Dimensionality reduction of passive impact-localization features in metallic and stiffened composite plates

Passive impact monitoring with bonded piezoelectric sensor networks produces feature vectors far larger than the information available for localizing the impact, yet reduction strategies have never been compared under controlled conditions for continuous coordinate regression. Six supervised and six unsupervised methods are benchmarked here on two specimens sharing the same eight-channel layout: an isotropic aluminum 7075-T6 plate and a quasi-isotropic AS4/8552 panel with an adhesively bonded stiffener. Dense 21x21 automated impact campaigns yield 14,846 and 53,361 labeled events, from which 233 descriptors are extracted per event. Each representation is evaluated over twenty retained dimensionalities and ten folds with a fixed multilayer perceptron, giving 2,400 models per specimen against a full-dimensional baseline. On the aluminum plate supervised selection dominates: recursive feature elimination with fifteen features lowers the mean Euclidean error by 11.2% to 12.1 mm, close to the 10 mm floor imposed by the acquisition rate. On the stiffened panel the ranking collapses: the families interleave, and no reduced configuration beats the 25.7 mm baseline, the best reaching only 27.2 mm while still retaining 95 features. Correlation analysis attributes the reversal to information structure: localization information is concentrated in twelve strongly correlated arrival-time descriptors in the isotropic plate but dispersed across many weak ones once anisotropy and stiffener scattering are present. The benefit of dimensionality reduction is therefore governed by the concentration of the feature-target correlation structure rather than by the algorithm.