Infrared spectroscopy (IR) serves as a cornerstone in analytical chemistry for molecular characterization, yet its machine learning (ML) applications are hindered by high-dimensional spectral data, nonlinear interactions, and imbalanced datasets. To address these challenges, this study proposes the Integrated Data Sampling, Augmentation, and CNN Framework (IDACF), which synergizes conditional tabular generative adversarial networks (CTGAN), adversarial training, and adaptive resampling to enhance spectral data quality and model generalizability. The framework introduces three innovations: (1) a conditional GAN-based mechanism for chemically meaningful spectral augmentation, (2) a multi-index evaluation system (Pearson correlation coefficient, structural similarity index, Chebyshev distance) to validate synthetic data fidelity, and (3) a hierarchical CNN architecture with dual-path feature extraction for capturing multi-scale stoichiometric patterns. Evaluated on three IR datasets—pharmaceutical tablet ingredient detection, licorice product quality assessment, and red wine composition analysis—IDACF demonstrates transformative performance gains. The CNN model achieves accuracy improvements of 18.66% (tablet), 44.17% (licorice), and 32.59% (wine), reaching up to 99.63% and 100% accuracy on augmented datasets. Comparative experiments across SVM, BPNN, and PLS-DA models further highlight IDACF’s robustness in mitigating data scarcity and class imbalance. This work establishes a scalable paradigm for chemometric analysis, with broad applicability in industrial quality control and spectroscopic diagnostics.