Edge-Cloud Continuum: Split Neural Inference and Adaptive Quantization for Industrial IoT

Modern Industry 4.0 smart manufacturing plants deploy high-resolution line-scan cameras, acoustic emission arrays, and thermal imaging sensors that generate gigabits per second of raw inspection telemetry per assembly line. Executing large Vision Transformers (ViTs) and multi-scale object detection backbones entirely on resource-constrained industrial edge programmable logic controllers (PLCs) or embedded System-on-Modules (SoMs) exceeds thermal design power envelopes and violates strict 20-millisecond conveyor actuation deadlines. Conversely, streaming uncompressed 4K 120-fps camera frames over factory-floor private 5G or Wi-Fi 6E links to centralized cloud GPU clusters induces severe uplink congestion, packet jitter, and production line halts during wireless fading or WAN outages. This paper proposes an Adaptive Split Neural Inference and Entropy-Guided Mixed-Precision Quantization architecture for the industrial edge-cloud continuum. Our framework introduces a learned Bottleneck Injection Module (BIM) and an early-exit confidence classifier at intermediate network layers on NVIDIA Jetson Orin industrial edge nodes, teamed with a Lagrangian Dynamic Split Scheduler (LDSS) that jointly selects the optimal layer partition index l^* and activation quantization bit-width b ∈ \2, 4, 8\ based on real-time wireless link capacity and cloud queue telemetry. Evaluated through a production-grade case study modeled on BMW Group iFactory Automated Optical Inspection (AOI) and Siemens Industrial Edge orchestrated via AWS IoT Greengrass across 64 automotive welding and paint-shop cells, empirical benchmarks demonstrate that the proposed continuum architecture resolves 76.4% of defect inspections locally via early exit in 5.8 ms, reduces 5G uplink bandwidth consumption by 88.2%, lowers end-to-end p99 inspection latency from 64.5 ms to 14.2 ms, and preserves 99.4% mean Average Precision (mAP@50) compared to full FP16 cloud execution. Research Domain & Specialization Track Master of Science in Information Technology (MSIT) — Edge Computing, Embedded AI, and Industrial Cloud Continuum. Core Technical Highlights Empirical Production Case Studies: Rigorous architectural blueprints and telemetry evaluation across enterprise cloud deployments. Four-Tier Reference Architecture: Formal specification from edge ingress to distributed hardware execution pools. Mathematical Formulation: Formally derived algorithmic and optimization models with closed-loop verification. Multi-Modal Benchmarking: Comparative empirical baselines under fault-injection and synthetic workload scaling. Author Information Author: Sonu Kumar SinghRole: Senior Consultant Affiliation: Independent Researcher Publication Status: Preprint / Working Paper (MSIT AI & Cloud Systems Research Series) Reproducibility & Open Science All included architecture diagrams, telemetry schemas, and benchmark curves are open-access under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.