ReSeleCT: A New Approach for Continual Learning With Application to Vehicle State Estimation

In time series and data stream analysis, neural networks (NNs) have demonstrated remarkable capacity in predicting current and future states. However, NNs often suffer from catastrophic forgetting (CF) when adapting to new tasks or data domains. This issue is particularly pressing in vehicle state estimation, given that new data domains are frequently invited to a pre-trained model. While memory-based techniques have been proposed to address CF in continual learning, they are less effective for time series regression problems due to the absence of a suitable subset selection strategy. This paper presents a novel method called ReSeleCT (Representative Selection for Continual learning in Time-series scenarios) for identifying and capturing a coreset of old datasets in memory-based continual learning with application to vehicle velocity estimation. The approach focuses on selecting a representative subset of historical data to retain key information from previous tasks. The framework is applied to estimate a vehicle's longitudinal and lateral velocities using neural networks, incorporating new maneuvers into the previously trained model. Experiments on sensor data from an electric Equinox vehicle demonstrate that ReSeleCT efficiently adapts to new data domains while preserving prediction accuracy on the old dataset, with the added benefit of fast model adaptation without significant computational overhead. Comparison against other algorithms in continual learning shows a superior performance of the proposed method in terms of prediction error while maintaining an acceptable training time.