Interpretable machine learning for assesting environmental and mounting effects on aeronautical glass wool absorption

Fibrous absorbers installed in aircraft fuselage linings are exposed to non-nominal mounting conditions and humidity variations that may alter their sound absorption performance. Although glass wool packages are commonly protected by surface layers, the combined influence of protective layers, installation-induced compression and moisture exposure remains only partially quantified under controlled impedance-tube conditions. This work investigates the acoustic variability of two aeronautical glass wool materials through sound absorption coefficient measurements performed over 250 Hz to 4250 Hz. Two experimental datasets are considered: the first examines material type, protective layer, mounting configuration and relative humidity, whereas the second addresses acoustic changes induced by repeated humidity cycles. Each factor is analyzed independently through a univariate neural-network classifier trained on pre-processed absorption curves. Classification accuracy is used as a qualitative indicator of feature detectability, while SHAP values identify the frequency regions responsible for the classification outcome. Results show that protective layers, material type, relative humidity and humidity cycles leave clear acoustic signatures, with univariate accuracies up to 99.5%, 96.0%, 69.6% and 82.5%, respectively. Installation effects are weaker but remain detectable above the random baseline. Humidity-cycle exposure produces progressive acoustic changes, supporting the relevance of moisture-driven degradation mechanisms in aircraft insulation systems.