Design of AI-Driven Predictive Maintenance for Additive Manufacturing Systems

Thesis Abstract

The rapid evolution of additive manufacturing (AM) technologies has intensified the demand for reliable, down-time minimizing maintenance strategies to sustain production quality and throughput. Traditional maintenance approaches are ill-suited for AM systems due to complex, multi-physics processes, high equipment investment, and variable operating conditions. This study proposes an AI-driven predictive maintenance (PdM) framework tailored to AM systems, aimed at reducing unexpected failures, extending machine life, and optimizing maintenance scheduling under real-world manufacturing constraints. The objectives are to (i) develop a multidimensional data architecture integrating machine telemetry, process parameters, sensor fusion, and material quality outcomes; (ii) implement machine learning models for early fault detection and remaining useful life (RUL) estimation of critical components such as laser sources, cooling systems, and drive electronics; (iii) validate the framework using a longitudinal dataset from a production AM facility and a controlled lab testbed; and (iv) assess economic and operational impacts through a cost-benefit analysis and lifecycle performance metrics. The methodology employs a mixed-methods research design grounded in the theory of maintenance optimization and reliability-centered maintenance (RCM), drawing on1321 Benford’s anomaly detection approach for sensor data, and a systems engineering perspective on digital twins. The population comprises industrial AM systems operating in a mid-size manufacturing plant and a university lab-scale additive manufacturing platform. A stratified random sample of 40 fused deposition modeling (FDM) and selective laser sintering (SLS) machines from the plant, together with 6 high-precision laser SLM (selective laser melting) units, will be monitored over 12 months. Data collection instruments include high-frequency telemetry sensors (temperature, vibration, power, and acoustics), process monitoring logs (print speed, layer thickness, hatch distance), maintenance records, non-destructive evaluation results, and material property outcomes. Instrumentation will be complemented by semi-structured interviews with maintenance engineers to capture tacit knowledge. Model development will proceed in three stages (1) anomaly detection using unsupervised learning (Isolation Forest, Autoencoders) to flag deviations from nominal behavior; (2) supervised learning for RUL prediction using time-series models (LSTM, Temporal Convolutional Networks) and survival analysis (Cox proportional hazards) to integrate covariates; and (3) a fusion layer employing Bayesian updating to continuously refine predictions as new data arrive. Data analysis will entail preprocessing, feature extraction from multivariate time-series, and cross-validation with a hold-out test set. Model performance will be evaluated using metrics such as area under the ROC curve for fault detection, root mean square error (RMSE) for RUL estimates, and calibration plots for probabilistic outputs. A comparative assessment will be conducted against baseline maintenance schedules and standard condition-based maintenance (CbM). The study will also perform a cost-benefit analysis incorporating maintenance costs, downtime losses, and potential yield improvements, alongside a sensitivity analysis to examine robustness under varying data quality and sensor failure scenarios. The expected findings include (i) a high-precision PdM model achieving >85% detection accuracy and RUL MAE under 15% of component useful life, (ii) demonstrable reductions in unplanned downtime by 25–40% and maintenance labor costs by 15–25%, and (iii) a validated digital twin prototype enabling real-time prognostics for AM processes. Contributions to knowledge are anticipated in (a) the integration of AI-driven PdM within heterogeneous AM environments, (b) a scalable, data-centric framework for predicting failures in optical, thermal, and mechanical subsystems of AM equipment, and (c) empirical evidence linking sensor-based degradation patterns to process quality outcomes, thereby informing RCM methodologies for advanced manufacturing. The study concludes that combining anomaly detection, predictive modeling, and Bayesian data fusion within a digital twin context yields substantial reliability and economic benefits for AM systems. Recommendations include extending the framework to other AM modalities (binder jetting, multi-material printing), developing industry-standard data schemas for cross-facility collaboration, and integrating operator training with the PdM platform to enhance adoption and operational resilience.