Edge Computing-Enabled Predictive Maintenance for Heterogeneous Manufacturing Lines

Thesis Abstract

The rapid evolution of manufacturing ecosystems toward highly interconnected, heterogeneous lines poses significant challenges for maintenance strategies due to variance in equipment, sensors, and control frameworks, which hinders timely fault detection and increases unplanned downtime. This study addresses the problem by developing an edge computing-enabled predictive maintenance framework that integrates heterogeneous machine data streams, local decision logic, and cloud-assisted analytics to reduce downtime and maintenance costs in multi-vendor production environments. The aim is to design, implement, and validate an ICT-driven solution that enables real-time anomaly detection, prognostic health assessment, and prescriptive maintenance scheduling across diverse equipment fleets. Specific objectives include (1) to characterize data sources, sensor modalities, and communication protocols across three representative production lines; (2) to develop edge-processed predictive models that fuse vibration, temperature, current, and qualitative operator inputs for remaining useful life estimation; (3) to implement lightweight, fault-tolerant inference engines at edge gateways with model adaptation for heterogeneity; (4) to evaluate predictive performance against centralized benchmarks using a mixed-methods, multi-site pilot; and (5) to assess operational impact on downtime, maintenance labor, and throughput through a cost-benefit analysis and stakeholder interviews. The methodology adopts a mixed-methods research design combining quantitative predictive analytics with qualitative validation. The population comprises three manufacturing lines within an automotive components facility characterized by diverse machine ages, brands, and control architectures. A purposive sample of 120 machines will be selected, with 40 machines per line representing varying duty cycles and failure histories. Data collection will utilize on-board sensors (vibration accelerometers, thermocouples, current sensors), edge gateway logs, maintenance records, and operator notes over a 12-month observation window, supplemented by 15 semi-structured interviews with maintenance engineers and line supervisors. Instrumentation will include a standardized data collection schema, calibrated vibration analysis tools, and a survey instrument for qualitative insights. Validation of instruments will involve pilot testing with 10 machines and reliability analysis (Cronbach’s alpha for survey items, test–retest for sensor-derived features). Analytical approaches will include (i) edge-enabled feature extraction and time-series modeling (autocorrelation, spectral analysis, and wavelet features) to feed predictive models; (ii) machine learning techniques such as gradient boosting (XGBoost) and recurrent neural networks (LSTM) for remaining useful life and fault probability estimation, with transfer learning to accommodate heterogeneous equipment; (iii) model validation through cross-validation and rolling-origin evaluation, with performance metrics including RMSE, MAE, AUC-ROC, and prognostic calibration. A comparative analysis will be conducted against a centralized cloud-based baseline to quantify benefits of edge processing in latency, bandwidth usage, and resilience. Theoretical framing will integrate Two-Factor Theory of technology adoption and the Technology-Organization-Environment (TOE) framework to interpret adoption and operationalization of edge computing in maintenance. Data fusion will leverage feature-level fusion techniques to reconcile heterogeneous data streams, with explainability analyses using SHAP values to interpret model decisions. Expected findings include (a) edge-processed models achieving comparable or superior predictive accuracy to centralized approaches while reducing data transmission by up to 65%; (b) faster maintenance decision cycles due to sub-second inference times at the edge; (c) improved prognostic reliability across machinery brands through transfer learning and recalibration; and (d) demonstrable reductions in unplanned downtime, maintenance labor hours, and production losses, with robust cost savings in the range of 12–18% per line per annum. The study will contribute to knowledge by (i) advancing practical methodologies for deploying predictive maintenance in heterogeneous manufacturing contexts using edge computing and multi-source data fusion; (ii) offering a validated framework for edge-to-cloud collaborative analytics and model governance across supplier ecosystems; and (iii) extending empirical evidence on the socio-technical implications of ICT-driven maintenance in real-world plants. Conclusions are anticipated to confirm that edge computing-enabled predictive maintenance enhances reliability and efficiency for heterogeneous lines without compromising safety or data governance. Recommendations will focus on scalable deployment strategies, governance of cross-vendor data standards, dynamic model update protocols, and policy guidance for industrial IoT architectures, including guidelines for operator training, cybersecurity, and continuous improvement cycles.