategy, guided by a complexity score incorporating factors such as wall intersections, proximity, and attenuation levels, optimizes model contributions to the global model. The system models radiation propagation over a 100m x 100m area, with sensor readings influenced by walls and other obstacles. Experimental results show localization errors between 24 meters in dense sensor networks and 3-6 meters in sparse networks. The FL approach demonstrated superior performance, outperforming individual environment-specific models 89% of the time in dense networks and 73% of the time in sparse networks. The global model consistently maintained robust performance, validating the effectiveness of the complexity score in facilitating generalization to unknown environments. This framework offers a novel solution to source localization in complex, non-IID settings by leveraging federated learning to balance adaptability and accuracy. The results highlight its potential for advancing localization techniques, particularly in unknown scenarios, scenarios with limited data and diverse environmental conditions.

Date of Award	11 Dec 2024
Original language	American English
Supervisor	Singh (Supervisor)