Edge-Computing-Driven Real-Time Sensor Fusion for Smart Factories

Thesis Abstract

In modern manufacturing, real-time decision-making relies on timely and accurate sensor data streams, yet edge-centrism remains underexplored for robust fusion under heterogeneous and noisy environments. This study investigates how edge-computing-enabled sensor fusion can enhance operational visibility, predictive maintenance, and adaptive process control in smart factories. The aim is to design, implement, and evaluate an edge-centric sensor fusion framework that fuses multimodal data streams—vibration, temperature, pressure, and vision—at the network edge to reduce latency, improve anomaly detection accuracy, and sustain performance under bandwidth-constrained conditions. Specific objectives include (1) developing a modular edge computing architecture that supports scalable data preprocessing, synchronization, and fusion algorithms; (2) implementing lightweight fusion models combining Kalman filtering, multi-sensor Bayesian networks, and lightweight deep learning (e.g., MobileNet-based anomaly detectors) suitable for edge deployment; (3) evaluating the framework in a pilot smart factory line with 1200 daily cycles across three production stations; (4) assessing end-to-end latency, fusion accuracy, and fault tolerance under varying network loads; and (5) formulating guidelines for deployment, governance, and security in edge-enabled manufacturing environments. The methodology adopts a pragmatist research design combining design science with empirical validation. The population comprises industrial sensors and edge gateways deployed in a mid-size automotive components plant. A purposive sample of three production lines, each instrumented with 8–12 heterogeneous sensors and two edge servers, yields approximately 3,600 hours of operational data collected over six months. Data collection instruments include time-synchronized sensor suites, edge-embedded fusion modules, and a central cloud-based repository for archival reference. Quantitative data analysis employs time-series analysis, Kalman filter-based optimal fusion, and Bayesian networks to quantify fusion accuracy improvements; regression analysis and ANOVA test the impact of edge deployment on latency, throughput, and anomaly detection precision. Qualitative insights derive from expert interviews with line engineers and maintenance personnel, analyzed via thematic analysis to capture practical challenges and governance considerations. Validity and reliability are addressed through cross-validation of fusion outputs against ground-truth events, calibration experiments, and repeated measures across lines. Ethical considerations include data privacy, safety compliance, and secure edge-to-cloud communications. The expected findings indicate that edge-enabled sensor fusion reduces end-to-end latency by 38–45% compared with cloud-only fusion, while increasing anomaly-detection precision by 12–18 percentage points and reducing false positives by 9–14 percentage points under congested network scenarios. It is anticipated that the fusion framework will maintain robust performance with up to 40% packet loss on peripheral links, owing to local preprocessing and fault-tolerant fusion strategies. The study also expects that integrating a Bayesian network with a lightweight CNN-based detector will outperform standalone methods in both predictive maintenance accuracy and real-time fault diagnosis, particularly for emergent faults with limited labeled data. The findings will be triangulated with expert qualitative insights to reveal actionable design considerations and security implications. The study contributes to knowledge by delivering a validated, modular edge-centric sensor fusion architecture tailored for smart factories, accompanied by empirical evidence of performance gains in latency, accuracy, and resilience under real-world conditions. It advances the theoretical understanding of distributed fusion under edge constraints by integrating concepts from Kalman filtering, Bayesian networks, and edge-AI approximations, and it extends practical guidance for deployment, governance, and cyber-physical security in edge-enabled manufacturing ecosystems. The principal conclusion is that edge-computing-driven sensor fusion offers a scalable, low-latency, high-accuracy solution for real-time factory intelligence, with clear design patterns for deployment, data governance, and continuous improvement. Recommendations include adopting standardized data schemas and time synchronization protocols, implementing runtime adaptivity in fusion models, and pursuing ongoing assessment of security and privacy risks in edge-to-cloud architectures.