基于神经网络—生长动力学模型对DM423生物量的软测量

以温度、培养时间、溶氧量、pH和葡萄糖浓度为输入变量,生物量浓度为输出变量,研究5-5-1静态多层前馈网络和神经网络—生长动力学两种模型估算的精确度。结果表明:静态多层前馈网络测试样本均方差为1.73×10-3,而神经网络—生长动力学混合模型测试样本均方差为0.25×10-3,其估算精确度优于单独使用静态多层前馈网络对生物量进行估算,动力学模型有较好的泛化能力。

Hybrid neural network based on software measurement for DM423 biomass during batch cultivation

The biomass concentration estimated by the static feedforward multiplayer neural network of 5-5-1 topology and hybrid neural network - microbe growth model with five inputs of culture time, temperature, pH and dissolved oxygen and glucose concentration. The result showed that the static feedforward multiplayer neural network mean squared error (MSE) of testing samples 1.73×10^-3. And hybrid neural network - microbe growth model offered a much better generalization accuracy than that of single neural network model, with MSE of testing samples of 0.25×10^-3, it was found that there was some deviation between estimated biomass and actual values while microbe were growing in the stationary phase.