A directed cross-task attention Transformer for joint fault diagnosis and prognostics of rail engine compressors

Reliable condition monitoring of multi-sensor mechanical systems requires the joint assessment of the fault state and its temporal progression. However, fault diagnosis and prognosis are commonly treated as separate tasks, whereas existing multitask approaches model their interaction implicitly through shared representations. It does not explicitly account for the relationship between diagnosis and prognosis, where information about the developing fault type can provide useful context for estimating time remaining before fault onset. To address the gap, this study introduces a directed cross-task attention within a multitask Transformer, in which the fault-diagnosis representation informs the time-to-onset prediction, rather than relying implicitly on a shared encoder. Learnable task tokens are used to derive task-specific representations through an attention mechanism, whereas an uncertainty-based auto-weighted loss adaptively balances the learning objectives during training. The framework is validated on multivariate sensor data from a rail engine air compressor comprising 14,712 air and 218,846 oil leak readings. The proposed framework achieves 95.2% classification accuracy and time-to-onset errors of 9.3 and 3.7 min for air and oil leaks, respectively. It achieves strong overall performance against deep learning baselines, with the component-wise ablation confirming the substantial contribution of both the directed cross-task attention and auto-weighted loss. Furthermore, the model detects air leak 96 min and oil leaks more than 2 days ahead of the built-in alarm system. The model also demonstrates robustness to missing data, noise, and drift while remaining sensitive to sparse outlier spikes. The results demonstrate the effectiveness of modeling dependency between fault diagnosis and time-to-onset prediction for integrated condition monitoring of multi-sensor mechanical systems.