Predicting the Remaining Useful Life (RUL) of Computer Numerical Control (CNC) machines is key in Predictive Health Management (PHM), where missing data is a persistent challenge. The spatio-temporal-duration coupling in sensor data complicates missing data recovery. This paper introduces a spatio-temporally decoupled RUL prediction method, STDP, that uses duration-focused imputation under missing data constraints. Employing a cross-dimensional self-attention (CDSA) approach for both imputation and prediction, we also developed a bidirectional reconstruction-prediction training strategy to dynamically adjust losses for data reconstruction and RUL prediction tasks. Our experimental results, across various missing rate scenarios, demonstrate that our method surpasses existing advanced methods in data imputation and RUL prediction, especially at high missing rates, thereby significantly enhancing prediction reliability and model stability in the manufacturing sector.