基于注意力机制的污水微型动物识别方法  

Identification of Sewage Microorganisms Using Attention Mechanism

作　　者：肖蕾 蓝宗苗 Xiao Lei;Lan Zongmiao(College of Automation,Guangdong Polytechnic Normal University,Guangzhou 510665,Guangdong,China)

机构地区：[1]广东技术师范大学自动化学院,广东广州510665

出　　处：《激光与光电子学进展》2023年第2期239-246,共8页Laser & Optoelectronics Progress

基　　金：2021年度广东省普通高校重点研究领域专项(新一代信息技术)(2021ZDZX1033)。

摘　　要：为了精准掌握污水处理系统活性污泥中微型动物的种类,及时调整污水处理工艺,针对传统机器学习需要人工设计特征、提取特征、设计分类器等复杂过程的弊端,提出一种基于注意力机制和迁移学习相结合的污水活性污泥中微型动物的深度学习识别方法。在迁移学习的基础上,通过对传统的VGG16模型添加注意力模块(SE-Net block),调整输出模块,采用数据增强方法扩充数据集。实验结果表明:相比于改进前的模型,改进后的模型(T-SE-VGG16)能够准确识别不同类型污水活性污泥中的微型动物,测试准确率为98.21%,提高了识别精度,缩短了训练时间,模型收敛速度快,泛化能力强。结果证实了T-SE-VGG16模型对污水活性污泥中的微型动物识别的可行性和可靠性。To accurately identify microorganism species in the activated sludge of sewage treatment systems and modify the wastewater treatment process in real-time,using traditional machine learning methods is a challenge because of various complicated processes.In this study,a deep learning approach based on the integration of attention mechanism and transfer learning is proposed to accurately identify the species of microorganisms in sewage-activated sludge by overcoming the requirements of developing features manually,extracting features,designing classifiers,and other complicated processes.On the basis of transfer learning,the conventional VGG16 model is enhanced by including the attention module(SE-Net block) and modifying the output module,and the dataset is expanded using the data improvement approach.Experimental findings demonstrate that compared with the model before the enhancement,the enhanced model(T-SEVGG16) can accurately recognize microorganisms in various types of sewage-activated sludge with a test accuracy of 98.21%,which enhances the recognition accuracy and reduces the training time.The model converges rapidly and has a strong generalization ability in terms of training time.Moreover,the T-SE-VGG16 model's feasibility and reliability for the identification of microorganisms in sewage-activated sludge are verified.

关 键 词：深度学习 迁移学习 注意力机制 活性污泥 微型动物识别 

分 类 号：TP389[自动化与计算机技术—计算机系统结构]