基于注意力Seq2Seq神经网络的生物强化系统厌氧发酵菌体质量预测研究  

Study on quality prediction of anaerobic fermentation bacteria in bio-enhancement system based on Attention Seq2Seq neural network

作　　者：毛腾跃[1,2,3,4] 李星星 占伟 杜亚光 贴军 郑禄 MAO Tengyue;LI Xingxing;ZHAN Wei;DU Yaguang;TIE Jun;ZHENG Lu(School of Computer Science,South-Central Minzu University,Wuhan 430074,China;School of Resources and Environment,South-Central Minzu University,Wuhan 430074,China;Hubei Engineering Technology Research Center for Heavy Metal Pollution Prevention and Control,Wuhan 430074,China;Hubei Engineering Technology Research Center for Intelligent Management of Manufacturing Enterprises,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,湖北武汉430074 [2]中南民族大学资源与环境学院,湖北武汉430074 [3]湖北省重金属污染防治工程技术研究中心,湖北武汉430074 [4]湖北省制造企业智能管理工程技术研究中心,湖北武汉430074

出　　处：《湖北师范大学学报（自然科学版）》2024年第2期37-44,共8页Journal of Hubei Normal University：Natural Science

基　　金：国家民委中青年英才培养计划(MZR20007);湖北省技术创新计划重点研发专项(2023BAB087);新疆维吾尔自治区区域协同创新专项(科技援疆计划)(2022E02035)。

摘　　要：生物强化厌氧发酵系统能够提高发酵效率和产物质量。然而,在生物强化甲烷厌氧发酵过程中,关键的生物参数难以实时在线测量。为了解决这一问题,提出了一种基于注意力融入Seq2Seq-LSTM模型的质量预测方法。通过编码器将时间序列数据输入,并引入注意力机制以增强对重要信息的关注,从而得到更新后的中间向量;在解码器中同样引入注意力机制,利用LSTM神经网络对当前时刻的中间向量和输入信息进行综合处理。同时,为了提高模型的稳定性,使用了Adamw梯度下降优化器进行训练。最后,将该方法与LSTM、AM-LSTM模型一同应用于甲烷发酵菌体质量预测并进行对比。实验结果表明,该模型拟合能力和预测准确性均优于其他两种模型,能够更好适用于甲烷发酵菌体质量的在线预测。Bio-enhanced anaerobic fermentation system can improve fermentation efficiency and product quality.However,in the process of bio-enhanced methane anaerobic fermentation,key biological parameters are difficult to be measured online in real time.To solve this problem,this study proposes a quality prediction method based on attention fusion into the Seq2Seq-LSTM model.The method inputs the time series data through an encoder and introduces an attention mechanism to enhance the focus on important information to obtain the updated intermediate vectors.The attention mechanism is also introduced in the decoder,and the LSTM neural network is utilized to synthesize the intermediate vectors and input information at the current moment.Meanwhile,in order to improve the stability of the model,Adamw gradient descent optimizer is used for training.Finally,the method is applied to methane fermentation bacterial mass prediction together with LSTM and AM-LSTM models.The experimental results show that the model fitting ability and prediction accuracy are superior to the other two models,and can be better applied to the online prediction of methane fermentation bacterial mass.

关 键 词：生物强化 厌氧发酵 质量预测 LSTM神经网络 注意力机制 Seq2Seq模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]