Attention-Based Deep Learning for Predictive Maintenance: A Case Study Using a Real-World Multivariate Sensor Dataset

Predictive maintenance (PdM) is a critical component of Industry 4.0 strategies, leveraging sensor-derived time-series data to forecast equipment failure and optimize maintenance scheduling. This study introduces an attention-enhanced multi-head LSTM (MH-LSTM) framework utilizing the real-world SCANIA Component X dataset. The dataset comprises over 10,000 multivariate instances capturing temporal sensor readings and repair records for heavy-duty truck engine components. The proposed MH LSTM model addresses two key tasks: remaining useful life (RUL) estimation and binary classification of failure within a ten-day forecast horizon. We benchmarked the performance against Linear Regression (LR), Random Forest (RF), and standard LSTM baselines. Experimental results demonstrate that the MH LSTM achieves the lowest forecasting error (RMSE ≈ 1.0 days, MAE ≈ 0.8 days), outperforming standard LSTM (RMSE ≈ 1.6 days, MAE ≈ 1.2 days) and classical models with substantially higher errors. In the classification task, MH LSTM reaches an accuracy of approximately 0.88, with balanced precision and recall metrics. Feature importance analysis and visual diagnostics, including scatter plots and confusion matrices, confirm that the attention mechanism enables the model to focus selectively on critical sensor channels and temporal segments. These findings underscore the advantages of attention mechanisms for capturing intricate temporal dependencies and enhancing predictive performance in real-world PdM settings. The methodology and codebase are fully reproducible, making this work a valuable reference for future industrial applications and academic comparisons.