NIR–MIR Data Fusion for Rapid Oil-Product Detection

The patent combines NIR and MIR spectrum technologies into a unified detection system. The NIR spectrum provides information about hydrogen-containing groups (C-H, O-H, N-H bonds) while the MIR spectrum provides fundamental frequency information of complementary groups. By merging both spectral data sources, the system achieves comprehensive chemical information retrieval without sacrificing detection speed, as both spectra can be collected and processed concurrently.

The detection system is designed with multi-functionality by integrating both NIR and MIR spectral analysis capabilities. This universal approach allows the same detection platform to access multiple types of chemical information simultaneously - both fundamental frequency vibrations and combination frequency information - making the system adaptable to various detection requirements for different oil products.

The patent transitions from single-spectrum prediction to multi-spectral data fusion prediction by adding another dimension of information. The detection model processes both NIR and MIR spectral data simultaneously, creating a two-dimensional feature space that captures more comprehensive chemical characteristics. This dimensional expansion enables the model to achieve superior prediction accuracy for detection indexes while maintaining manageable complexity through structured data fusion approaches.

The detection model functions as a composite system that integrates multiple spectral data sources. By treating NIR and MIR spectra as complementary components, the model creates a composite prediction framework that leverages the strengths of both spectral regions. The chemometric model processes fused spectral information to produce enhanced prediction results for oil product detection indexes.

The patent applies segmentation to the data processing workflow by dividing the multi-spectral processing into distinct stages: NIR spectrum processing, MIR spectrum processing, and data fusion processing. Each stage can be independently optimized and validated. The spectrum preprocessing, variable selection, and model training are segmented into manageable modules that handle complexity systematically while maintaining high prediction reliability.

The patent performs preliminary actions by conducting spectrum preprocessing (including derivative processing and vector normalization) and variable selection (using CARS and VIP methods) before the final data fusion and model training stages. These preliminary processing steps reduce data complexity and highlight important features, making the subsequent fusion and prediction processes more manageable and reliable.