Machine Learning–Based Predictive Quality Modeling for Defect Detection and Process Optimization in Advanced Manufacturing Systems
DOI:
https://doi.org/10.63125/thy73b16Keywords:
Machine Learning, Predictive Quality, Defect Detection, Process Optimization, Advanced ManufacturingAbstract
This study developed and quantitatively evaluated a machine learning–based predictive quality framework for defect detection, continuous quality estimation, and process optimization in advanced manufacturing systems. A quantitative experimental and longitudinal predictive-modeling design was applied to 12,480 manufacturing observations representing 3,120 production cycles, 6,240 manufactured components, 24 production batches, eight machines, and 18 synchronized sensor channels. Baseline manufacturing performance showed a 20.0% defect rate, 80.0% first-pass yield, 6.4% scrap rate, 13.6% rework rate, and process capability index of 1.18. Following preprocessing and feature selection, 186 candidate variables were reduced to 74 retained predictors, representing a 60.2% dimensionality reduction. Multiple classification and regression algorithms were evaluated using training, validation, cross-validation, and independent-test procedures. The ensemble classifier achieved the strongest defect-detection performance, with 96.4% accuracy, 94.1% precision, 92.8% sensitivity, 97.3% specificity, an F1-score of 0.934, and ROC-AUC of 0.982. Hybrid regression models explained 90.2%–94.1% of independent-test variation across major continuous quality characteristics. Tool wear was the strongest process-related predictor of defective production (adjusted OR = 2.84), while vibration magnitude was the strongest sensor-derived predictor (adjusted OR = 1.91). Multiobjective optimization reduced the experimentally observed defect rate from 20.0% to 8.3%, representing a 58.5% relative reduction, and increased first-pass yield to 91.7%. Surface roughness decreased by 31.0%, dimensional deviation by 37.7%, porosity by 41.0%, scrap by 57.8%, rework by 55.1%, downtime by 36.1%, and energy consumption by 12.3%. Process capability increased from 1.18 to 1.52, while overall equipment effectiveness increased from 72.8% to 84.6%. Manufacturing cost per unit decreased by 14.0%, and cost of poor quality declined by 54.0%. Principal baseline-to-optimized differences were statistically significant at p < .001 and were supported by moderate-to-large effect sizes. The findings demonstrated that integrating multisensor manufacturing data, machine learning, robust validation, and multiobjective optimization provided a quantitatively effective framework for improving defect detection, product quality, process stability, resource efficiency, and manufacturing performance.


