Topological AI-Analysis vs. Classical Cross-Correlation: Overcoming Legacy Defectoscope Vulnerabilities

This research paper presents a systematic audit of classical cross-correlation DSP algorithms for acoustic pipeline leak detection, evaluated on 371 field recordings (215 normal, 156 leaks). We demonstrate that conventional correlators suffer a critical specificity collapse at sensor distances exceeding 100 meters, generating 100% False Positive Rate due to probabilistic noise alignment within widened max_tau search windows. To resolve this scalability barrier, we developed and validated a hybrid topological AI classifier combining 1D-CNN (Acoustic1DNet) for time-domain waveform analysis and 2D-CNN (Acoustic2DNet) for Mel-Spectrogram frequency ridge detection. The hybrid architecture achieves distance-invariant 97.7% specificity (95% CI: [94.7%, 99.0%]) and 97.4% sensitivity (95% CI: [93.6%, 99.0%]) across all pipe spans from 10m to 250m, eliminating the spatial constraints of static DSP frameworks.