A Deep Time–Frequency Pattern Recognition Framework for Broadband Wireless Signal Detection

Broadband spectrum monitoring requires coexisting signals to be detected, localized, and classified simultaneously. Representing an in-phase/quadrature (I/Q) capture as a time–frequency image turns this task into a multimodal recognition problem. Classical detectors return only a binary occupancy decision, whereas deep classifiers return one label per input; neither reports the class, extent, and reliability of every coexisting signal. This paper formulates the task as object detection in the time–frequency plane. An STFT front-end, with its parameters selected through a sensitivity analysis, feeds a YOLOv8n detector that predicts the class and bounding box of every signal in one pass. The detector is standard; the contributions are the task formulation, the representation-selection procedure, and a validation covering STFT sensitivity, ablation, detector and task baselines, channel impairments, low SNR, cross-validation, and calibration. On a 1,250-image test set, the pipeline reaches [email protected] = 0.540; on a separate larger campaign, it achieves [email protected] = 0.772 and a mean localization IoU of 0.9234. Its F1-optimal operating point yields precision 0.70 and recall 0.90. At 10 dB, multipath fading changes object-level recall by at most 0.03 relative to AWGN, while impulsive and harmonic interference are the dominant failure modes. Isotonic recalibration reduces the expected calibration error (ECE) from 0.225 to 0.050. Because the larger campaign also differs in STFT and training settings, the scale result is reported as an indication rather than an isolated estimate; at the reduced scale, a detector with 3.7 times as many parameters brings no accuracy gain. The framework returns the class, extent, and confidence of every signal, together with a recalibration procedure that supports autonomous spectrum management.

Authors

Publication Details

Journal
International Journal of Pattern Recognition and Artificial Intelligence
Published
2026-11-19
DOI
https://doi.org/10.1142/s0218001426400811
Primary Topic
Wireless Signal Modulation Classification
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Deep Time–Frequency Pattern Recognition Framework for Broadband Wireless Signal Detection

Boyuan Liu, Yuegang Song
International Journal of Pattern Recognition and Artificial Intelligence
Wireless Signal Modulation Classification
article

A Deep Time–Frequency Pattern Recognition Framework for Broadband Wireless Signal Detection

Boyuan Liu, Yuegang Song
article en

Abstract

Broadband spectrum monitoring requires coexisting signals to be detected, localized, and classified simultaneously. Representing an in-phase/quadrature (I/Q) capture as a time–frequency image turns this task into a multimodal recognition problem. Classical detectors return only a binary occupancy decision, whereas deep classifiers return one label per input; neither reports the class, extent, and reliability of every coexisting signal. This paper formulates the task as object detection in the time–frequency plane. An STFT front-end, with its parameters selected through a sensitivity analysis, feeds a YOLOv8n detector that predicts the class and bounding box of every signal in one pass. The detector is standard; the contributions are the task formulation, the representation-selection procedure, and a validation covering STFT sensitivity, ablation, detector and task baselines, channel impairments, low SNR, cross-validation, and calibration. On a 1,250-image test set, the pipeline reaches [email protected] = 0.540; on a separate larger campaign, it achieves [email protected] = 0.772 and a mean localization IoU of 0.9234. Its F1-optimal operating point yields precision 0.70 and recall 0.90. At 10 dB, multipath fading changes object-level recall by at most 0.03 relative to AWGN, while impulsive and harmonic interference are the dominant failure modes. Isotonic recalibration reduces the expected calibration error (ECE) from 0.225 to 0.050. Because the larger campaign also differs in STFT and training settings, the scale result is reported as an indication rather than an isolated estimate; at the reduced scale, a detector with 3.7 times as many parameters brings no accuracy gain. The framework returns the class, extent, and confidence of every signal, together with a recalibration procedure that supports autonomous spectrum management.

International Journal of Pattern Recognition and Artificial Intelligence
Wireless Signal Modulation Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon