The propagation of IoT devices in healthcare, environmental monitoring, and smart cities has created the demand for robust anomaly detection mechanisms for security against failure and cyber-attacks. Most of these devices, however, must run within power-constrained environments, and energy efficiency takes the center stage. Conventional anomaly detection techniques, relying on power-intensive operations and periodic transmission of data, have the potential to drain IoT devices’ short power supplies in no time, rendering them inappropriate for continuous use. To overcome this, this paper suggests a new energy-aware anomaly detection framework tailored to IoT systems with short power supplies. Our approach uses lightweight machine learning with adaptive sampling schemes to facilitate continuous monitoring without wasting power. The model makes use of in-device feature selection and edge computing models to lessen data transmission requirements and processing requirements, with a hierarchical detection framework where initial screening is done at the device level and only high-priority anomalies are passed on to edge processors for in-depth analysis. Results show significant energy saving without compromising accuracy, proving the applicability of the model to real-world IoT environments with power limitations.