Evaluating Transferability and Validation Protocols of Convolutional Neural Networks for Acoustic Water Leak Detection Using SebaKMT Logger Recordings: Code, Manifests, and Predictions

This preprint evaluates direct transfer, supervised target adaptation, and validation protocol design of convolutional neural networks for acoustic water leak detection using field recordings from SebaKMT correlation loggers. A single-scale CNN (25,633 parameters) is compared with a three-branch multiscale time–frequency network (TFCNN, 76,769 parameters) and classical machine learning baselines across multiple experimental protocols. In direct transfer from an unstructured multi-instrument collection, multiscale TFCNN increased sensitivity from 78.20% to 86.40% compared to the single-scale network, but exhibited low specificity (16.50%). While supervised adaptation of the output classifier improved test score ranking (ROC-AUC 94.16%), threshold selection subject to high validation recall retained 20 alarms among 40 negative field recordings. When trained from scratch on 180 SebaKMT recordings, random file partitioning yielded 100% accuracy in all ten runs, which is shown to be inflated by shared recording sessions between training and test sets. Under site-disjoint grouped cross-validation across six registered site groups, TFCNN out-of-fold accuracy was 87.83 ± 3.22% (mean within-fold ROC-AUC 94.28 ± 3.87%). A classical logistic regression baseline using nine handcrafted signal descriptors without digital amplitude achieved an accuracy similar to TFCNN (87.78% vs. 87.83%), with higher sensitivity (98.33% vs. 92.67%) and lower specificity (82.50% vs. 85.42%), demonstrating no detectable accuracy advantage for the deep architecture on this corpus. Furthermore, logistic regression on digital RMS alone reached 94.44% accuracy at threshold 0.50 (within-fold ROC-AUC 99.67 ± 0.58%), although its sensitivity was lower at 83.33%. This strong predictive association motivates investigation of acquisition-related shortcuts rather than confirming a physical acoustic mechanism. The study concludes that acoustic diagnostics on small field datasets require strict site-disjoint validation protocols and transparent classical feature baselines alongside deep neural networks to prevent session leakage and shortcut learning.