Low-power Internet of Things systems depend on accurate sensor fault detection to ensure reliable operation in critical applications like industrial monitoring and aerial systems. However, existing feature extraction methods struggle to efficiently capture complex time-series patterns under resource constraints, hindering detection reliability. This study aims to enhance sensor fault detection in low-power Internet of Things systems under resource constraints. Our approach derives multi-scale, interpretable features from raw sensor data, capturing temporal dynamics and local statistics efficiently. Experimental results demonstrate that our enhanced features outperform benchmark techniques, improving fault identification and providing robust signal representations. Our work provides a practical, efficient solution, enhancing data integrity and resilience across diverse Internet of Things deployments.