Spatio-Temporal Missing Data Imputation With Cross Modality in HCPSs

Abstract

Time-series data missing is a common problem, which often happens with irregular sampling in sensor device failure in human-cyber-physical systems (HCPSs). The generation of networked time-series data is conducive to achieving real-time perception in HCPSs. Many methods exist for imputing random or non-random missing data, but their accuracy is often inadequate at high missing rates. We propose a cross-modality approach using dense spatio-temporal transformer networks (DSTTN) to impute high-rate missing data in time series. The DSTTN merges the spatio-temporal modal data by cross-modality data fusion technique, and then constructs an end-to-end transformer pipeline with dense skip connections to recover the corrupted data accurately. We have conducted many comparative experiments to assess DSTTN imputation performance in the MAR and missing not at random (MNAR). Cross-modality data fusion offers a new solution for complete data missing, a specific case of MNAR. Furthermore, we also compare and analyse the various recent models, and the particularities between them. Based on the comparative analysis, the application value and working conditions of the DSTTN are demonstrated in detail by the results of rich experiments.