Gas identification from sulfide all-solid-state batteries: a novel open-set domain generalization method with multi-waveform temperature modulation and generalizable metric learning

Accurate and non-destructive identification of characteristic gases released during early failure of sulphide all-solid-state batteries is critical for assessing their safety risks. However, existing gas identification methods based on electronic noses remain constrained by limited response information, sensor drift and unknown-gas interference, resulting in limited identification performance. Therefore, this paper proposes a novel open-set domain generalisation method (NOSDG) that integrates a multi-waveform temperature modulation strategy (MWTM) with a generalisable metric learning network (GMLN). First, MWTM is used to excite the sensors to acquire dynamic response data rich in discriminative information. Subsequently, within GMLN, a joint sensitivity-driven self-evolving mask mechanism is proposed to separately suppress highly drift-sensitive responses and overly class-discriminative responses, thereby learning features with cross-domain invariance and discriminative robustness. Meanwhile, a boundary feature-driven optimisation strategy is proposed to generate latent boundary features that deviate from known-class feature distributions and employ them as unknown-class references, thereby optimising the decision boundary by contracting the known-class acceptance regions from the outside inward. Consequently, NOSDG can identify known gases and reject unknown gases under sensor drift conditions. Experimental results on a laboratory-collected gas dataset show that NOSDG achieves a mean accuracy of 87.06% and a mean H-score of 85.52%, outperforming the comparison methods.