Institutional Repository of Key Laboratory of Experimental Marine Biology, Institute of Oceanology, Chinese Academy of Sciences
循环水养虾系统智能化水质调控和生物量判别技术研究  
Alternative Title  Research on water quality regulation and biomass evaluation technique of recirculating aquaculture system of Litopenaeus vannamei 
陈福迪  
Subtype  博士 
Thesis Advisor  孙建明 
20200519  
Degree Grantor  中国科学院大学 
Place of Conferral  中国科学院海洋研究所 
Degree Name  理学博士 
Degree Discipline  海洋生态学 
Keyword  循环水养殖系统 凡纳滨对虾 水质调控 深度学习 生物量评估模型 
Abstract  传统的粗放型养殖模式由于人工调控能力小、占地广、污水排放量大等问题迫切需要升级转型。在此背景下，工厂化循环水养殖模式综合利用工程学、现代信息化技术和水产养殖技术，能够有效地解决土地和用水的局限。凡纳滨对虾（Litopenaeus vannamei）是我国对虾养殖产量最高的品种，由于其生长速度快、抗病能力强、苗种供应稳定等特点尤其适宜工厂化循环水的高密度养殖模式。 本研究以凡纳滨对虾循环水养殖为背景，开展了智能化水质调控和生物量判别技术研究。研究分为五个阶段，首先对养殖池曝气布型进行优化，以老旧育苗室中矩形池为基础进行改造，利用计算流体力学对矩形池中微孔纳米管的曝气布型进行分析和优化，确定最适合矩形池的曝气方式和进气流量，解决养虾池内的排污问题，保证养虾池内水环境的清洁。在保证养殖池内水质清洁的前提下，第二阶段开展基于过程控制的RAS水质调控技术研究，设计并构建基于过程控制的变流式循环水养殖系统，通过机器学习算法构建微滤机循环泵联动调控模型，保证循环水系统内水环境的清洁和稳定。第三阶段通过图像处理技术提取对虾图像中的眼球作为特征点并提取眼球长轴长度，构建眼球与体重之间的回归模型，该模型能够作为智能化生物量评估技术的应用基础；第四阶段通过深度学习模型构建水下对虾识别和数量检测模型，以第三阶段构建的生物量模型为基础构建生物量评估模型，从而计算系统内的养殖量。第五阶段构建了基于机器学习的循环水养殖系统中凡纳滨对虾产量预测模型，采用多元线性回归、人工神经网络和支持向量机等方法构建水质指标与生物量之间的关系，对模型进行对比和评价，筛选出适合预测对虾产量的最优模型，通过模型的输出结果能够有效对水环境进行有效评估。 本文的主要研究结果如下： 1.基于计算流体力学优化养殖池水体增氧性能和集污效果 设计了三种微孔曝气管的布局方式（曝气盘、四角和点状布气），研究其在进气口气体流量为6，18和30 m^{3}/h时对矩形池水体的集污和增氧性能。通过计算流体力学模型构建和饵料集污实验验证，发现曝气盘布气和点状布气集污效果较差，四角布气的集污效果最好；在18 m^{3}/h气体流量时可以在池底设计管状集污器实现集污，在30 m^{3}/h气体流量时可以在池底设计圆盘状集污器实现集污。在三种气体流量下，点状布气的氧气转移系数（K_{L}a20）、充氧能力（SOTR）和标准氧转移效率（SOTE）值均为最高，其次为四角布气，最后是曝气盘布气，结果表明点状布气的增氧性能最佳。但是在18 m^{3}/h和30 m^{3}/h气体流量下，四角布气和点状布气的K_{L}a_{20}、SOTR和SOTE已经没有显著性差异（p > 0.05）。因此在用于循环水养殖的矩形池中，综合考虑集污和增氧的需要，四角布气的方式是最优选择。 2.变流式循环水养殖系统设计与构建 设计并构建了微滤机—循环泵联动工作模型。将RAS设置为3个不同循环量档位（55 m^{3}/h、65 m^{3}/h和75 m^{3}/h），开展60d内的变流式循环水养殖实验。结果发现，氨氮和亚硝酸盐氮的质量浓度随着循环量的增加逐渐降低，亚硝酸盐氮的转化具有滞后性，高循环量有利于保持水质清洁。三个不同循环量档位下的浊度反冲次数拟合关系分别为y=0.2883x+0.6476、y=0.3356x+2.9733和y=0.502x+1.9866回归系数R^{2}分别为0.8974、0.9029和0.9256。在三档循环量下的每个单位时间段（0.5 h）微滤机反冲次数的阈值分别为4、9、11次，以阈值作为三档循环量的升档（1）降档（1）的分类依据，采用传感器数据作为解释变量构建支持向量机的二分类模型，采用网格寻优方式优化SVM模型的参数，训练集和测试集的综合评价指标（Fscore）分别为100%和97.83%，模型构建结果良好，适合用于变流式循环水养殖系统的智能化流量调节。 3.基于机器视觉技术的凡纳滨对虾生物量计算模型 基于机器视觉技术建立了对虾眼球体重关系模型。共采集295尾对虾并测量其眼球长轴长度、体长和体重作为数据样本。对虾眼球长轴长度采用机器视觉技术识别和测量，对数据集采用连续拟合与分段拟合两种方法构建了对虾眼球（D）体重（W）关系模型，连续拟合关系模型为：W=38.865d^{2.7914}；分段拟合模型为：W=177.41D^{3.7363}（D<0.2 cm），W=51.003D^{3.104}（0.2 4.基于深度学习的对虾水下图像识别模型 基于RAS中采集的水下图像构建对虾个体检测深度学习模型，模型的平均精度均值为86.21%。使用YOLOv3模型能够准确检测完整的对虾个体，且准确度高，响应速度快。使用Canny算子提取图像中完整对虾的眼球轮廓，与第三章的眼球体重分段回归模型配合能够有效估算生物量，采集235尾对虾样品有针对性地测试模型，在30%容忍度下，对虾体长体重的均方根误差分别为0.3446和0.6269，因此本章构建的混合模型能够有效用于对虾生物量估算。 5.基于机器学习的凡纳滨对虾产量预测 构建基于传感器采集的水质数据、日常养殖操作数据和对虾体长体重数据构建基于机器学习方法的对虾产量预测模型。多元线性回归、反向传播神经网络和支持向量机作为构建预测模型的方法。采用平均绝对误差（Mean Absolute Error，MAE）、平均绝对百分比误差（Mean Absolute Percentage Error）和均方根误差（Root Mean Square Error, RMSE）三个指标对建立的模型评估，支持向量机的MAE、RMSE和MAPE指标分别为0.90、1.06和21%，与另两种方法所构建的模型相比，支持向量机模型鲁棒性更强，预测精度更高，适合作为循环水养殖系统中预测凡纳滨对虾产量的应用模型。 
Other Abstract  With the development of environmental protection supervision, the traditional extensive aquaculture needs to be upgraded and transformed due to the problems of manual control difficulty, wide land occupation and large sewage discharge. The comprehensive utilization of engineering, modern information technology and aquaculture technology in the industrial recirculating aquaculture can effectively solve the limitations of land and water. Litopenaeus vannamei is the species with the highest shrimp production in China. Due to its fast growth rate, strong disease resistance, and sustainable larva supply, it is particularly suitable for highdensity culture in the recirculating aquaculture mode. In this study, the water quality regulation and biomass evaluation of the recirculating aquaculture system (RAS) of Litopenaeus vannamei were studied. The study was divided into five stages. Firstly, the cultural pond was optimized based on the rectangular pond in the old seedling workshop, the aeration pattern of microporous nanotubes in the rectangular pond is analyzed and optimized by using computational fluid dynamics, so as to determine the aeration mode and inlet flow rate most suitable for the rectangular pond. The sewage problem in the shrimp pond was solved, and the water environment in the shrimp pond was kept clean. In the second stage, water quality control technology based on process control was applied to design and construct a recirculating aquaculture system with variable flow based on microscreen filter detecting turbidity. Machine learning algorithm was used to develop a drum filter and recirculating pump united controlling model to ensure the clean and stable water environment in the recirculating aquaculture system. In the third stage, the eyeballs in shrimp image were extracted as the feature points and the length of the eyeball axis was extracted by image processing technology. Regression models were constructed between the long axis and the body weight. This model is the application basis of intelligent biomass assessment technology. In the fourth stage, the deep learning was used to construct the underwater shrimp recognition and quantity detection model. The biomass model constructed in the third stage was used to calculate the cultured biomass in the system. In the fifth part, the production prediction model of Litopenaeus vannamei in recirculating aquaculture system based on machine learning was constructed. Multiple linear regression, artificial neural network, and support vector machine were used to construct the relationship between water quality and biomass. By comparing and evaluating the models, the optimal model for predicting shrimp yield was selected. The output of the model can effectively evaluate the water environment. The main results of this dissertation are as follows:
The performance of sewage collection and aeration in rectangular water tanks were compared among three finepore aeration tube (disctype diffuser, four cornertype diffuser and distributetype diffuser) layouts at three airflow rates, 6, 18, and 30 m^{3}/h. The results of computational fluid dynamics (CFD) modeling and bait collection tests revealed that sewage collection using the four cornertype diffuser was better than using the disctype diffuser and the distributetype diffuser. Using the four cornertype diffuser, a stripshaped sewage collector could be set up at the bottom of the tank to collect sewage at an airflow rate of 18 m^{3}/h and a discshaped sewage collector could be set up at the bottom of the tank to collect sewage at an air flow rate of 30 m^{3}/h. In terms of aeration, volumetric oxygen transfer coefficient (K_{L}a_{20}), standard oxygen transfer rate (SOTR), and standard oxygen transfer efficiency (SOTE) of the distributetype diffuser were all the highest at the three air flow rates, followed by four cornertype diffuser, and finally the disctype diffuser, indicating that the aeration performance in the distributetype diffuser was optimal. However, there were no significant differences in K_{L}a_{20}, SOTR, and SOTE between the four cornertype diffuser and the distributetype diffuser at both 18 and 30 m^{3}/h airflow rates. Therefore, the four cornertype diffuser was the optimal choice for recirculating aquaculture, considering the sewage collection and aeration requirements.
The recirculating volume was set as three gradients (55 m^{3}/h, 65 m^{3}/h, and 75 m^{3}/h), and the experiment of variable flow recirculating aquaculture was carried out within 60d. When the recirculation volume were 55 m^{3}/h, 65 m^{3}/h, and 75 m^{3}/h respectively, the mass concentration of ammonia nitrogen and nitrite nitrogen can be maintained at low levels under high recirculating volume. The results showed that high circulating volume was beneficial to keep water clean. The fitting relationship between turbidity and backwash number under three different flow are y=0.2883x+0.6476, y=0.3356x+2.9733, and y=0.502x+1.9866, with the regression coefficient of 0.8974, 0.9029, and 0.9256 respectively. The threshold of the average backwash number of the microscreen filter in each unit time period (0.5h) under the three recirculation gradients were 4, 9 and 11 times, respectively. The threshold value was used as the classification basis of the three gradients, 1 was labeled as increasing and 1 was labeled as decreasing. The classification model of support vector machine was constructed with sensor data as the explanatory variables. The parameters of SVM model were optimized by grid optimization and the Fscore of the training set and the test set were 100% and 97.83, respectively. The results of the model construction evaluated well and can be used for the intelligent flow regulation of the recirculation aquaculture system with variable flow.
Based on machine vision technology, the eyeballweight relationship model of shrimp was established. A total of 295 shrimp were sampled from a recirculating aquaculture system (RAS). The longaxis length (D), body length (L), and body weight (W) of each individual was measured. The long axis length of the shrimp eyeball was identified and measured using machine vision technology. Continuous fitting and piecewise fitting models were used to construct the eyeballweight relationship model for L. vannamei. The continuous fitting relationship model was described as: W = 38.865D^{2.7914}, while the piecewise model was described as: D < 2 mm, W = 0.0326D^{3.7363}, R² = 0.9288; 2 mm ≤ D < 3.9 mm, W = 0.0401D^{3.104}, R² = 0.9629; 3.9 mm ≤ D < 5.8 mm, W = 0.0421D^{3.0311}, R² = 0.9216; 5.8 mm < D, W = 0.103D^{2.6226}, R² = 0.9457. The root mean square error (RMSE) of the piecewise fitting model (0.0244, 0.1575, 0.5034, 0.7072) was smaller than the continuous fitting model (0.8229). The correlation coefficient (R^{2}) of the piecewise model (0.9288, 0.9629, 0.9216, and 0.9457) was similar to that of the continuous fitting model (R^{2 }= 0.9621). The results indicated that the piecewise fitting model is suitable for calculating the biomass of L. vannamei in RAS and provides a novel way of estimating the biomass of L. vannamei cultured in RAS.
A deep learning model for Litopenaeus vannamei detection was constructed based on underwater images collected in RAS, and the mean average precision of the model was 86.21%. YOLOv3 model can accurately detect complete shrimp individuals with high accuracy and fast response time. Canny operator was used to extract the eye contour of the complete shrimp in the image. Then the biomass can be estimated effectively by the stepwise eyeball weight regression model in chapter 3. 235 shrimp samples were collected to test the model. Under the 30% tolerance, the RMSE of shrimp body length and weight prediction were 0.3446 and 0.6269, respectively. Therefore, the hybrid model constructed in this chapter can be used to estimate shrimp biomass.
The water quality data collected by sensors, the data of RAS daily operation and the data of shrimp body length and weight were constructed prediction models of the shrimp production based on the machine learning methods. Multiple linear regression, back propagation neural network, and support vector machine were used to construct the prediction model. Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) were used to evaluate the established model. For SVM, MAE, RMSE and MAPE were 0.90 1.06 and 21, respectively. Compared with the models constructed by the other two methods, the support vector machine (SVM) model was more robust and had higher prediction accuracy, so it is suitable for predicting the production of L. vannamei shrimp in the recirculating aquaculture system. 
Subject Area  海洋科学其他学科 
MOST Discipline Catalogue  理学::海洋科学 
Pages  135 
Language  中文 
Table of Contents  第二章 矩形养殖池中水体集污效果和增氧性能的优化研究... 18 2.2.2 微孔曝气管三种布置方式集污性能的比较... 20 2.2.3 微孔曝气管三种布置方式增氧性能的比较... 22 第四章 基于机器视觉技术的凡纳滨对虾生物量计算模型研究... 53 第六章 基于机器学习方法的凡纳滨对虾产量预测研究... 81

Document Type  学位论文 
Identifier  http://ir.qdio.ac.cn/handle/337002/164638 
Collection  实验海洋生物学重点实验室 
Recommended Citation GB/T 7714  陈福迪. 循环水养虾系统智能化水质调控和生物量判别技术研究[D]. 中国科学院海洋研究所. 中国科学院大学,2020. 
Files in This Item:  
File Name/Size  DocType  Version  Access  License  
陈福迪201718006812065.（10830KB）  学位论文  暂不开放  ODC PDDL 
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.
Edit Comment