Predicting Product Failure Across Sequential Manufacturing Steps

Abstract

Modern manufacturing systems face increasingly rigorous quality requirements, with sustainable production demanding a minimization of waste at lower production costs. As manufacturing systems become increasingly automated and complex, maintaining consistently high product quality depends to a greater extent on the effective use of process and sensor data. Machine learning (ML) offers promising tools for detecting hidden patterns in industrial process data, however its application to fault prediction across sequential multi-step production lines remains underexplored. This thesis investigates whether direct and recursive modeling strategies can reveal predictive potential in a cable manufactuing line by identifying those steps of the production process that contribute most strongly to product failures. The study develops and evaluates four models (direct and recursive variants of XGBoost and Multi-Layer Perceptron) on a dataset collected from an actively operating industrial cable production line. Models were optimized using cross-validation with Optuna, and their performance was assessed on unseen test data through confusion matrices, precision–recall analysis, and ROC-AUC metrics. SHAP feature importance analysis was used to determine which process variables excerted the most influence on fault predictions. Results show that all models struggled to reliably detect defective parts, with a persistent bias toward the majority class in the direct approach. Recursive models produced slightly more balanced predictions than direct models, offering a modest stabilizing effect across process slots. However, none of the approaches achieved the required robustness suitable for deployment. Feature importance analysis further revealed the absence of strong predictive signals within the dataset: while direct models highlighted some potentially relevant process parameters, recursive models relied exclusively on their own propagated predictions, indicating an inability to incorporate new slot-specific information. The findings confirm that the modeling pipeline itself is valid, as demonstrated on artificially altered datasets with amplified signals, but that the original production data lacks sufficient predictive quality. Methodologically, the comparison of direct and recursive strategies demonstrates that both approaches are theoretically capable of localizing predictive potential. Conceptually, this study contributes to a better understanding of how ML can be utilized in sequential manufacturing contexts and provides a foundation for future research directions, such as optimizing the prediction horizon or focusing on critical process blocks.

Date of Award	2025
Original language	English
Supervisor	Bogdan Burlacu (Supervisor)