Spectral Shortcut Learning and Generalization Degradation in 2D-CNN Spectrogram Classifiers Trained from Scratch on Procedural Synthetic Contaminated Data

Applying 2D Convolutional Neural Networks (2D-CNNs) for acoustic leak detection in fluid-transport pipelines is fundamentally hindered by the acute shortage of field positive leak recordings. An increasingly common practice among NDT practitioners is the rapid generation of surrogate acoustic WAV files using heuristic scripts produced via Large Language Models (LLMs). In this work, employing a rigorous multi-seed statistical benchmark (10 independent Random Seeds, 40 complete training-evaluation cycles), we systematically examine the generalization dynamics of 2D-CNN spectrogram classifiers under progressive contamination by LLM-assisted procedural synthetic data (ε ∈ [0.0, 0.20]). Validation is executed on a clean held-out physical validation corpus (N=92, 37 leak targets, 55 normal pipes). We establish that synthetic LLM-generated surrogate injection triggers Spectral Shortcut Learning and an acute Recall Collapse: at ε = 20%, physical leak detection Recall collapses from 94.05 ± 6.37% down to 66.76 ± 12.33%, while missed physical leaks surge 5.6-fold. Convolutional filters selectively overfit to idealized high-contrast Gaussian bursts produced by the generator, ignoring the broadband stochastic signatures of actual fluid discharge.