A data-driven approach to anomaly prediction in automotive display manufacturing

Abstract

This study investigates the Wet Optical Bonding (WOB) process in automotive display manufacturing within an Industry 4.0 framework. The objective is to enable early detection of display defects, such as air bubbles and particle contamination, by leveraging tabular WOB input features available before the curing stage, including glass measurements. The anomaly detection task is approached using a range of machine learning (ML) methods. These include Bayesian optimized binary classifiers such as XGBoost, CatBoost, and Deep Feedforward Neural Network; the Automated ML (AutoML) H2O tool; and Bayesian optimized one-class learners, including Isolation Forest and deep Autoencoders. A large industrial dataset of approximately 64,000 WOB records was used to conduct extensive predictive experiments. The evaluation followed a rigorous protocol with internal 3-fold cross-validation for validation and external 10-fold cross-validation for testing, assessing both predictive accuracy and computational efficiency. The ML models demonstrated strong discriminatory performance while maintaining reasonable computational requirements. In addition, a deployment analysis illustrated the potential for reducing the cost and cycle time of the WOB process. Finally, a sensitivity analysis using explainable artificial intelligence (XAI) techniques was conducted to highlight the relevance and influence of key WOB input features.