A Transformer-based Multi-Platform Sequential Estimation Fusion

This paper considers estimation fusion problem in the case of unknown correlations among local estimates, motivated by multi-sensor target tracking with correlated measurement noises. A Transformer-based sequential multi-platform fusion method is put forward by learning data features of historical local tracks, instead of numerical optimization in existing weighting fusion. Firstly, a neural network-based sequential fusion framework is proposed, where it owns a hierarchical structure and sequential training process to adapt to different numbers of local tracks without changing network parameters and retraining. Secondly, the Taylor expansion-based positional encoding in Transformer network is constructed, by using a third-order Taylor expansion to approximately replace original sin and cos functions to better extract aperiodic variation features of input sequence. Thirdly, by arranging different local estimates of input sequence in time order, a max–min normalization-based data pre-processing and its inverse process are presented, to prevent precision truncation and retain data diversity. An example of target tracking with multiple sensors show that the proposed method owns superior fusion precision than that of the sequential filter, simple convex combination, covariance intersection and Long Short-Term Memory-based sequential fusion methods, in terms of different correlation coefficients. And its fusion precision is also improved with the increasing of sensor numbers.