Аннотация
Dense industrial sensor networks increasingly generate high-frequency telemetry that cloud-centric architectures cannot absorb without prohibitive latency and bandwidth costs. This paper presents an edge-deployed Digital Twin service architecture for real-time orchestration of heterogeneous industrial sensor streams, validated on two public benchmarks and one operational brownfield deployment. A lightweight dual-rule filter, implemented on a Raspberry Pi 5 edge gateway, evaluates each incoming record against a statistical deviation criterion and a digital state-change criterion; only flagged records are forwarded upstream while steady-state telemetry is aggregated locally. Applied to the MetroPT-3 compressor benchmark (15,169,480 records, 15 sensors at 1 Hz over 180 days), the filter reduces upstream volume by 87.2% while retaining all four documented air-leak failures. Validation on the AI4I 2020 predictive maintenance benchmark (10,000 records, four failure modes, 3.39% failure rate) confirms 93.7% reduction with 100% fault retention across all failure types. An operational case study on a brownfield smart manufacturing facility, instrumented with 17 ESP32-based IIoT nodes generating approximately 6.32 million raw records per monitoring period, demonstrates 99.97% system availability, end-to- end latency below 2 seconds at the 95th percentile, and sustained write throughput of 1,220 metrics per second. Edge-informed oven state control at the same facility further reduced equipment energy consumption by 23.06% and shortened anomaly reaction time by 38.8%. These results confirm that a service-oriented edge Digital Twin layer substantially lowers communication overhead, preserves real-time fault responsiveness, and scales to heterogeneous industrial sensing workloads without cloud dependency for time-critical decisions.