ENERGY-EFFICIENT TIME SERIES CLASSIFICATION ON IOT DEVICES WITH SENSOR-AWARE EARLY-EXIT MACHINE LEARNING

Lubah Nelson

Time-series data processing is critical in several high-impact applications including mobilehealth, environmental monitoring, and digital agriculture. Energy efficiency represents a fundamental bottleneck for continuous time-series classification on edge IoT devices, partic- ularly in wearable health monitoring where battery constraints severely limit deployment. Traditional machine learning models for time-series processing wait for complete sensor data windows before classification, resulting in excessive energy consumption regardless of task difficulty. This thesis introduces Sensor-Aware Early-Exit (SEE) classifiers that jointly optimize sensing and computational energy through progressive inference with partial data windows. The proposed framework includes multiple architectural innovations: (1) SEE-CNN, a con- volutional neural network with late input blocks that preserve temporal information while enabling early exits; (2) SEE-vRF, a vertically-split random forest architecture; and (3) SEE- hRF, a sequential random forest with cascaded sub-forests. These architectures dynamically terminate inference and transition sensors to low-power mode once prediction confidence exceeds learned thresholds. A comprehensive evaluation methodology was developed encompassing six diverse health and activity recognition datasets (PAMAP2, WESAD, UCI-HAR, OPPORTUNITY, EMG, and Epilepsy), systematic ablation studies, and comparative analysis against state-of-the-art baselines including NMEC and computational-only early-exit methods. Threshold optimiza- tion was performed through Design Space Exploration combined with Bayesian Optimiza- tion, balancing accuracy-energy trade-offs across the solution space. Experimental results demonstrate that SEE classifiers achieve 50-70% sensing energy savings while maintaining accuracy within 2% of standard models that use complete data windows. The sequential RF architecture (SEE-hRF) achieves up to 70% energy reduction on health monitoring datasets. End-to-end validation on Raspberry Pi hardware with real sensors confirms 29-42% system-level energy savings, validating the practical feasibility of the approach. Calibration analysis demonstrates that entropy-based confidence metrics provide reliable exit decisions with Expected Calibration Error below 0.05. This work establishes that joint sensor-compute optimization through early-exit architec- tures fundamentally outperforms approaches targeting only computational or only sensing efficiency. The proposed methods enable longer battery lifetime for wearable devices, poten- tially improving user compliance and expanding the viability of continuous health monitoring applications.

ENERGY-EFFICIENT TIME SERIES CLASSIFICATION ON IOT DEVICES WITH SENSOR-AWARE EARLY-EXIT MACHINE LEARNING

Files and links (1)

Abstract

Metrics

Details