Abstract
Large-scale mining equipment operates under extreme conditions, making accurate fault prediction and remaining useful life (RUL) estimation essential for predictive maintenance strategies. This paper proposes a novel deep learning framework that integrates temporal convolutional networks (TCN) with multi-head attention mechanisms for prognostic applications in mining machinery. The TCN backbone employs dilated causal convolutions to capture long-range temporal dependencies from multivariate sensor data, while a dual-branch attention module adaptively emphasizes informative features along both temporal and channel dimensions. A multi-task learning architecture with uncertainty-based loss weighting enables simultaneous optimization of fault classification and RUL regression objectives. Experimental validation on real-world data collected from haul trucks and hydraulic excavators demonstrates superior performance compared to baseline methods. The proposed model achieves 92.47% accuracy in fault prediction and 98.45 h RMSE in RUL estimation, with an R² coefficient of 0.912. Ablation studies confirm the contribution of each architectural component, while robustness testing reveals graceful degradation under sensor dropout and measurement noise conditions. The framework provides mining enterprises with a practical solution for enhancing operational reliability and maintenance scheduling efficiency. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1038/s41598-026-43145-z.