AI-Driven Performance Optimization of NB-IoT for Real-Time Industrial Monitoring in Industry 4.0: A Critical Review

Narrowband Internet of Things (NB-IoT) is an important cellular low-power wide-area network (LPWAN) technology for industrial sensing, offering wide-area coverage, licensed-spectrum operation, low device complexity, and long battery life. However, real-time industrial monitoring requires simultaneous consideration of latency, reliability, coverage, energy efficiency, security, and scalability. Although artificial intelligence (AI) has been proposed to adapt NB-IoT operation to changing channel and traffic conditions, the practical benefits and limitations of these approaches remain insufficiently consolidated. This paper presents a critical, PRISMA-inspired review of AI-driven NB-IoT performance optimization for industrial monitoring in Industry 4.0. Following the screening of literature from IEEE Xplore, SpringerLink, ScienceDirect, MDPI, ACM, Wiley, Scopus, and relevant 3GPP documents, 43 sources were retained and classified as core AI-driven NB-IoT studies, non-AI comparator studies, or supporting literature. The review maps AI paradigms to six performance objectives and examines their potential deployment across device, edge, and cloud layers. The synthesis reveals three key insights. First, direct evidence remains limited, with core studies relying predominantly on reinforcement-learning-related methods and simulation-based evaluation. Second, the available evidence does not establish a single AI paradigm as suitable for the entire monitoring pipeline; a layered combination of device-level TinyML, edge-level prediction and constrained control, and cloud-level training represents a promising direction, although end-to-end validation remains lacking. Third, AI overhead—including training, inference, model updates, memory, energy, and communication costs—must be evaluated alongside network performance because it may offset optimization gains. The review links six research gaps to seven future research directions and identifies reproducible testbeds, open datasets, overhead-aware multi-KPI benchmarking, and safe, explainable control as priorities for credible industrial deployment.