<![CDATA[Wireless Sensor Networks (WSNs) are widely used in various monitoring applications, including environmental observation, smart infrastructure, and Internet of Things (IoT) systems. Despite their widespread adoption, WSNs are highly susceptible to data errors caused by sensor degradation, hardware malfunctions, environmental disturbances, and communication issues. These faults can significantly reduce data reliability and lead to incorrect system decisions if not properly handled. This study proposes a multi-class data-fault detection approach for WSNs under imbalanced data conditions by integrating Random Undersampling (RUS) with the Extra-Trees classification algorithm. The proposed framework aims to address the class imbalance problem commonly found in sensor fault datasets while improving fault detection performance across multiple fault types. Experiments were conducted using a WSN dataset containing temperature and humidity measurements, in which three fault types: Bias, Drift, and Spike were analyzed alongside normal sensor data. The experimental results demonstrate that Random Undersampling leads to a substantial improvement in classification performance. Without RUS, the Extra-Trees classifier achieved an accuracy of 48% and failed to detect spike faults. After applying RUS, classification accuracy increased to 91%, accompanied by balanced precision, recall, and F1-score values across all classes. These findings indicate that the combination of Random Undersampling and Extra-Trees provides an effective and reliable solution for multi-class data fault detection in WSN environments.]]>
