An Explainable Artificial Intelligence Framework for Predictive Maintenance Using Hybrid Machine Learning and Multi-Task Transformer Models

Abstract

An integrated explainable artificial intelligence (XAI) framework for predictive maintenance is proposed, focusing on failure prediction, Remaining Useful Life (RUL) estimation, and real-time asset monitoring. The framework is evaluated using a synthetically generated predictive maintenance data simulating 50 machines over time, producing realistic multivariate sensor signals that degrade with wear and occasional shocks, then converted into a supervised dataset for failure prediction and RUL estimation. The study evaluates several classical machine learning (ML) models, such as XGBoost (Extreme Gradient Boosting), Random Forest (RF), Gradient Boosting (GB), Logistic Regression (LogR), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN), for classification and regression tasks. XGBoost demonstrated strong overall performance, achieving the best RUL regression performance and competitive classification results, while RF and GB achieved comparable or slightly better results for individual classification metrics. A multi-task transformer is introduced to jointly model failure probability, uncertainty-aware RUL, anomaly reconstruction, and survival-based time-to-failure estimation, thereby providing a unified framework for modeling multiple temporal predictive-maintenance objectives. An autoencoder is included as a supplementary reconstruction-based component for anomaly scoring, while a risk function combines failure probability and RUL to enable real-time decision-making. To ensure interpretability, SHapley Additive exPlanations (SHAP) analysis is used to quantify feature importance at the sensor level. The framework is evaluated using Receiver Operating Characteristic (ROC) curves and Precision-Recall curves, along with Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R²) metrics. These results demonstrate the strong performance of ensemble models, particularly XGBoost. Overall, the framework demonstrates the potential of combining tree-based models, transformer-based temporal learning, and SHAP-based interpretability for predictive maintenance under synthetic sensor conditions.