WARDEN: Wiener-Adaptive Residual Dynamic ENcoder for Channel Charting in Future Cellular Networks

Channel Charting is a self-supervised learning technique that maps high-dimensional Channel State Information into a low-dimensional latent space preserving the spatial topology of user equipment trajectories, without relying on ground-truth position labels. While recent work has extended Channel Charting with neural temporal predictors, these approaches learn latent dynamics entirely from data, ignoring the analytical structure available from classical estimation theory and offering limited interpretability. In this paper, we propose Wiener-Adaptive Residual Dynamic ENcoder (WARDEN), a novel Channel Charting architecture that decomposes latent-space dynamics into two components: an analytically optimal linear transition matrix, derived via the Wiener-Hopf solution to the minimum mean-square error prediction problem on the latent sequence, and a lightweight neural residual network that captures non-linear corrections beyond the linear model. Training proceeds in two stages: a bootstrap phase that estimates the transition matrix from a pre-trained encoder, followed by joint end-to-end optimisation with periodic re-estimation. WARDEN is validated on four environments, comprising three real-world massive MIMO measurement datasets with distributed and co-located arrays and a synthetic outdoor ray-tracing scenario, using identical hyperparameters throughout, and is compared under a unified protocol against representative learned temporal mechanisms based on recurrent and joint-embedding predictive architectures. Across all environments and over multiple training seeds, chart Trustworthiness improves consistently with the regularisation strength across all environments (up to +12.7%), the linear prediction error decreases by 35% to 61%, and the localisation error on the real datasets is reduced by up to 11.4%, with statistical significance confirmed by paired tests. Compared with learned temporal predictors, WARDEN exhibits the lowest seed variance and is the only method among those compared that improves latent dynamics without degrading any geometric quality metric, while adding no inference overhead. We further demonstrate that the eigenstructure of the transition matrix provides interpretable spectral diagnostics of user equipment kinematics, and that the residual norm serves as a supervision-free measure of trajectory non-linearity.