Remaining useful life prediction enhanced by multi-task gradient modulation

In the field of remaining useful life (RUL) prediction, multi-task learning (MTL) has proven effective in enhancing model generalization by utilizing complementary degradation knowledge extracted from related auxiliary tasks. However, learning multiple tasks simultaneously can be challenging due to potential conflicts between tasks, which may lead to gradient interference and result in suboptimal performance. To address this issue, this paper proposes a novel RUL prediction method based on multi-task gradient modulation (MGM). The method effectively captures complementary degradation knowledge by incorporating three auxiliary tasks: health state classification, health indicator prediction, and future sequence prediction . Additionally, a new MGM module is developed to modulate the gradients of these auxiliary tasks. By dynamically adjusting the contributions of auxiliary tasks to minimize discrepancies with the RUL prediction task and applying orthogonal projectio n to mitigate conflicting gradient components, the MGM module promotes knowledge complementarity between tasks. As a result, the proposed method achieves higher precision and stability in predicting RUL. The experimental study based on the turbofan engine dataset verifies the method's effectiveness and advantages.