International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

TriFuse-PdM: High-Fidelity Machine Failure Prediction Using Hybrid Resampling and Model Calibration

Authors :

Saghar Shafaati¹ Javad Mohammadzadeh²

1- Department of Computer Engineering, South Tehran Branch,Islamic Azad University, Tehran, Iran 2- Department of Computer Engineering, Ka.C., Islamic Azad University, Karaj, Iran

Keywords :

Predictive Maintenance (PdM)،Class Imbalance،Ensemble Learning،SHAP Explainability،Hybrid Resampling (SMOTE-Tomek)

Abstract :

Abstract— Predictive maintenance (PdM) is vital for modern industrial systems, enabling timely failure prediction through data-driven methods. Yet, class imbalance in failure datasets significantly hinders the accurate detection of rare but critical events. This paper presents TriFuse-PdM, a unified and interpretable PdM framework that integrates SMOTE-Tomek resampling, cost-sensitive learning, ensemble models (Random Forest, XGBoost), and a deep Multi-Layer Perceptron. SHAP is used for interpretability, and performance is assessed via ROC-AUC, PR-AUC, MCC, and calibration analysis on the AI4I 2020 dataset. TriFuse-PdM achieves high predictive accuracy (ROC-AUC ≈ 0.995, PR-AUC > 0.99) while ensuring reliable uncertainty estimation and actionable insights, offering a scalable solution for Industry 4.0 maintenance challenges. Specifically, the framework tackles class imbalance through SMOTE-Tomek resampling and cost-sensitive training, enhances interpretability via SHAP analysis, and improves prediction reliability through probability calibration.