Signal as Point: Deep Learning Signal Detector on Time Domain

Precise signal detection on time domain is a vital task in the information processing of modern wireless systems, but challenges persist in accuracy. Conventional time-domain detection with Spectrum Sensing (SS) can only identify the presence of a signal without pinpointing its exact start and end. Other methods employ Short-Time Fourier Transform (STFT) to transform samples into spectrograms, followed by deep-learning-based object detectors for signal detection, they inherently present flaws mapping back to the time domain. Our study proposes a novel time-domain signal detection approach by modeling signals as center points and durations. Our detector can precisely and directly locate signals in I/Q sampling data without prior knowledge of channel conditions or time-frequency transformation. Our method particularly excels under low signal-to-noise ratio (SNR) conditions.