Spatial-Narrative-Aware Cooperative Edge Caching and Adaptive Layered Video Delivery for Industry 5.0

Abstract

INTRODUCTION: Industry 5.0 brings worker context, system resilience, and resource efficiency into the design of industrial digital services. Training video is a useful edge-delivery case because demand varies with where work is performed and with the type of instruction being requested, whereas conventional cache policies usually reduce demand to a single global popularity profile.

OBJECTIVES: We propose a spatial-narrative-aware delivery framework that combines zone-category-conditioned demand, layer-level cooperative caching, and adaptive Scalable Video Coding (SVC) transmission. Here, “spatial-narrative-aware” refers only to factory-zone and instructional-category metadata; it is not a claim of temporal shot segmentation or automatic narrative parsing.

METHODS: Evaluation uses 20 independent runs with fixed random seeds. Each run contains 1,500 requests, including a 300-request warm-up, across 20 edge nodes and 1,000 three-layer SVC items. All methods receive the same request sequence and per-request Rayleigh-fading realization. Experiments cover component ablation, a 10–100 MB cache-capacity sweep, θ2 sensitivity, controlled node failures, worker zone transitions, and an adaptation of a 2024 cooperative-SVC method under matched conditions. A separate public-data validation uses JSVM 9.19.15 on a 10-s construction-work clip and evaluates PPE detection on the 141-image Construction-PPE test split.

RESULTS: With 10 MB per node, Full ALT-PACC attains a 72.14% cache hit rate (95% CI: 71.60–72.67%), compared with 65.74% for ALT+LRU and 63.28% for CCo-Fog-adapted. Modeled energy is 80.62 mJ per served segment, 7.7% lower than ALT+LRU, and mean delay is 955.0 ms versus 1009.3 ms. In the auxiliary JSVM experiment, BL, BL+E1, and Full require 141.82, 319.16, and 606.34 kb/s and yield 29.57, 32.21, and 35.03 dB Y-PSNR. PPE mAP@0.5 is 0.391 on the original test images and 0.390 after base-layer coding; the paired-bootstrap retention interval includes 100%.

CONCLUSION: Layer-aware cooperation accounts for most of the cache-efficiency gain, while zone-category context and replica-diversity terms add a smaller increment under tight storage. The public-data checks show a measured SVC rate-quality progression and no detectable mAP@0.5 loss for PPE detection at the base layer in this external dataset. These results support network-side caching and safety-object semantic retention, not worker learning or task completion.