Neural-Network-Switched Kalman Filters as Novel Trackers for Multipath Channels

High mobility leads to fast-varying, non-stationary channels in some modern applications, such as wireless communications for High-Speed Railways (HSR). This results in nonlinear transitions, namely the potential birth of a new tap in a multipath channel or an active tap's death. A pressing question then is how to make use of unexploited correlations, such as time correlation in each tap of a multipath channel, when the Wide-Sense Stationary Uncorrelated Scattering (WSSUS) condition can no longer be assumed. Whereas Kalman filtering (KF) has been proposed to exploit such time correlation in each tap under WSSUS scenarios, a capital disadvantage of KF is its weak performance when non-linear transitions are considered. This work reviews previous proposals to tackle this birth-death nonlinearity problem, as well as their drawbacks, and derives from them a new neural-network switching concept, called Neural-Network-switched Kalman Filter (NNKF). This novel tracker is computationally inexpensive and its simulations hereby show that it outperforms all previously known multipath channel tracking systems. The proposed tracker achieves the performance of the ideal birth/death detection case, thus approximately halving squared error w.r.t. Least-Squares (LS) estimation in Orthogonal Frequency Division Multiplexing (OFDM) systems.