基于长短期记忆神经网络的反应液葡萄糖含量预测

Content prediction of glucose in reaction solution based on long short term memory neural network

中文关键词:  双酶法  葡萄糖酸锌  反应液葡萄糖含量  灰色关联分析  长短期记忆神经网络  

英文关键词:double-enzyme method  zinc gluconate  glucose content in the reaction solution.  Pearson correlation coefficient  long short term memory neural network

中文摘要:

      目的 建立皮尔逊相关系数(Pearson correlation coefficient, PCC)和长短期记忆(long short term memory, LSTM)神经网络的反应液葡萄糖含量预测模型用以实时预测葡萄糖酸锌生产过程中反应液葡萄糖含量。方法 通过葡萄糖酸锌制备实验, 结合PCC理论确定对反应液葡萄糖含量有较大影响的因素, 对这些因素进行数据采集并将其作为神经网络的输入变量, 采集反应液葡萄糖含量数据并进行处理, 将其作为神经网络的输出变量, 进而建立反向传播神经网络(backpropagation neural network, BP)和LSTM神经网络的反应液葡萄糖含量预测模型。结果 通过100次模型迭代训练, 对照BP反应液葡萄糖含量预测模型可以看出LSTM反应液葡萄糖含量预测模型在测试集的误差约为0.45%, 误差较小, 准确度较高。结论 基于LSTM反应液葡萄糖含量预测模型显著提高了预测精度, 相比现有检测方法更加智能高效, 能够有效辅助生产进行。

英文摘要:

      Objective To establish Pearson correlation coefficient (PCC) and long short term memory (LSTM) neural network model for predicting the glucose content in the reaction solution during the production process of zinc gluconate. Methods Through the preparation experiment of zinc gluconate, combined with the PCC theory, the factors that have a significant impact on the glucose content of the reaction solution were determined. These factors were collected and used as input variables for the neural network. The glucose content data of the reaction solution was collected and processed as output variables of the neural network, and then a prediction model for the glucose content of the reaction solution was established using the backpropagation neural network (BP) and LSTM neural networks. Results Through 100 iterations of model training and comparing with the BP reaction solution glucose content prediction model, it could be seen that the LSTM reaction solution glucose content prediction model had an error of about 0.45% on the test set, which was relatively small and had high accuracy. Conclusion The glucose content prediction model based on LSTM reaction solution significantly improves the prediction accuracy and is more intelligent and efficient compared to existing detection methods, which can effectively assist production.