The ability to detect, classify, and characterize radio signal transmissions is an important task with diverse applications in defense, networking, and communications. This task is made complex when multiple simultaneous signals may be present. Finding useful signal features can go beyond just classifying the signal modulation type, to include estimates of: the signal-to-noise ratio (SNR), the confidence in the classification, and the presence of multiple signals. Such tasks have traditionally been done by human analysts, often relying on ad-hoc techniques. Recent advances have shown that applying deep learning methods to radio signal modulation classification tasks is feasible. But current recognition-only networks are limited in several ways. They can classify modulation for just one signal, produce limited feature outputs, and may give over-confident results in noisy environments. The researchers built a robust radio frequency signal classifier with multiple outputs, providing not just modulation classification, but also SNR estimation, classification confidence, and multiple signal detection (even when these signals are located at the same transmit frequency). Compared to the state-of-the-art deep learning classifier in the single signal case, our system obtains better accuracy (up to 91.6% accuracy at 20 dB SNR), while providing additional useful outputs.
