Machine learning for zero defect manufacturing

Achieving Zero Defect Manufacturing (ZDM) requires the integration of intelli-gent, data-driven systems capable of fault detecting and prevention. Deep learning models particularly convolutional neural networks (CNNs), and ensemble learning tech-niques enhance the reliability of industrial processes by automating complex data analy-sis. This study addresses critical challenges, including data imbalance, model inter-pretability, and real-time deployment. On industrial image datasets, EfficientNet-B3 achieved 99.97% accuracy, ResNet-50 reached 99.89%, and MobileNetV3 delivered 99.65% accuracy with the fastest inference (119.88 FPS). For predictive maintenance, GAN-augmented XGBoost and SimCLR-enhanced XGBoost achieved F1-scores of 0.993 and 0.992 for Random Failures. A Few Shot Prototypical Network delivered 0.988 for Random Failures and near-perfect detection of Power (0.999) and Tool Wear Failures (0.991). Explainability was ensured through Grad-CAM visualizations, enabling transpar-ent and traceable model decisions. Weighted sampling and custom loss functions effec-tively addressed data imbalance while lightweight CNNs and augmented predictive models were optimized for real-time, low-power deployment on Raspberry Pi 5, directly supporting industrial ZDM. By bridging the gap between theoretical advances in machine learning and practi-cal industrial implementations, this study contributes to the evolution of flexible, scala-ble, and sustainable manufacturing systems within the industry 4.0 paradigm.