Machine failure prediction and interpretation for operator support in industrial production settings : A methodological study using the AI4I 2020 predictive maintenance dataset

Sammanfattning

This thesis investigates how process and machine variables can be used to predict machine failure and interpret failure-related operating conditions in a predictive maintenance setting. The study is situated within the broader context of operator support in industrial production settings. Since real industrial data were not available, the AI4I 2020 Predictive Maintenance Dataset was used as a synthetic industrial proxy dataset for methodological development. The analysis focused on identifying variables associated with machine failure and evaluating supervised learning models under class imbalance. To support this aim, an analytical workflow combining exploratory data analysis, supervised machine learning, class-sensitive evaluation and model interpretability was developed. Logistic Regression, Decision Tree, Random Forest and XGBoost were compared using PR-AUC, ROC-AUC, recall, precision, F1-score and macro-F1, while F2-score supported threshold tuning. Permutation importance and SHAP were then used to interpret model behaviour. A synthetic-rule sensitivity audit was also conducted to examine how much of the model's performance reflected the deterministic failure-generation structure of the AI4I dataset rather than transferable sensor patterns. The results showed that torque, tool wear, rotational speed and air temperature were the variables most consistently associated with machine failure in the selected dataset. Among the evaluated models, XGBoost achieved the strongest overall balance across the selected metrics and was therefore identified as the most suitable model in this study. This thesis contributes an interpretable analytical workflow for machine failure prediction and operator-support-oriented interpretation. However, the findings should not be interpreted as direct evidence of industrial transferability, as the dataset is synthetic and does not represent real industrial production data. Future work should therefore apply and validate the workflow using real industrial data and feedback from domain experts.