Edge Machine Learning for Intelligent Internet of Things Applications

Abstract

The Internet of Things (IoT) has evolved from a conceptual paradigm into a pervasive infrastructure connecting billions of sensors, actuators, and embedded devices across smart homes, industrial automation, healthcare monitoring, autonomous transportation, precision agriculture, and urban infrastructure. The intelligence of IoT systems increasingly depends on the ability to perform machine learning (ML) inference and, in some cases, training directly at the network edge, close to the data source, rather than relying exclusively on centralized cloud processing. Edge machine learning addresses the fundamental constraints of IoT environments, including limited network bandwidth, stringent latency requirements, intermittent connectivity, data privacy concerns, and the need for autonomous real-time decision-making, by deploying optimized ML models on resource-constrained edge devices and gateways. This paper presents a comprehensive survey of recent advances in edge machine learning for intelligent IoT applications. We examine the enabling techniques for deploying ML on edge hardware, including model compression, neural architecture search for efficient designs, on-device training, and specialized hardware accelerators. The paper reviews edge ML frameworks and deployment toolchains that bridge the gap between model development and embedded execution. We provide detailed analyses of edge ML applications across smart home and building automation, industrial IoT and predictive maintenance, wearable health monitoring, autonomous vehicles and robotics, precision agriculture, and smart city infrastructure. The distributed intelligence paradigm is explored through federated learning, split computing, and hierarchical edge-cloud architectures that partition ML workloads across the IoT-edge-cloud continuum. Critical challenges including energy efficiency, memory constraints, model robustness, security of edge ML models, over-the-air model updates, and heterogeneous device ecosystems are systematically discussed. Finally, we outline future research directions including tiny foundation models, neuromorphic edge processors, on-device continual learning, and the convergence of edge ML with 5G/6G communication networks.