An effective NIR analysis tool: CNN assisted by novel spectrogram reconstruction technology based on recurrence plot

Near-infrared (NIR) spectroscopy has become a widely utilized analytical technique for the rapid quality assessment of products in petrochemical and other process industries. Deep learning has garnered increasing attention in industrial NIR detection, primarily attributed to its flexibility and broad adaptability. However, existing spectroscopic studies have predominantly concentrated on creating 1D deep learning models tailored to 1D spectral signals, neglecting the potential to unleash the powerful 2D image analysis capabilities of deep learning. This paper discussed recent 2D CNN model and presented a new viewpoint of NIR 2D spectrogram reconstruction using Recurrence plot (RP) derived from time series mathematical transformations. The proposed NIR spectrogram reconstruction technology was assessed in real NIR scenarios that comprise classification and prediction. Empirical analysis involving CNN models based on RP spectrograms and 1D NIR signals, alongside other machine learning approaches, demonstrates that CNNs augmented with RP spectrogram reconstruction significantly enhance performance in both classification and prediction tasks. With its distinctive advantages of novelty, simplicity, and precision, 2D spectrogram reconstruction technology proves to be an effective enhancement for CNN-based NIR analysis in real-world industrial scenarios.