Multi-Expert Feature Fusion for Signal Out-of-Distribution Detection

Abstract

Out-of-distribution (OOD) detection is critical for the reliable deployment of radio-frequency (RF) signal recognition systems, as novel signals may severely degrade performance. We propose a multi-expert framework that integrates time-, frequency-, and time-frequency-domain experts to construct a richer feature space. Within this space, a normalized Mahalanobis distance is employed to jointly assess feature similarity and cross-view consistency. This design enables more robust separation between known and unknown signals compared with single-domain approaches. Experiments on both a synthetic Automatic Modulation Recognition (AMR) dataset and a real-world Specific Emitter Identification (SEI) dataset show that our method significantly outperforms comparison approaches in OOD detection while maintaining high closed-set classification accuracy.