Abstract
We consider data-driven Bayesian state estimation from compressedmeasurements (BSCM) of a model-free process. The dimension of the temporalmeasurement vector is lower than that of the temporal state vector to beestimated, leading to an under-determined inverse problem. The underlyingdynamical model of the state's evolution is unknown for a 'model-free process.'Hence, it is difficult to use traditional model-driven methods, for example,Kalman and particle filters. Instead, we consider data-driven methods. Weexperimentally show that two existing unsupervised learning-based data-drivenmethods fail to address the BSCM problem in a model-free process. The methodsare -- data-driven nonlinear state estimation (DANSE) and deep Markov model(DMM). While DANSE provides good predictive/forecasting performance to modelthe temporal measurement data as a time series, its unsupervised learning lackssuitable regularization for tackling the BSCM task. We then propose asemi-supervised learning approach and develop a semi-supervised learning-basedDANSE method, referred to as SemiDANSE. In SemiDANSE, we use a large amount ofunlabelled data along with a limited amount of labelled data, i.e., pairwisemeasurement-and-state data, which provides the desired regularization. Usingthree benchmark dynamical systems, we empirically show that the data-drivenSemiDANSE provides competitive state estimation performance for BSCM using ahandful of different measurement systems, against a hybrid method calledKalmanNet and two model-driven methods (extended Kalman filter and unscentedKalman filter) that know the dynamical models exactly.