In order to solve the problem of missing measurement data in cyber physical systems (CPSs) after being attacked by denial of service (DoS), a missing data recovery strategy on data imputation is proposed. First, the aperiodic, resource constrained DoS attack model is constructed, and random packet loss is introduced to simulate the complexity of cyber-attacks in actual CPSs. For the problem of system measurement data recovery, a novel data imputation algorithm that leverages a diagonal masking self-attention mechanism is introduced. To enhance the accuracy of data imputation and expedite the training process, a diagonal masking mechanism is employed to mitigate the model's reliance on its own values. Subsequently, a weighted combination of the diagonal masking self-attention modules is utilized to further refine the imputation results. The experimental results of power CPSs show that, compared with several deep learning algorithms, the proposed data recovery method can improve the accuracy and efficiency of the system measurement data recovery in complex communication environment, and enhances the anti-attack ability and stability of the system.