目的 聚焦于热毒宁注射液（Reduning Injection，RI）金银花浓缩工序，旨在开发并验证一套数据驱动的技术框架，以实现对“黄金批次”的客观评价、智能识别与工艺优化，并从根本上揭示影响“黄金批次”形成的关键工艺机制与量化控制策略，以期提升生产过程的效率与稳定性。方法 收集了170批金银花浓缩工序生产工艺数据，创新性地构建了“综合浓缩效能比”（E）作为工艺性能评价指标。采用一种潜空间引导的效能阈值自适应划分策略（latent-space guided adaptive performance thresholding，LGAPT），客观地将生产批次划分为黄金批次与非黄金批次。随后，构建了基于极限梯度提升决策树算法（extreme gradient boosting，XGBoost）的分类预测模型，并结合SHAP（Shapley加性解释）与部分依赖图（partial dependence plots，PDP）等模型解释技术，深入挖掘影响批次质量的关键工艺特征及其最优调控区间。最后通过非参数检验、效应量分析及核密度估计对模型发现进行统计学验证。结果 确定黄金批次的E分界阈值为0.136 3 kg/(m³∙min)，据此将170批数据划分为61个黄金批次和109个非黄金批次；XGBoost模型在识别黄金批次方面表现优异（测试集的F1分数为0.84）；模型解释分析识别出LT_std-2、LT_skew-1和T3_abd-2等是影响黄金批次形成的核心特征，并明确了其优化区间。统计验证结果与模型解释高度一致，证实了该结论的可靠性。结论 建立了一套从性能量化、智能划分到模型寻优的完整技术框架，成功识别了影响金银花浓缩工序质量的关键特征及其控制范围；该框架为中药生产过程的标准化与智能化控制提供了有力的理论依据，推动了从“经验判断”到“数据驱动”的工艺优化模式转变。

Objective This study focuses on the Jinyinhua (Lonicerae Japonicae Flos, LJF) concentration process in the production of Reduning Injection (热毒宁注射液, RI), with the objective of developing and validating a data-driven technical framework. The framework is designed to enable objective evaluation, intelligent identification, and process optimization of “golden batches”, while fundamentally elucidating the key process mechanisms and quantitative control strategies influencing their formation. Ultimately, this approach aims to enhance the efficiency and stability of the production process. Methods This study collected production process data for 170 batches of LJF concentrate, and innovatively established “comprehensive concentration efficiency ratio” (E) as an evaluation index for process performance. A latent-space guided adaptive performance thresholding (LGAPT) strategy was employed to objectively classify the production batches into golden and non-golden categories. Subsequently, a classification prediction model based on extreme gradient boosting (XGBoost) was constructed. By integrating model explanation techniques such as SHAP (Shapley additive explanations) and partial dependence plots (PDP), the key process features affecting batch quality and their optimal control ranges were systematically investigated. Finally, the model findings were statistically verified through non-parametric tests, effect size analysis, and kernel density estimation. Results The study established a threshold of E = 0.136 3 kg/(m³∙min), classifying the 170 batches into 61 golden and 109 non-golden batches. The XGBoost model demonstrated excellent performance in identifying golden batches, achieving an F1-score of 0.84 on the test set. Model interpretation identified features such as LT_std-2, LT_skew-1, and T3_abd-2 as core determinants for golden batch formation and defined their optimal ranges. The statistical validation results were highly consistent with the model interpretations, confirming the reliability of the findings. Conclusion This study established a comprehensive technical framework encompassing performance quantification, intelligent classification, and model-based optimization. By successfully identifying the key process features and their optimal control ranges, this work lays a methodological foundation for the standardization and intelligent control of traditional Chinese medicine (TCM) manufacturing, driving a paradigm shift in process optimization from an empirical-based to a data-driven approach.

R283.6

国家工信部产业基础再造和制造业高质量发展专项（TC2308068）；国家科技部长三角科技创新共同体联合攻关项目（2023CSJGG1700）