A federated learning framework with GAF-CNNs for robust sensor fault detection in WSNs

The widespread adoption of the Internet of Things (IoT) has made sensor data essential for numerous applications, with Wireless Sensor Networks (WSNs) serving as key enablers. However, their deployment in dynamic and harsh environments makes them vulnerable to faults such as communication failures, hardware malfunctions, and software errors, threatening both data integrity and overall system performance. Traditional machine learning methods for fault detection often rely on centralized training, which raises concerns regarding scalability and data privacy. In this work, we simulate five common fault types of bias, drift, spike, erratic, and stuck by injecting them into normal temperature sensor data. The temperature sensor readings are transformed into two-dimensional Gramian Angular Field (GAF) images. To mitigate the issue of data scarcity a Wasserstein Generative Adversarial Network (Wasserstein-GAN) is employed to generate synthetic samples. An optimized Convolutional Neural Network (CNN) is then deployed within a Federated Learning (FL) framework to perform decentralized fault detection while preserving data privacy across distributed nodes. Experimental results demonstrate that the integration of GAF, Wasserstein-GAN, and FL yields a robust, scalable solution for accurate fault classification in IoT-enabled WSNs.