Time Series Class-Incremental Learning with Residual Mamba Encoder

Class-incremental learning aims to enable systems to adapt to new classes over time without forgetting existing knowledge, a critical challenge in dynamic, non-stationary environments such as time series classification. While significant progress has been made in class-incremental learning for images and text, time series data remains underexplored. In this paper, we investigate the effectiveness of the Mamba model, based on Structured State Space Models (SSMs), in time series class-incremental learning. We utilize a residual Mamba encoder and compare it with mainstream 1D-CNN and PathTST’s TST encoders across various class-incremental learning algorithms. To our knowledge, this is the first application of the Mamba model to time series class-incremental learning. Extensive experiments demonstrate that the residual Mamba encoder achieves comparable or superior performance and computational efficiency, particularly on time-series-specific algorithms.