Multi-level Deep Learning Kalman Filter

The well-known Kalman filter and its adaptive variants belong to model-based optimization, and their optimality depends on reliable prior information such as system models, which is sometimes hard to obtain. To reasonably introduce prior domain knowledge on the basis of offline data learning, a multi-level deep learning Kalman filter is designed in this paper with dynamic model parameter learning for evolution trend prediction, process noise covariance learning to obtain the optimal gain, and compensation term learning to correct the errors after the filtering update. The gated recurrent unit is used to construct offline learning modules, which endow the multi-level filter with nonlinear model fitting and memory iterative learning capabilities. The proposed algorithm is validated in maneuvering target tracking tasks, showcasing significant enhancements.