DTIN: Dual Transformer-based Imputation Nets for multivariate time series emitter missing data

Abstract

As a kind of multivariate time series (MTS) data, emitter signals often exhibit missing or corrupt values, posing serious challenges to emitter data research such as specific emitter identification (SEI). Existing multivariate missing data imputation (MDI) methods are deficient in two aspects when applied to MTS emitter data: first, single-channel models cannot handle variables with varying numbers of complete samples; second, they cannot efficiently impute MTS data that are out of the model’s domain. To address these issues, a dual channel architecture tailored for MTS emitter data was devised in this study, which is called Dual Transformer-based Imputation Nets (DTIN). DTIN processes different types of variables through two parallel channels to extract different spatiotemporal features. Furthermore, drawing inspiration from image style manipulation, multivariate time series pivotal tuning inversion (MTSPTI) techniques are employed for better imputation performance, in which an in-domain pivotal code is created and input into the generator that is tuned for out-of-domain MTS emitter data. Extensive experiments on two real-world emitter datasets demonstrate that DTIN outperforms several existing MDI models.