Machine Learning-Based Failure Analysis in Industrial Manufacturing: A Comparative Evaluation under Realistic Data Constraints

Industrial failure analysis (FA) increasingly relies on data-driven approaches, particularly machine learning (ML), to enhance production efficiency, ensure process stability, and support decision-making of human operators. However, these methods face challenges in complex manufacturing environments, including limited data availability, data quality, process dynamics, and data variability across production stations. While prior research has demonstrated the potential of ML algorithms for FA, the impact of real-world data constraints on model performance remains insufficiently explored. This study presents a comparative evaluation of state-of-the-art ML models, including decision tree, random forest, logistic regression, and artificial neural networks, under realistic industrial limitations. Using a real-world dataset from automotive final assembly, we develop an intelligent FA system that leverages ML to assist assembly line workers in interpreting error protocols and identifying roo t causes. To emulate industrial conditions, targeted perturbation functions are implemented that simulate data scarcity, noise, concept drift, and increased analytical complexity. ... mehrAcross the analyzed perturbation settings, random forest shows highest robustness. Yet, all models degrade substantially under realistic challenges, with drops in the F1-score by over 50 % under heavy drift. These results quantify how real-world data-constraints affect different classifiers and provide guidance for stress-testing models before deployment.