Open-Metric for Unknown Signal Inference

The closed-set assumption that only known classes can be classified during testing is usually adopted in traditional signal recognition missions. It has been thoroughly explored and made huge progress in the past decade. However, it cannot cope with the unknown categories coming from the changing environment. As waveform diversity technology evolves by leaps and bounds, the existence of novel unknown classes is inevitable in the open-world environment, imposing a huge challenge to the noncooperative electronic support system. To further satisfy the demands of modern electronic intelligence systems and sense new threats, it is urgent to find an approach that can tackle unknown signals. An open-metric strategy is proposed in this article to deduce the unknown classes from the intercepted signals. By virtue of metric learning methods, a plain backbone can be trained to offer more discriminative representations and a consistent proxy for each class. According to extreme value analysis, information about known classes can be converted into an estimated score of unknown classes, creating an open decision layer. The investigated open-metric approach demonstrates its superiority in unknown signal inference tasks, facilitating improvements in the intelligence and smartness of the electronic reconnaissance agents to cope with future cognitive electronic warfare.