Abstract:Background: The extraction process is a critical unit operation in the manufacturing of Traditional Chinese Medicine (TCM), which concentrates active pharmaceutical ingredients (APIs) and profoundly impacts downstream processes and final product quality. However, the inherent variability of raw medicinal materials and fluctuations in operating conditions pose significant challenges to achieving batch-to-batch consistency. The traditional quality control model, which relies on time-consuming and destructive off-line testing, cannot provide the real-time feedback required for effective process control and optimization. This study aims to develop a comprehensive on-line monitoring framework based on the principles of Process Analytical Technology (PAT) to address this gap. Objective: Focusing on the extraction process of Senecio cannabifolius, this study aims to establish an on-line monitoring scheme based on near-infrared (NIR) spectroscopy combined with chemometric methods. This framework is dedicated to achieving three core objectives: rapid quantification of multiple active pharmaceutical ingredients, real-time determination of the optimal extraction endpoint, and robust process monitoring and fault detection. Methods: An in-situ transflectance fiber optic probe was installed directly into the extraction tank to acquire spectra every 60 seconds. For quantitative analysis, a Partial Least Squares Regression (PLSR) model was developed to predict the concentrations of six marker compounds (Protocatechuic acid, Cryptochlorogenic acid, Hyperoside, Isoquercitrin, Neochlorogenic acid, and Dicaffeoylquinic acid). The spectra were preprocessed by Savitzky-Golay (SG) smoothing and first derivative, with the 5349–6749 cm?1 band selected for modeling. The Moving Block Standard Deviation (MBSD) method was used to analyze continuous spectral data for real-time determination of the extraction endpoint. Concurrently, a Multivariate Statistical Process Control (MSPC) model based on Principal Component Analysis (PCA) was established using data from multiple normal operating condition (NOC) batches. The model was used to monitor and evaluate validation batches using Principal Component scores, Hotelling's T2, and Distance to Model X (DModX) control charts. Results: The established PLSR model exhibited excellent correlation and predictive accuracy compared to the HPLC reference values. The MBSD method determined the extraction endpoint at approximately 90 minutes, which was highly consistent with the time at which the concentration of each API reached its plateau as measured by off-line analysis. The MSPC model demonstrated high efficiency in process monitoring: normal batches consistently remained within the established control limits, confirming process stability. More importantly, the model showed high sensitivity to process anomalies: species substitution was rapidly identified by a significant increase in the DModX value; deviations in the solid-to-liquid ratio were reflected in a shift in the PC1 score trajectory; and a heating interruption caused an instantaneous increase in the Hotelling's T2 value. Under the tested conditions, changes in soaking time did not trigger an alarm, indicating their limited impact on the overall process. Conclusion: The process monitoring strategy proposed in this paper successfully integrates rapid quantitative analysis, qualitative endpoint determination, and multi-scenario fault detection for a complex TCM extraction process. This method enhances real-time process understanding, improves the timeliness of endpoint decision-making, and helps to reduce quality fluctuations, providing a methodological reference for advancing the intelligent manufacturing of TCM and ensuring the quality consistency of medicinal materials.