Abstract: In the fermentation state,the online prediction accuracy of product concentration is unstable due to differences between batches.To address this problem,a Dynamic Prediction approach for fermentation product concentration using Incremental Learning(ILDP)was proposed,and a similarity calculation module based on feature dimension reduction was presented for historical and new samples,which helped address the issue of fewer newly added label samples during fermentation.An adaptive update module based on incremental learning was proposed to update model parameters by calculating the loss gradient both new and training samples,enabling quick adaptation to updates with limited newly added label samples.Finally,experiments on the IndPenSim dataset validated the predictive performance of the proposed method on fermentation data from various batches.

Key words: fermentation process, soft sensor, incremental learning, self-adaption

摘要： 针对实际发酵过程中不同批次之间的差异性导致产物浓度的在线预测精度不稳定的问题,提出一种基于增量学习的发酵过程产物浓度动态预测方法,并提出基于特征降维的相似度计算模块,对历史样本与新增样本进行特征降维,从历史样本中选取与新增样本相似的样本填充新增样本集,解决发酵过程新增标签样本少的问题;提出基于增量学习的自适应更新模块,通过计算新增样本与模型训练样本的损失梯度来更新模型参数,使模型在新增标签样本少的情况下具备快速自适应更新的能力;在青霉素公开数据集IndPenSim上进行实验,验证了该方法在不同批次发酵数据上的预测性能。

关键词: 发酵过程, 软测量, 增量学习, 自适应