The Smart Kalman Filter: A Deep Learning-Based Approach for Time-Varying Channel Estimation

In digital wireless communications, the received signal can be strongly altered by the environment and may contain Inter-Symbol Interference (ISI). To remove or reduce the ISI i.e. equalize, the impulse response of the propagation channel can be estimated. The Kalman Filter (KF) is an inescapable estimation algorithm in linear systems because of its optimality in terms of Minimum Mean Square Error (MMSE) under certain assumptions. However, in real conditions, implementations of KF are often difficult because of the necessity of hand-tuning parameters. In this paper, we present the Smart Kalman Filter (SKF), a hybrid architecture that combines a KF and the power of neural networks to extract relevant features from data to benefit from an adaptive KF that is automatically well tuned over time. We demonstrate in this paper that the proposed SKF is up to 5dB better at low Signal-to-Noise Ratio (SNR) and 3dB better at high SNR than Least Square (LS) algorithm in a time-varying channel estimation context with abrupt Doppler frequency variations.