Abstract:

Aiming at the problem of concentration prediction in the penicillin fermentation process, an online soft measurement method based on temporal difference memory network (TDMN) was proposed. In view of the multi-batch and non-stationary characteristics of penicillin fermentation process data, this paper uses a sliding window for preprocessing, uses the TDMN model to capture local features, and realizes real-time update of the model through incremental learning to achieve online soft measurement of penicillin concentration. The experimental results show that the model exhibits significant accuracy in penicillin concentration prediction. Compared with traditional methods such as artificial neural network (ANN), recurrent neural network (RNN) and long short-term memory network (LSTM), it has higher real-time and adaptability, and can effectively track the concentration changes during the fermentation process. Through comparison with other models, it is found that the method based on incremental learning significantly reduces the prediction error and improves the prediction accuracy. This method can not only dynamically update the model, but also effectively track changes in penicillin concentration, providing a new solution for monitoring and controlling the status of the fermentation process, and has important application value and industry significance.

Key words: fermentation process, neural networks, non-stationary, prediction, model

摘要：

针对青霉素发酵过程中的浓度难题，提出了一种基于时序差分记忆网络（temporal difference memory network，TDMN）的在线软测量方法。鉴于青霉素发酵过程数据的多批次、非平稳特性，采用滑动窗口进行预处理，利用TDMN模型捕捉局部特征，并通过增量学习实现模型的实时更新，以实现青霉素浓度在线软测量。实验结果表明，该模型在青霉素浓度预测中表现出显著的准确性，与人工神经网络（artificial neural network，ANN）、循环神经网络（recurrent neural network，RNN）和长短期记忆网络（long short-term memory，LSTM）等传统方法相比，具有更高的实时性和适应性，能够有效跟踪发酵过程中的浓度变化。通过与其他模型的比较，发现基于增量学习的方法显著降低了预测误差，提高了预测精度。该方法不仅能够动态更新模型，还能有效跟踪青霉素浓度的变化，为发酵过程状态的监测与控制提供了新的解决方案，具有重要的应用价值和行业意义。

关键词: 发酵过程, 神经网络, 非平稳, 预测, 模型