LWGAIN: An Improved Missing Data Imputation Method Based on Generative Adversarial Imputation Network

In the field of petroleum exploration and development, the location of reservoir is intrinsically linked to lithology. Therefore, accurately identifying the underground lithology using well logging data is of paramount importance. However, due to unforeseen circumstances such as equipment failures, there may be amounts of missing log data. Hence, addressing these missing data is crucial in lithology identification. This paper introduces a method for imputing missing data, known as labels Wasserstein generative adversarial imputation network (LWGAIN). This method builds upon the traditional generative adversarial imputation network (GAIN), incorporating Wasserstein distance as a loss function expression, and adding labels to the input of generator to enhance the available information of network. Furthermore, a method for selecting labels is proposed. Experiments conducted on the Kansas logging dataset indicate that the root mean square error (RMSE) and mean absolute error (MAE) of LWGAIN are lower than those of other methods. Moreover, the imputed data can be effectively used for lithology identification with satisfactory accuracy.