Seeing Sound: A Computer Vision Approach to Ultrasonic Leak Detection in Industrial Pipelines

This paper presents an end-to-end computer vision pipeline for ultrasonic leak detection in pressurized industrial pipelines. Raw high-frequency acoustic signals from Kaskad-3 dual sensors are transformed into Mel-Spectrograms using STFT (n_fft=2048, hop_length=512, n_mels=128) and classified using a ResNet-18 CNN trained on over 60,000 field audio segments. A hybrid decision system combining CNN inference with Phase Coherence analysis, Peak Sharpness, and Spatial Dispersion metrics reliably rejects distributed noise sources (including boiling groundwater on district heating pipes at 130–150°C) that confuse traditional correlation algorithms. The hybrid system achieved ROC-AUC of 0.92, Accuracy of 94.2%, and reduced the false positive rate by 68% compared to legacy heuristics. Ablation analysis confirms that Phase Coherence integration was specifically responsible for an additional 42% reduction in false positive excavations on boiling water scenarios. An open-source demo is available at https://github.com/rivixi/ultrasonic-cv-demo and an interactive engine at https://uzk.rivixi.ai