ECGAIN: An Enhanced Conditional Missing Data Imputation Method For Process Industries

Missing data is a quite pervasive phenomenon in process industries due to the failure of signal transmission and data storage, which brings a great challenge to the downstream data-driven modeling. Existing missing data imputation (MDI) methods do not consider the sparsity of observed data which would affect the imputation performance and impair the modeling accuracy of downstream applications further. Therefore, a novel Enhanced Conditional Generative Adversarial Imputation Network (ECGAIN) is proposed in this paper to improve the MDI performance when the observed data is sparse. Specifically, the generator observes some observed data and outputs a vector. The discriminator takes the vector as the input to judge which one is actually observed and which one is imputed by generator. During the above process, two data enhancement mechanisms termed adaptive update mechanism (AUM) and adaptive attention mechanism (AAM) are designed to improve the MDI performance. Meanwhile, label-relevant information is taken as the condition of both the generator and discriminator to guide the imputation. Case studies on two University of California Irvine (UCI) public datasets and a real-word thermal power process dataset have proved the reliability and effectiveness of the proposed ECGAIN by comparing with other state-of-the-art MDI methods.