Robust geometric preprocessing reveals architecture-conditional spectral collapse in heterogeneous fleet anomaly detection

Deploying pre-trained audio models for anomaly detection in heterogeneous industrial fleets induces geometric pathologies in the representation space. This study diagnoses the architecture-conditional nature of Spectral Collapse (SC) and evaluates a robust preprocessing protocol—the Geometric Approximation Protocol (GAP)—comprising coordinate-wise median centering, interquartile range (IQR) normalization, and whitening principal component analysis (PCA). The evaluation spans 14,400 trials per condition on the MIMII/DCASE benchmark (16 physical machines \(×\) 3 SNR levels, 12 pre-trained architectures, 5 anomaly detectors), validated by 10,000-iteration permutation tests with Benjamini–Hochberg correction (91.4% retain significance; \(n = 48\) machine-SNR analytical units). Transformer-based models exhibit geometric stability (effective linear dimensionality 97–114), whereas convolutional architectures undergo Spectral Energy Concentration under the Robust condition (ELD: 294 \(→\) 9.5)—a Spectral Collapse pattern consistent with IQR-based suppression of outlier-driven variance rather than heterogeneity alone. GAP yields negligible practical effect in 96.7% of model \(×\) detector combinations (median Cohen’s \(d = -0.004\) , IQR \([-0.011, +\,0.003]\) ); practically meaningful improvement is confined to CNN architectures (PANNs, YAMNet) paired with OC-SVM. Transformer and SSL architectures show no practically meaningful preprocessing effect. Within this single-benchmark scope, these results reframe GAP not as a universal normalization solution but as a selective geometric diagnostic layer whose benefit is architecture-conditional.