Current Concepts in Engineering Research and Technology Vol. 4

While sensor data are central to any IoT monitoring system, conventional approaches that use fixed thresholds to identify faults often fail to detect subtle faults. Readings that freeze, drift slowly, or depart from normal behavioural patterns may remain unnoticed even when they stay within valid operating ranges. In this chapter, a behaviour-based sensor fault-detection framework is proposed, primarily for deployment on resource-constrained edge devices in IoT environments, by combining temporal and spatial feature extraction with a lightweight Random Forest classifier. The system continuously gathers synchronised multi-sensor time-series data and then extracts features based on the delta (difference) between current and previous readings to reflect the temporal characteristics of the data before passing them to a supervised classification system trained to recognise normal behaviour or fault patterns. The proposed method classifies data with a very low computational burden using a dataset of approximately 12,500 sensor readings and an 80/20 train-test split, achieving near-perfect classification accuracy at the edge. Experimental results verify the effectiveness of the system in detecting subtle sensor faults that simple threshold-based approaches fail to identify.