Answers are grounded in this paper and cite their sources.
This article investigates real-time state estimation on a state-space model with unknown noise statistics. The estimation problem is formulated as a noise covariance parameter learning problem, aiming to learn suitable noise covariance matrices from data through deep neural networks. In general, this formulation suffers from catastrophic forgetting, where learning noise statistics from current data usually results in a dramatic performance degradation on previously learned settings. We present lifelong learning Kalman filter (LLKF), a deep learning-based Kalman filter that continually learns noise covariances from data to support filtering. To achieve continual learning of noise statistics in LLKF, we derive a loss function to provide a training paradigm without catastrophic forgetting. This bridges the gap between postlearning filtering and lifelong filtering that allows the estimator to acquire, update, accumulate, and exploit knowledge throughout its lifetime. Both numerical simulations and real-world experiments illustrate the effectiveness and continual learning performance of the proposed LLKF.
No takes yet. Share an insight, caveat, or question.
Lin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: