Title: Merging Statistical Metrics and Deep Encodings for Rehearsal-Driven Class-Incremental Time Series Classification

Abstract: Real-world systems frequently need to integrate new categories and information without causing the classification model to forget previously acquired knowledge. This challenge, known as class-incremental continual learning, becomes particularly complex when dealing with multivariate time series due to their inherent temporal structures. To address this, we introduce a novel methodology for class-incremental continual learning tailored to multivariate time series classification. Our approach relies on a dual-stream feature extraction pipeline that leverages both deep temporal embeddings—generated through a pre-trained, frozen foundation model—and statistical features. Testing across five benchmark datasets demonstrates that our proposed system delivers competitive average accuracy while effectively minimizing forgetting rates in all experimental setups.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC