Integration of independently preprocessed spectra to improve classification of repeatedly fried durian chips using portable near-infrared spectrometer and machine learning

Durian chips fried in repeated-use oil can be harmful to health, as prolonged use of the oil leads to excessively high levels of total polar compounds. This study explored the use of a portable near-infrared (NIR) spectrometer combined with machine learning techniques to classify chips into two groups based on frying duration. An alternative fusion technique was proposed by integrating various independently preprocessed spectral data before model analysis. Binary classification was performed to distinguish between chips fried in oil used for 48 hours or less and those fried in oil used for more than that duration. Based on the results, the model developed using partial least squares discriminant analysis with fusion-preprocessed spectral data achieved superior classification performance, with a correlation coefficient of prediction (rp) of 0.810 and a root mean square error of prediction (RMSEP) of 0.294, compared to the model using traditional preprocessed spectral data (rp of 0.791 and RMSEP of 0.310). Subsequently, the fusion-preprocessed spectral data were used to develop a support vector machine (SVM) model, which demonstrated strong performance, achieving a precision of 91.9% in minimizing false positives and reducing the misclassification of chips fried in oil for a longer duration compared to those fried for a shorter duration. However, the SVM model lacked robustness when applied to different datasets. When trained and tested on the same dataset, the classification performance was highest for distinguishing between chips fried in oil used for 36 h or less and those fried in oil used for more than 36 h, achieving a precision of 97.8% on the test set. These findings suggested that the combination of a portable NIR spectrometer and the fusion preprocessing technique held promise for rapid, on-site assessment of chips fried in oil used for up to 36 h. Nonetheless, further studies should be undertaken to enhance the robustness and generalizability of the SVM model.