基于LSTM-GAN的青霉素发酵过程异常检测  

Anomaly Detection in Penicillin Fermentation Process Based on LSTM-GAN

作　　者：张秀清[1] 刘哲鹏 王晓君[1] 彭亮亮 Zhang Xiuqing;Liu Zhepeng;Wang Xiaojun;Peng Liangliang(School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang 050018,China;Hebei National Pharmaceutical Group Weiqida Pharmaceutical Co.,Ltd.,Datong 037000,China)

机构地区：[1]河北科技大学,石家庄050018 [2]河北国药集团威奇达药业有限公司,大同037000

出　　处：《信息化研究》2024年第6期22-26,65,共6页INFORMATIZATION RESEARCH

摘　　要：青霉素生产过程是典型的间歇过程,它的发酵过程呈现时变性强和非线性强的特征,在发酵工艺过程中及时发现隐藏的异常行为,可以避免重大经济损失。考虑到发酵过程的复杂性,本文提出一种嵌入了长短期记忆网络(LSTM)的生成对抗网络(GAN)异常检测模型,以发酵培养基中残糖浓度作为研究对象,GAN负责捕获变量之间的潜在关联,进一步提升LSTM的检测能力。同时,使用二元交叉熵损失训练网络,结合判别损失以及重构损失实现发酵过程异常检测。在药厂数据集上的实验结果表明,使用GAN进行青霉素发酵过程单变量异常检测在准确率、召回率的得分上有不错的表现。Penicillin production is a typical intermittent process,and its fermentation process presents strong time-varying and nonlinear characteristics,and the detection of hidden anomalous behaviors during the fermentation process can avoid significant economic losses.Considering the complexity of the fermentation process,a GAN anomaly detection model embedded with LSTM is proposed to take the residual sugar concentration in the fermentation medium as the research object,and the GAN is responsible for capturing the potential correlation between the variables,which further enhances the detection capability of LSTM.Meanwhile,the network was trained using binary cross-entropy loss,and the discriminative loss as well as the reconstruction loss were combined to realize the anomaly detection of the fermentation process.The experimental results on the pharmacy dataset show that univariate anomaly detection in penicillin fermentation process using generative adversarial network performs well in terms of accuracy and recall scores.

关 键 词：异常检测 时间序列 生成对抗网络 长短期记忆网络 青霉素发酵 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]