Ideal time series classification should capture invariant representations, but achieving reliable performance on out-of-distribution (OOD) data remains a key challenge. This is because models often entangle domain-specific features with label-related features, leading to erroneous correlations. Current feature separation methods lack the semantic direction needed to truly separate universal features due to their lack of guidance. To address this issue, we propose the Energy-Regularized Information for Shift-Robustness (ERIS) framework, which enables guided and reliable feature separation. ERIS achieves this goal through an energy-guided correction mechanism, a weight-level orthogonality strategy, and an auxiliary adversarial generalization mechanism. Experimental results on four benchmarks demonstrate that ERIS consistently achieves state-of-the-art performance, achieving statistically significant improvements over state-of-the-art baselines.
