准确预测复杂装备的剩余使用寿命(Remaining Useful Life, RUL)，直接关系其运行安全与运维成本。面对实际工程中普遍存在的数据缺失问题，现有两步式修复预测策略误差传递放大、联合训练中梯度冲突与退化语义漂移等核心难题，严重影响了剩余使用寿命预测的准确性和可靠性。针对上述问题，提出了一种数据缺失场景下数模联动的剩余寿命预测方法。首先，在完整监测数据上预训练融合卷积神经网络与Wiener退化过程的数模联动预测模型，建立多源传感数据、复合健康指标演化与剩余寿命之间的内在映射关系；其次，采用基于自注意力的时序插补网络作为修复基础架构，构建由数据重构损失、复合健康指标特征损失和剩余寿命误差损失组成的三层一致性约束，从数据层、特征层和任务层对修复过程进行联合优化；同时，引入投影冲突梯度协调策略解决多目标训练中的梯度冲突问题，并通过半监督轨迹对齐提升方法在无标签在线场景下的适用性。最后，基于NASA C-MAPSS航空发动机数据集的对比实验与消融实验结果表明，所提方法在不同缺失率下均表现出较高的RUL预测精度和鲁棒性。

Accurate prediction of the Remaining Useful Life (RUL) of complex equipment is directly related to their operational safety and operation and maintenance costs. The pervasive problem of missing data in practical engineering, together with core challenges such as error propagation and amplification in existing two-stage repair-then-prediction strategies, gradient conflicts, and degradation semantic drift in joint training, has seriously affected the accuracy and reliability of RUL prediction. To address these issues, a data-model interactive RUL prediction method for missing data scenarios is proposed. Firstly, a data-model interactive prediction model integrating Convolutional Neural Network (CNN) and Wiener degradation process is pre-trained on complete monitoring data to establish the intrinsic mapping relationship among multi-source sensor data, Composite Health Index (CHI) evolution, and RUL. Secondly, the Self-Attention-based Imputation for Time Series (SAITS) is adopted as the basic repair architecture, and a three-layer consistency constraint composed of data reconstruction loss, CHI feature loss, and RUL prediction error loss is constructed to jointly optimize the repair process from the data layer, feature layer, and task layer. Meanwhile, the Projecting Conflicting Gradients (PCGrad) strategy is introduced to address the gradient conflict problem in multi-objective training, and semi-supervised trajectory alignment is used to enhance the applicability of the method in unlabeled online scenarios. Finally, comparative experiments and ablation studies on the NASA C-MAPSS aero-engine dataset demonstrate that the proposed method achieves high RUL prediction accuracy and strong robustness under different data missing rates.