Adaptive Predictive Maintenance for Additive Manufactured Parts via Data-Driven Digital Twin

Thesis Abstract

This study addresses the reliability and maintenance challenges of additively manufactured (AM) components by developing an adaptive predictive maintenance framework implemented through a data-driven digital twin. The problem stems from process-induced defects, anisotropic material properties, and residual stresses in AM parts, which compromise structural integrity and lead to unexpected downtime. The aim is to design a real-time, self-learning maintenance system that predicts failure modes and schedules interventions to minimize downtime while preserving part performance. Specific objectives include (i) constructing a digital twin that fuses in-situ process data, post-build non-destructive evaluation (NDE) signals, and operational load histories; (ii) developing machine learning models that map AM process parameters and fatigue indicators to deterioration trajectories; (iii) integrating physics-based damage models with data-driven predictors to enable adaptive maintenance decisions; (iv) validating the framework on a representative AM parts cohort under variable service conditions; and (v) evaluating cost-benefit implications for maintenance planning in a production environment. The methodology combines a mixed-methods research design with empirical testing in both lab and industrial contexts. The population comprises metal and polymer AM parts (n=120) produced via laser powder bed fusion and directed energy deposition, sampled across three geometries and two material systems. A stratified sampling approach yields a dataset of 120 units for initial model development and an independent test set of 40 parts for out-of-sample validation. Data collection instruments include in-situ process monitors (laser power, scan speed, hatch distance), build chamber temperature profiles, post-build XCT scans for porosity and lack-of-fusion defects, surface roughness measurements, microstructural characterizations, and operational load histories captured via embedded sensor networks. The data collection plan also encompasses periodic non-destructive inspections (NDE) and accelerated life testing under calibrated machined-fatigue protocols. The study employs a design science research approach to iteratively develop and test the digital twin components, complemented by a theoretical lens drawn from reliability-centered maintenance (RCM) and the dynamic Bayesian network (DBN) framework for sequential decision making. Data analysis employs a sequence of techniques (i) feature engineering and time-series analysis of process and sensor data; (ii) regression-based survival models and random forest survival to predict remaining useful life (RUL); (iii) Bayesian updating for adaptive model recalibration as new data arrive; (iv) physics-informed neural networks to couple damage mechanics with observed degradation; and (v) cost-benefit analysis to quantify maintenance optimization. Model validation includes cross-validation, out-of-sample testing, and sensitivity analyses to assess robustness under measurement noise and process variability. Expected findings include (a) quantified relationships between AM process parameters, defect characteristics, and degradation rates under service loading; (b) a predictive maintenance policy that reduces downtime by at least 20% and extends mean time to failure (MTTF) by 15–25% relative to baseline scheduled maintenance; (c) a digital twin architecture that seamlessly ingests multi-modal data and delivers actionable maintenance recommendations in near real-time; and (d) evidence that integrating physics-based damage models with data-driven predictors yields superior predictive accuracy (increase in R2 by 10–20% and reduction in mean absolute error by 15–25%) compared with purely data-driven or purely physics-based approaches. The study contributes to knowledge by operationalizing adaptive predictive maintenance for AM parts within a data-driven digital twin framework, advancing integration of process-structure-property-performance (PSPP) linkages, and offering a transferable methodology for diverse AM technologies. The theoretical contribution lies in extending DBN-based sequential decision models to AM-specific degradation processes and in demonstrating how RCM-informed constraints enhance maintenance decision efficacy in digital twin environments. The main conclusion anticipates that AM-driven adaptive maintenance, grounded in hybrid physics-informed machine learning, can deliver measurable reliability and cost benefits in high-variability manufacturing contexts. Recommendations include scaling the framework to multi-site production, incorporating supply-chain risk into maintenance optimization, and exploring federated learning to protect proprietary process data while maintaining predictive performance.