The field data of wind turbine data acquisition and supervisory control system are commonly missing, which have a certain negative influence on the downstream condition monitoring task. To address this issue, a mask autoencoder network with attention mechanism is proposed to impute missing values in panel data samples, increase the number of available samples, and improve the accuracy and continuity of condition monitoring results. The method takes the denoising autoencoder network as the overall framework. In the encoding stage, the missing values are masked by the attention mechanism, and the missing values are given a higher weight to strengthen the attention of the network. In the decoding stage, the complete data samples are output after missing values imputation. Then, the parameter of the target variable is predicted by using the sample features extracted by long short-term memory network, and the condition monitoring is realized according to the prediction residual. This method is evaluated by the operation data of a wind turbine gearbox. Results show that the data imputation bias of the proposed method is at least 17. 2% better than that of the comparison method. Compared with before data imputation, the number of samples increased significantly after data imputation, which makes the prediction residual of normal data decreased by 37. 4% on average and the detection rate of fault data increased by 6. 8% .