Abstract:

Establishing chemical process models using deep learning often faces challenges such as high costs of data acquisition and scarcity of data. In response to this, this paper proposed a low-cost data augmentation modeling method called mGAN-NN, which was based on generative adversarial networks (GANs). The method introduced the maximum mean discrepancy (MMD) into the loss function of the generator of GANs, thereby improving the similarity of the distribution features between the generated data and the real data. Then, the generated data were used to establish a neural network model, which was subsequently fine-tuned using real experimental data. The significant advantage of this method was its ability to construct robust models with limited data. The proposed method was applied to establish a prediction model for the active content of fatty acid methyl ester sulfonate (MES) in a microreactor. The model achieved a coefficient of determination (R²) of 0.91, representing a 236% improvement over the traditional neural network (ANN) and a 32% improvement over the support vector regression (SVR). Moreover, the mean absolute error (MAE) was reduced to 3.38, which was 58% lower than that of ANN and 45% lower than that of SVR, demonstrating excellent generalization and prediction accuracy.

Key words: generative adversarial networks, data augmentation, microreactor, modeling, algorithm

摘要：

使用深度学习建立化工过程模型往往面临数据获取成本高、数据短缺等问题。对此，本文提出了一种基于生成对抗网络（GANs）的低成本数据增强建模方法mGAN-NN。该方法在GANs生成器的损失函数中引入最大均值差异（MMD）判据，提高了生成数据与真实数据分布特征的相似度；然后利用生成的数据建立神经网络模型，进而使用真实实验数据对所建模型进行调整。这种方法明显的优势在于，即使在数据量有限的情况下，也能构建出鲁棒性强的模型。采用此方法建立了微反应器中制备脂肪酸甲酯磺酸盐（MES）活性物质量分数的预测模型，其拟合系数（R²）可达0.91，相较神经网络（ANN）提高了236%，相较支持向量机回归（SVR）提高了32%。同时，平均绝对误差（MAE）降至3.38，相较ANN减小了58%，相较SVR减小了45%，展示出了较高的泛化性以及预测精度。

关键词: 生成对抗网络, 数据增强, 微反应器, 建模, 算法