摘要
分别用常规BP神经网络、贝叶斯正则化BP神经网络及遗传算法-贝叶斯正则化BP神经网络,对多组分有机酸的滴定数据进行主成分非线性拟合。结果显示,贝叶斯正则化能限制网络权值,避免过拟合;遗传算法则使网络的全局优化能力和稳健性提高。对26个测试样本中的乙酸、乳酸、草酸、琥珀酸、柠檬酸和乌头酸6种组分,以及柠檬酸和乌头酸的总量进行了15次拟合预测,平均预测相对误差(RSE)分别为10.02%,9.34%,10.66%,12.18%,29.81%,31.94%和3.80%;性质相似的柠檬酸和乌头酸的拟合预测能力较差,但其总量可得较好的预测结果。应用本法对两种糖蜜中有机酸进行了分析,并与离子色谱分析结果进行了对比。
Based on a back-propagation neural network（BP） integrated with Bayesian regularization and genetic algorithm,a nonlinear fitting of the principal component for the data obtained from titrating multi organic acids was proposed.Results reveal that the combination of the advantages from Bayesian regularization to adjust the effectively network parameters（weights and biases） adaptively for improving generalization and genetic algorithm to find the optimal initial weights and thresholds of neural network ensures global optimum solution with good performance.The method was applied to simultaneously determine acetic acid,lactic acid,oxalic acid,succinic acid,citric acid and aconitic acid in a converted titration data set.It was found that the more the similarities among the organic acids,the worse their predictive performance by models when they are treated individually,however,the result was good when they were treated together.For above six organic acid in sample set,their average relative mean square root errors of predicting results were 10.02%,9.34%,10.66%,12.18%,29.81%,30.94%,respectively,and 3.8% for total amount of citric acid and aconitic acid.Some organic acids in two sugarcane molasses samples are determined and compared with results from ion chromatography.
出处
《分析化学》
SCIE
EI
CAS
CSCD
北大核心
2011年第5期743-747,共5页
Chinese Journal of Analytical Chemistry
基金
广西区科技厅基金(No.0778006-16)资助项目
关键词
遗传算法
神经网络
贝叶斯正则化
糖蜜
有机酸分析
Genetic algorithm
Neural network
Bayesian regularization
Sugarcane molasses
Mixed organic acids
