Smart Factory Digital Twin for Predictive Maintenance and Optimization

Thesis Abstract

The rapid digitization of manufacturing environments and the expansion of interconnected assets have intensified the need for integrated intelligence to ensure operational reliability, minimize downtime, and optimize production throughput. This study addresses the gap between traditional predictive maintenance approaches and the dynamic, real-time decision support required by modern smart factories. The aim is to develop and validate a Digital Twin-based framework that enables predictive maintenance and production optimization through continuous24/7 monitoring, simulation, and prescriptive guidance. Specific objectives include (i) designing a modular digital twin architecture that integrates IoT sensor streams, historian data, and enterprise resource planning (ERP) data, (ii) developing machine learning models for remaining useful life (RUL) estimation and anomaly detection for critical equipment, (iii) implementing a real-time optimization engine that balances maintenance schedules with production targets, and (iv) evaluating the framework’s impact on downtime reduction, maintenance cost, and overall equipment effectiveness (OEE) in a defined manufacturing cell. The study adopts a mixed-methods, embedded research design anchored in the Resource-Based View (RBV) and the Technology-Organization-Environment (TOE) framework to capture both quantitative performance metrics and qualitative insights from practitioners. The population consists of a medium-sized automotive components assembly facility comprising 180 machines across three production lines, with a purposive sample of 50 critical assets selected by failure history and criticality. Data sources include five years of historical sensor data (vibration, temperature, force, pressure), maintenance logs, production throughput records, and ERP data. Instrumentation combines calibrated IoT sensors, standardized maintenance request forms, and structured interview guides for maintenance engineers and line supervisors. The validation plan employs cross-validation for predictive models, Lloyd’s reliability analysis for sensor data, and content validity checks with domain experts. The data analysis sequence includes (i) data preprocessing and feature extraction using time-series engineering and wavelet transforms, (ii) supervised learning for RUL prediction employing gradient boosting and deep recurrent networks, (iii) unsupervised clustering for anomaly pattern discovery, (iv) a Bayesian network-based causal model linking maintenance actions to downtime outcomes, (v) a mixed-methods integration using convergent parallel design, and (vi) a scenario-based evaluation of the real-time optimization engine using discrete-event simulation and linear programming. The conceptual framework integrates Turban’s IT-enabled organizational performance theory with the prognostic/prescriptive analytics paradigm, underpinned by the Theory of Constraints to ensure feasible production scheduling within maintenance windows. Expected findings include (a) improved accuracy of RUL estimates for high-impact assets (target RMSE under 8%), (b) reliable detection of incipient faults with a false-positive rate below 5% through multi-sensor fusion, (c) measurable reductions in unscheduled downtime and maintenance costs (projected decreases of 25–35%), and (d) enhanced OEE due to synchronized maintenance planning and production scheduling, validated through pre/post implementation comparisons and simulation-based what-if analyses. The study anticipates that the digital twin will demonstrate robustness to data gaps and noise, with imputed estimates guided by Bayesian priors and physics-informed constraints. Contribution to knowledge includes (i) a scalable, field-tested digital twin architecture tailored for mid-sized manufacturing environments, (ii) an integrated modeling pipeline combining ML-based RUL prediction, anomaly detection, and prescriptive optimization, (iii) empirical evidence on the socio-technical factors influencing adoption, and (iv) a demonstrable link between digital twin maturation and measurable improvements in maintenance efficiency and production performance. The research extends existing literature by operationalizing a holistic twin that couples asset health with manufacturing scheduling, rather than treating maintenance and production optimization in isolation, and by providing a replicable methodology for empirical validation in real-world settings. The main conclusion is that a modular smart factory digital twin, when embedded with real-time data streams and an optimization layer, can significantly enhance predictive maintenance effectiveness and production resilience. Recommendations include prioritizing data governance and sensor calibration protocols, adopting a phased deployment with pilot lines before full-scale rollout, investing in cross-functional teams to manage the digital twin lifecycle, and extending the framework to include supply chain and quality management domains to achieve enterprise-wide optimization.