A predictive analytics framework for early detection of production halts and quality issues

Abstract

This study presents a Machine Learning (ML) framework for an Ahead-of-Time (AoT) prediction of production halts and defects in particleboard manufacturing that uses only pre-production input variables. The proposed approach incorporates both Single-Task Learning (STL) and Multi-Task Learning (MTL) paradigms, which are evaluated across three production lines under two modeling strategies: Line-Specific Modeling (LSM) and Line-Agnostic Modeling (LAM). The experimental evaluation benchmarks a lightweight Logistic Regression (LogR) model against three Automated Machine Learning (AutoML) techniques: H2O AutoML, Ludwig, and a Bayesian-optimized Deep Feedforward Network (DFFN). Results show that the STL-LSM combination using LogR achieves the highest overall predictive performance. To enhance model interpretability, we apply two model-agnostic eXplainable Artificial Intelligence (XAI) techniques: SHapley Additive exPlanations (SHAP) and One-Dimensional Sensitivity Analysis (1DSA). These methods generate feature importance rankings across targets and production lines, which are evaluated using quantitative (normalized distance metrics) and qualitative measures (alignment with domain expert insights). The XAI findings reveal a strong consistency between SHAP and 1DSA, with 1DSA requiring a substantially lower computational cost. Moreover, the convergence between model-derived interpretations and expert feedback highlights the practical relevance of the proposed ML framework in supporting data-driven decision-making for particleboard production planning.