Real‐Time Evaluation and Maintenance Method of Online Near‐Infrared Detection Model

Near‐infrared spectroscopy analysis has been widely applied in online detection due to its advantages of being rapid, nondestructive, and environmentally friendly. In such applications, the stability and accuracy of the detection models are critical. However, online NIR models are often affected by environmental fluctuations, instrument drift, and variations in sample characteristics. To address these issues, this study proposes a real‐time evaluation and maintenance method for online NIR detection models based on correlation analysis. The method analyzes the correlation between predictions from the offline and online models to assess the online model’s performance. Based on the correlation results, slope‐intercept correction is then applied to improve model accuracy. A total of 1187 samples collected from online NIR instruments across three periods between 2022 and 2023 by Yunnan Tobacco Company were used, along with their corresponding offline reference data. The correlation coefficients between online and offline predictions for total sugar and nicotine before correction were 0.867, 0.949, 0.768 and 0.942, 0.981, 0.976, respectively. After applying slope‐intercept correction, the mean absolute error, root mean square error, and mean absolute percentage error for total sugar decreased from 1.205%, 1.526%, and 4.28% to 1.063%, 1.345%, and 3.75%, respectively. For nicotine, the same metrics were reduced from 0.148%, 0.188%, and 5.23% to 0.135%, 0.171%, and 4.81%, respectively. The results demonstrate that the proposed method effectively reduces discrepancies between online and offline models, enhances model evaluation and maintenance, and offers a viable solution for ensuring the long‐term stability of NIR online detection systems.