机构地区:[1]华南理工大学化学与化工学院,广东广州510640 [2]无限极(中国)有限公司,广东广州510623 [3]晶格码(青岛)智能科技有限公司,山东青岛266109 [4]北京石油化工学院化学工程学院,恩泽生物质精细化工北京市重点实验室,制药和结晶系统工程中心,北京102617
出 处:《化工进展》2020年第8期3263-3272,共10页Chemical Industry and Engineering Progress
基 金:无限极(中国)有限公司的资助;国家自然科学基金(61633006)。
摘 要:应用近红外光谱技术在线监测工业产品质量时,会出现环境条件变化或仪器的部件如探头或光纤更换的情况,使原模型不再具有原来的预测效果,但是完全从头开始采集数据重新建立新模型工作量大,造成原来宝贵的模型和数据的浪费。为了解决这一矛盾,本文以一种中草药口服液中多糖含量、可溶性固形物含量及pH为研究对象,利用近红外光谱技术对其进行实时在线检测,研究了主从机分辨率不同的光谱之间的模型转移。模型转移过程利用已建模完成并成功上线应用的模型为原模型,在不能获取原主仪器和从仪器一一对应的标准标样的条件下,找到虚拟标样建立转移矩阵。以直接标准化法结合主成分分析降维作为模型转移方法,以质量指标化学参考值与预测值间的相对误差为指标筛选最佳模型。模型转移结果显示,多糖模型预测值与化学参考值间的相对误差可控制在10%以内,可溶性固形物相对误差在5%以内,pH相对误差在3%以内。在线生产使用表明,转移的模型同原模型一样可有效应用于在线、快速对质量指标做出准确的预测。结果表明,本文提出的采用虚拟标样的模型转移方法对于无法获得主从机一一对应的标准标样的情况下的模型转移,是一个可行的有效方法。In the industrial application of near infrared spectroscopy(NIR) to on-line monitoring of product quality,it is not uncommon that the application environment is changed or instrument components e.g.the probe head or the fiber are replaced,leading to a situation that the original model no longer performs as satisfactory as before,but building completely new models by collecting new calibration data means huge task and waste of the previous valuable model and data.In order to solve the dilemma,by reference to the study on NIR application to on-line real-time monitoring of the polysaccharide content,soluble solid content and pH in the manufacture of a Chinese herbal medicine oral liquid,this paper has investigated model transfer methods between the master(original) NIR spectrometer and the slave(new)spectrometer in which the resolution of the two is different.Using the model that had been built and successfully applied to the original spectrometer as the master model,in the absence of a reference sample to which both the original(master) and the new(slave) spectrometers could be applied,a virtual reference between the two spectrometers was established,and a matrix for mode transfer was thus constructed.The method of direct standardization(DS) combined with principal component analysis(PCA) dimension reduction was used as the model transfer approach,and the best model was selected by using the relative error between the chemical index and the predicted value of the quality index.The result indicates that the relative errors can be controlled within 10% for polysaccharides between the predictions of the model and the chemical reference values,within 5% for soluble solids,and within 4% for pH.The online use shows that the transferred model can be effectively applied to the online and rapid prediction of quality indicators like the original model.The result proves the effectiveness of the model transfer approach presented in this work that makes use of virtual reference samples in applications where no one to one corres
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...