Accurately predicting the remaining useful life (RUL) of machines is vital for assessing machine health and minimizing economic losses resulting from downtime in sensor-equipped machines. However, real-world applications often encounter challenges such as rapid production cycles and unstable network conditions, inevitably leading to significant amounts of missing data. This challenges data-driven machinery RUL prediction, as conventional deep learning methods may struggle with missing data, impacting prediction accuracy. To address the issue, a missing data reconstruction method based on self-learning of mask spatial distribution is proposed. The structured spatial distribution characteristics of the mask, learned by the autoencoder, serve as self-supervised information for the imputation network to improve the data reconstruction performance. Meanwhile, a multi-task learning-enhanced prediction network architecture with adaptive weight adjustment is designed, defining tasks by RUL prediction under different data reconstruction accuracies. After pre-training on multiple tasks, the prediction network's learning efficiency benefits from incorporating both common and task-specific rules for feature extraction from similar reconstructed data distributions. The proposed method is evaluated through ablation and comparative tests on application scenarios and standard datasets. Experimental results show that the proposed algorithm performs competitively against state-of-the-art data reconstruction algorithms on these test suites.