Partial least squares (PLS) regression coupled with near-infrared (NIR) spectroscopy is central to real-time pharmaceutical quality monitoring under the Process Analytical Technology (PAT) framework. While Hotelling T² and Q residuals are routinely used to monitor deployed models, the spectral regions and perturbation types that drive exceedances in each statistic are not directly provided by regression coefficients or variable importance in projection (VIP) scores.,This paper presents a framework in which a variational autoencoder (VAE) generates a model-consistent synthetic population for systematic perturbation experiments. Applied to the IDRC 2002 shootout NIR dataset, 50,000 synthetic spectra were filtered to a 350-spectrum model-consistent subset using T² and Q thresholds. Three perturbation types were applied: Gaussian peak injection, scatter and baseline variation, and wavelength axis shifts. Mean T², Q residual responses, and exceedance rates were mapped across the full spectral range (742.9–1739.8 nm).,The 1150–1220 nm and 1450–1600 nm regions showed consistently elevated sensitivity across all perturbation types. A 1 nm global wavelength shift increased mean T² by approximately 25 units, with localized sensitivity at 1150 nm and 1550 nm. Terminal regions below 900 nm and above 1650 nm were insensitive throughout. The framework is directly transferable to any PLS deployment using T² and Q monitoring and is not specific to the dataset or instrument used here.