<p>Remaining Useful Life (RUL) prediction of stochastic degrading devices is crucial for Prognostics and Health Management (PHM). Traditional methods often rely on empirically predefined failure thresholds, which are challenging to determine for complex systems, and hybrid-driven approaches struggle with integrating Composite Health Indicator (CHI) construction and degradation modeling. To address these limitations, this paper proposes a novel threshold-independent data-model interactive approach. The key innovation lies in the integration of an improved Transformer model for CHI extraction and a linear Wiener process for degradation modeling, with the failure threshold treated as an independent condition. This approach establishes a comprehensive feedback loop between CHI construction and degradation modeling by minimizing an optimization objective function that incorporates the failure threshold as a conditional constraint. The improved Transformer model enhances the traditional architecture by discarding the decoder layer and introducing a gated convolutional unit to better incorporate locally relevant information from multi-sensor data into the attention mechanism. While the Wiener process provides an analytical expression for RUL prediction, including uncertainty quantification. The proposed method is validated on the C-MAPSS turbofan engine dataset and the PHM2010 CNC milling tool dataset. For engine prognostics, it achieves a Score of 1.42 and an RMSE of 8.7, outperforming existing methods. For tool wear prediction, it attains an RMSE of 5.27 and a MAPE of 7.52%, demonstrating superior accuracy and robustness. These results highlight the method’s versatility and effectiveness across diverse degradation scenarios, offering a significant advancement in RUL prediction for industrial applications.</p>