Parallel Generative Adversarial Imputation Network for Multivariate Missing Time-Series Reconstruction and Its Application to Aeroengines
Song Ma, Zeng-Song Xu, Tao Sun
Abstract
The reconstruction and imputation of missing values in multivariate time series (MTS) is a pressing issue in the field of industrial artificial intelligence. To address this problem, an end-to-end deep learning model called the “time series generative adversarial imputation network (TSGAIN) with pre-imputation, parallel convolution and transformer encoder (PPCTE)” is proposed in this paper. Specifically, the proposed TSGAIN framework is an improved generative adversarial network for imputation of missing time series, which can take into account both distribution and time information. Next, a pre-imputation module is advanced to obtain more accurate input information for the proposed model. Then, as two crucial components of the proposed model, the generator and discriminator are further improved. In particular, in terms of feature correlation, a parallel convolution module is designed, which further enhances the potential of the proposed model to extract feature correlation in multivariate data. In terms of temporality, by introducing the transformer encoder module with multi-head attention mechanism, the ability of the proposed model to mine the temporal correlation of time series data is further strengthened. Finally, a series of simulation experiments are carried out on the commercial modular aviation propulsion system simulation (C-MAPSS) experimental data provided by NASA, verifying the effectiveness of the proposed modules and the superiority of the proposed method in the imputation process of aero-engine missing sensor data.