Energy-Aware Edge Intelligence for Real-Time Industrial Automation

Thesis Abstract

Industrial facilities increasingly rely on real-time automation systems powered by edge computing to meet stringent latency and reliability requirements; however, the energy cost and environmental impact of edge-enabled industrial workloads remain incompletely understood, particularly when balancing local processing, communication, and maintenance overheads. This study investigates energy-aware edge intelligence for real-time industrial automation by examining how adaptive offloading, model compression, and workload scheduling influence both performance metrics (latency, throughput, reliability) and energy consumption in heterogeneous edge-cloud architectures. The aim is to design and validate an integrated framework that minimizes energy while preserving or enhancing real-time control quality in industrial environments. Specific objectives include (1) to quantify the energy-performance trade-offs of edge inference for predictive maintenance, anomaly detection, and closed-loop control in a representative manufacturing line; (2) to develop an adaptive offloading strategy that selects computation location (edge node, gateway, or cloud) based on context, workload characteristics, and network conditions; (3) to evaluate model compression and quantization techniques for resource-constrained edge devices without compromising control accuracy; (4) to formulate a dynamic resource provisioning and scheduling algorithm that co-optimizes compute, network, and storage energy consumption; (5) to validate the framework in a real-world pilot with a mid-sized assembly line, involving diverse heterogeneous devices and industrial protocols. The methodology adopts a mixed-methods research design grounded in the Resource-Based View and Dynamic Capabilities Theory to explain how energy-efficient edge intelligence can sustain competitive advantage in manufacturing. The population comprises industrial facilities deploying edge-empowered automation across North American and European sites, with a purposive sample of three pilot environments representing varying scales, network topologies, and control schemas. A total of 12 edge devices, 4 gateway controllers, and 2 local data centers will be instrumented to collect empirical data over a six-month deployment. Data collection instruments include high-frequency telemetry from edge nodes (CPU/GPU utilization, memory, power draw, and thermal metrics), network measurements (latency, jitter, packet loss, bandwidth), and process-level indicators (control loop error, energy per operation, and uptime). Complementary qualitative data will be gathered through semi-structured interviews with system engineers and operators to capture contextual factors affecting energy decisions. Analytical methods comprise (i) regression analysis and multivariate time-series forecasting to model energy consumption and latency across configurations; (ii) ANOVA and post-hoc tests to identify statistically significant differences among offloading strategies and model compression settings; (iii) mixed-effects models to handle hierarchical data from multiple devices and sites; (iv) optimization-based simulations to derive the co-optimized scheduling policy, incorporating stochastic network delays and failure modes; (v) thematic analysis of interview transcripts to extract governance and organizational factors influencing adoption; (vi) validation against a formal model of the control loop employing Lyapunov-based stability checks to ensure real-time performance is not degraded. Expected findings indicate that context-aware offloading, coupled with targeted model compression and intelligent scheduling, can reduce total energy consumption of edge-enabled automation by 25–40% without compromising latency or control quality. It is anticipated that network-aware decisions will be most beneficial under intermittent connectivity, whereas on-premise edge processing with lightweight models will dominate when ultra-low latency is required. The study also expects to reveal trade-offs between local inference accuracy and communication energy, highlighting scenarios where modest reductions in model precision yield substantial energy savings with negligible control impact. The contribution to knowledge includes a transferable framework for energy-aware edge intelligence in real-time industrial automation, an adaptive offloading algorithm with provable performance bounds, and empirical evidence linking organizational factors to technology uptake in energy-efficient automation. The main conclusion is that an integrated strategy combining adaptive offloading, model compression, and dynamic scheduling can achieve substantial energy savings while meeting stringent industrial performance requirements. Recommendations include adopting standardized energy metrics across industrial edge ecosystems, deploying pilot programs to calibrate offloading thresholds for specific control loops, and incorporating energy-aware policies in industrial IT governance to sustain long-term efficiency gains.