Federated Learning-Based Sensor Fault Classification in Label-Scarce Sensor Networks

Ensuring reliable data collection from sensor nodes is vital for the success of Internet of Things (IoT) applications such as smart cities, industrial automation, and other data-driven systems. As sensor data underpins key decision-making processes, undetected faults can compromise system performance, leading to misinterpretations and costly operational disruptions. Therefore, timely and accurate fault classification is essential to preserve the integrity and resilience of IoT deployments. This article proposes a novel method for classifying five common types of sensor faults across homogeneous sensor nodes using a combination of federated learning (FL) and transfer learning (TL). The approach supports two categories of client nodes: source nodes with labeled data and target nodes without labels. Federated training at source nodes exploits dataset similarity with the unlabeled target node, while a central server adaptively weights each client’s contribution based on its relevance to the target sensor before aggregation. Across various source-target configurations, the proposed method consistently outperforms baseline approaches. It yields 2%–11% higher accuracy compared to the widely used FedAvg method. In addition, it improves over a centralized LSTM-based model by +3.5% to +5.4%, and surpasses traditional classifiers—such as support vector machine (SVM), decision tree, random forest, and K-nearest neighbors (KNN)—by up to +33.6% in accuracy and +40.2% in ${F}1$ -score, especially in target nodes lacking labeled data. These results highlight the effectiveness of the proposed approach in enabling robust, scalable, and privacy-preserving sensor fault classification in real-world IoT systems.