Abstract
An acoustic leak-detection framework is presented for industrial ventilation ducts by combining physically motivated spectrogram remapping with attention-guided object detection. Small, pressurized leaks generate broadband turbulent-jet noise, whereas blowers, airflow fluctuations, and structural vibration dominate the lower-frequency field. Preliminary spectra from the test duct showed repeatable leakage energy from 10 to 20 kHz, so this band was expanded during short-time Fourier transformation while low-frequency and out-of-band components were compressed. The remapped spectrograms were processed by a YOLO v5 detector with a Squeeze-and-Excitation channel-attention module. A field dataset of 4780 paired linear and nonlinear spectrograms was collected from a 320-m duct section at the China Jinping Underground Laboratory under three pressures and two aperture sizes. Matched data partitions and ablation tests were used to isolate the effects of the acoustic representation and network architecture. Across the mixed-condition test set, the proposed method achieved 97.1% precision, 95.6% recall, and 95.2% mAP@0.5:0.95, improving precision by 2.7% points over the corresponding linear-spectrogram baseline.