Lightweight deep learning models for real-time IoT data analysis in resource-constrained environments

The rapid growth of Internet of Things (IoT) deployments has intensified the need for real-time data analytics using lightweight machine learning models that can operate under strict energy, latency, and reliability constraints. While existing studies largely focus on model compression and inference optimization, they often overlook the impact of unreliable wireless communication on learning performance. This paper presents a comprehensive cross-layer framework that integrates lightweight ensemble learning with PHY-MAC layer-aware IoT network simulation to evaluate the end-to-end performance of real-time data analytics in resource-constrained environments. We investigate three data handling strategies-Baseline transmission, Duty-Cycle-aware reporting, and a proposed Rolling Window aggregation method-combined with multiple MAC protocols (TDMA, CSMA, and DutyCycleMAC) and PHY channel models (Ideal, AWGN, and Rayleigh). Extensive MATLAB-based simulations quantify key Quality-of-Service (QoS) metrics, including packet delivery ratio (PDR), latency, jitter, throughput, and energy consumption, alongside machine learning metrics such as accuracy, F1-score, and inference time. Results show that TDMA consistently achieves the highest PDR (up to 1.0 under Ideal PHY), while the proposed Rolling Window variant improves robustness and energy efficiency without sacrificing predictive accuracy. The findings demonstrate that communication-aware data aggregation is critical for deploying reliable and energy-efficient lightweight AI in real-world IoT systems.