Multiple features fusion and mixup with conditional decoder for

In recent years, the combination of vibration spectral data and data-driven methods has dominated the development and application of close spectral recognition. Nevertheless, in practical applications, open spectral categories (i.e., novel/unknown spectral categories) may be encountered, as collecting comprehend-sive categories is time-consuming and requires professional expertise. The intuitive solution is to obscure features of different categories, but relevant exploratory experiments yield unsatisfactory open-set performance, which may be attributed to sparse spectral features and high inter-class similarity. To remedy this issue, we innovatively propose an end-to-end scheme combining Multiple Features Fusion and Mixup with Conditional Decoder (MFFMCD) in this paper. In particular, to enhance feature representation, MFFMCD adopts two auxiliary feature extraction modules and fuses different branch features. Additionally, to cope with high inter-class similarity, the enhanced features are obscured within a mini-batch and restored to corresponding class samples through a conditional decoder to mimic the feature distribution of unknown classes. Experiments on three publicly available spectral datasets show that the proposed MFFMCD significantly outperforms existing methods. In the end, extensive ablation studies are conducted to investigate the effectiveness, correctness, and robustness of our proposal.