Optimizing Edge AI for Smart Factories: A Siemens Case Study

Thesis Abstract

Edge AI technologies promise to elevate operational efficiency, predictive maintenance, and energy optimization in modern manufacturing by enabling real-time decision-making at the device layer. This study addresses the gap between theoretical edge AI benefits and practical deployment within Siemens’ smart factories, where heterogeneous industrial environments, strict latency requirements, and rigorous safety standards challenge scalable adoption. The aim is to optimize edge inference pipelines and orchestration across multi-site manufacturing lines to reduce downtime, improve defect detection, and lower total cost of ownership. Specific objectives include (1) characterizing the current edge AI architecture and data flows within Siemens’ digital factories; (2) developing a unified framework for end-to-end edge inference optimization, including model selection, quantization, pruning, and on-device serving; (3) evaluating the impact of edge deployment strategies on latency, throughput, and energy consumption; (4) assessing governance, security, and reliability implications in industrial environments; and (5) producing actionable recommendations for scalable deployment aligned with Siemens’ product lifecycle and cybersecurity standards. The methodology adopts a mixed-methods design anchored in both empirical measurement and theory-driven analysis. The population comprises three Siemens production facilities, with a purposive sample of 12 production lines and 36 edge devices (including industrial cameras, PLCs, and gate controllers) deployed for defect detection and predictive maintenance tasks. Data collection combines (i) quantitative telemetry from edge devices over a six-month period, including CPU/GPU utilization, inference latency, bandwidth usage, and energy consumption, and (ii) qualitative interviews with 18 engineers and operations managers to capture governance, safety, and organizational readiness factors. Instruments include a standardized telemetry logger, a structured survey instrument validated for reliability (Cronbach’s alpha > 0.80), and semi-structured interview guides. Validity and reliability are ensured through triangulation, pilot testing, and inter-rater reliability checks for qualitative coding. Analytical approaches integrate both statistical and computational methods. Descriptive statistics summarize baseline performance; inferential analyses employ multiple regression and ANOVA to quantify the effects of edge optimization interventions (quantization levels, model pruning ratios, and edge-cloud offloading policies) on latency, throughput, and energy metrics, controlling for line complexity and workload variance. A time-series analysis (ARIMA/Prophet) assesses performance trends post-implementation. On the qualitative side, thematic analysis of interview transcripts identifies organizational enablers and barriers, mapped to established theories of technology acceptance and sociotechnical systems. A conceptual model drawing on the Technology Acceptance Model (TAM) and the Socio-Technical Systems theory frames the integration of technical and organizational factors, with structural equation modeling (SEM) used to test hypothesized relationships between perceived usefulness, ease of use, governance quality, and deployment success. Key expected findings include (i) a scalable edge optimization framework with recommended quantization/pruning configurations that meet latency targets (<20 ms per inference) and energy budgets suitable for edge-only processing in high-mpeed production lines; (ii) evidence that selective offloading to a centralized inference server yields diminishing returns beyond a 60/40 edge/cloud split for specific defect-detection tasks; (iii) demonstrable improvements in mean time between failures (MTBF) and defect detection accuracy when edge-optimized models are deployed, with statistically significant gains (p < 0.05) compared to baseline models; (iv) insights into governance and security practices that correlate with sustainable deployment and risk reduction. The study contributes to knowledge by bridging industrial practice and edge AI theory, offering a replicable framework for optimizing edge inference in complex manufacturing ecosystems, and providing empirical evidence on how architectural choices interact with organizational factors to determine deployment outcomes. The main conclusion anticipates that a holistic, theory-informed approach to edge AI—integrating model-level optimizations, edge-centric orchestration, and robust governance—substantially enhances operational performance in Siemens’ smart factories. Recommendations include adopting the proposed optimization blueprint across similar high-volume manufacturing settings, investing in telemetry-enabled feedback loops for continual model refinement, and strengthening cybersecurity and safety protocols to sustain scalable, reliable edge deployments.