Combination of markov transition field and multi-scale feature extraction for cetane number prediction in diesel using near-infrared spectroscopy.

Diesel, a crucial energy source for transportation and industrial applications. The cetane number (CN) plays a key role in assessing diesel quality and combustion performance. In this study, near-infrared spectroscopy (NIRS) data of diesel were encoded into images using the Markov transition field (MTF). A hybrid CNN-BiLSTM model with multi-head attention was developed to achieve multimodal feature extraction and fusion by simultaneously inputting spectral data and MTF images, enabling accurate CN prediction. For comparison, PLSR, 1D-CNN, and 2D-CNN models were established using feature wavelengths selected by variable combination population analysis (VCPA), SG1-preprocessed full spectra, and MTF images, respectively. Results demonstrated that the CNN-BiLSTM achieved the best performance (Rp2 = 0.9824, RMSEP = 0.3106), whereas the VCPA-PLSR performed the worst (Rp2 = 0.7541, RMSEP = 1.6425). The 1D-CNN outperformed the VCPA-PLSR, and the 2D-CNN further improved upon the 1D-CNN. This study demonstrates that converting NIRS data into images via MTF effectively leverages the potential of CNNs and improves CN prediction performance without feature wavelength extraction, while the introduced BiLSTM channel mitigates the distortion of original data caused by MTF encoding. The approach provides a non-destructive and reliable framework for liquid samples' quality evaluation in industrial applications.