Traditional offline monitoring of Chinese hamster ovary cell fermentation processes suffers from severe systemic time delays. To achieve real-time monitoring of key physiological and biochemical parameters, this study proposes a non-invasive soft sensing framework based on cellular micro-morphology. Phase-contrast microscopic images were acquired throughout the fed-batch cultivation cycle, systematically extracting multidimensional features at the single-cell level, including geometric dimensions, shape, and internal optical texture. Given the highly non-linear mapping relationship between microscopic phenotypes and macroscopic metabolism revealed by univariate analysis \( \left( {\left| r \right| < 0.5} \right) \) , a multivariate predictive model was constructed using the Random Forest (RF) algorithm, which was evaluated in parallel with a Partial Least Squares Regression model. Blind testing using an independent test batch subjected to an abnormal temperature disturbance demonstrated that the RF model effectively overcame the limitations of linear algorithms, achieving high-precision, cross-batch predictions for viable cell density (R2 = 0.86), glucose (R2 = 0.77), and lactate (R2 = 0.90). Furthermore, from the perspective of model interpretability, feature analysis indicated that standard deviation features, representing the morphological heterogeneity of the cell population, showed a higher predictive contribution in characterizing the evolution of metabolic states than mean features. This study suggests the tremendous potential of microscopic population morphological heterogeneity as a digital biomarker for bioprocess monitoring, providing a reliable data-driven strategy for the at-line evaluation of industrial bioreactors.
