A Parameter Extraction Net for Wireless Wideband Received Signals Based on UNet

Abstract

Reliable wideband spectrum monitoring requires signal detection, modulation discrimination, and parameter extraction. This paper presents a segmentation-driven hybrid pipeline for synthetic wideband time-frequency representations. Received waveforms are converted into spectrograms, and a task-oriented UNet variant termed HOC-UNet combines high-frequency perception (HFP), OmniAttention (OA), and Coordinate Attention (CA) to predict pixel-level signal masks. Deterministic connected-component analysis and time-frequency coordinate mapping then recover start time, duration, center frequency, and bandwidth. The contribution is not a new attention operator; it is a stage-specific integration in which boundary-sensitive processing is placed in shallow encoder and skip features, coordinate attention reweights the two highest-resolution skips, and an Omni-Dimensional Dynamic Convolution (ODConv)-derived channel/filter attention plug-in refines all four decoder fusion stages. Experiments use a controlled synthetic benchmark with nine modulation types and eight additive white Gaussian noise (AWGN) power settings (−5, 0, 5, 10, 15, 20, 25, and 30 dBW). The reported aggregate binary mean intersection over union (mIoU) is 88.65%, while the reported multi-class mIoU is 52.47%. On successfully matched components at 0 dBW, HOC-UNet obtains start-time, duration, center-frequency, and bandwidth normalized root mean square error (NRMSE) values of 0.00406, 0.00940, 0.00057, and 0.02510 in the binary task and 0.00379, 0.00759, 0.00075, and 0.00707 in the multi-class task, respectively. These conditional errors do not include missed detections or false alarms. The conclusions are, therefore, limited to the constructed synthetic AWGN benchmark; measured channels, fading, interference, frequency offset, and hardware impairments remain to be validated.