Adaptive Inference Telemetry: A Cascaded Edge-to-Cloud Dynamic Routing Architecture for Constrained Industrial IoT Telemetry

Modern Industrial Internet of Things (IIoT) edge deployments face a fundamental operational trade-off: localized execution on micro-edge architectures is restricted by hardware compute and memory bounds, whereas continuous cloud telemetry offloading incurs prohibitive API token expenditures, network transmission latency, and operational fragility under network jitter. This paper introduces an adaptive inference telemetry framework utilizing a two-tier cascaded decision pipeline. Telemetry signals and raw sensor packets are first evaluated on-device using an INT8-quantized lightweight neural network running in an ONNX runtime environment. Observations yielding predictive confidence below an empirically tuned entropy threshold are compressed and conditionally dispatched to upstream cloud transformer endpoints for deep verification. By evaluating the system against real-world degradation telemetry benchmarks, we demonstrate that dynamic confidence thresholding isolates anomalous edge cases, diminishes cloud compute expenditures by over 70%, and maintains sub-25ms median inference latency on local edge cores. The architecture demonstrates a viable engineering standard for balancing token economics and inference reliability across distributed sensing infrastructure. Full open-source pipeline, evaluation scripts, and reproducible models are documented.