Improving Root Cause Analysis of Production Defect Using AI: A Case Study in an Automotive Manufacturing Plant

Abstract

In automotive manufacturing, repetitive defects often occur across different time periods, creating a valuable historical dataset containing defect names and their corresponding root causes. Traditionally, identifying the root cause of a production defect relied heavily on human analysis, requiring significant time and on-site inspection. This often led to delayed countermeasures, increased production downtime, and additional issues such as line stops. This study presents an AI-based approach to assist root cause analysis using historical defect data, aiming to reduce the analysis time and improve feedback accuracy. The implementation focused on enabling faster and more accurate identification of root causes by integrating a machine learning model into the factory’s defect recording system (ATPPM, Analisa Tindakan Penanggulangan dan Pencegahan Masalah). The development process involved data preprocessing, model training, and API deployment. The original dataset consisted of 3,128 records, which were cleaned and reduced to 1,449 labeled entries, each annotated with one of 161 unique root cause labels. Eleven machine learning models were evaluated, including Logistic Regression, Random Forest, SVM, and RNN. Initial evaluation using F1-score, precision, and recall showed Logistic Regression achieving the best F1-score of 0.83. Further validation using 5-Fold Cross Validation identified the Support Vector Machine (SVM) as the best-performing model, with an average accuracy of 89.1%. This model was deployed via a Python Flask API and integrated into the existing ATPPM system. The AI-powered system significantly accelerated the root cause analysis process, reducing the average analysis time by 228 minutes. Potential future enhancements involve automating the model’s training process on a regular schedule (weekly or daily), integrating additional data sources including big data and quality management systems, and scaling the current API implementation to multiple production lines for wider impact.