基于径向基函数(RBF)神经网络的红鳍东方鲀体质量预测  

Prediction of Takifugu rubripes weight based on radial basis function neural network

作　　者：王新安[1,2] 马爱军[1,2] 赵艳飞[1] 岳亮[1] 孙建华[1,2] 孟雪松 刘圣聪 

机构地区：[1]中国水产科学研究院黄海水产研究所,农业部海洋渔业可持续发展重点实验室,青岛市海水鱼类种子工程与生物技术重点实验室,山东青岛266071 [2]青岛海洋科学与海洋技术国家实验室生物与技术功能实验室,青岛266071 [3]大连天正实业有限公司,辽宁大连116000

出　　处：《水产学报》2015年第12期1799-1806,共8页Journal of Fisheries of China

基　　金：国家现代农业技术体系资助项目(CARS-50-G01);中央级公益性科研院所基本科研业务费专项(20603022012005);大连金州新区科技计划(2012-B1-012)~~

摘　　要：为解决基于表型性状预测红鳍东方鲀体质量时,由于不同表型性状间的自相关、部分性状和体质量之间的非线性关系以及线性回归方法自变量间的共线性,导致根据表型性状预测体质量误差过大的问题,本研究根据人工神经网络(artificial neural networks,ANN)建模原理,采用径向基函数(radial basis function,RBF)神经网络模型,利用72个红鳍东方鲀样本的表型数据,通过最近邻聚类算法,构建了基于RBF神经网络的红鳍东方鲀体质量预测模型,并采用线性回归检验法对所构建模型的可信度进行检验。结果显示,基于RBF神经网络预测模型的确定系数R^2为0.992,接近于1,而线性回归模型的确定系数为0.949,比线性回归预测模型的确定系数提高4.53%;此外,线性回归共线性检验显示,模型自变量间存在一定的共线性,性状体周长1和体高的自相关性较大。研究结果表明,通过RBF神经网络方法构建的预测模型,消除了线性回归分析中自变量的共线性问题,预测精度显著高于线性回归方法构建的预测模型,对红鳍东方鲀体质量的预测效果优于线性回归模型。基于RBF神经网络体质量预测模型的构建,为利用表型性状精确评估红鳍东方鲀的体质量提供了一种新的方法。Takifugu rubripes belongs to the bony fishes, Tetraodontiformers, Tetraodontoidei,Tetraodontidae,Takifugu. It is distributed mainly in Japan of the western north Pacific,the Korean peninsula and China ＇s coastal areas. Due to its appealing taste,rich nutrition,lowfat content and numerous trace elements,Takifugu rubripes represents one of the fish species with high economic value. In recent years,Takifugu rubripes are farmed in large numbers in Dalian,Qinhuangdao,Tangshan,and Tianjin regions,and has become the main cultured species of puffer fishery in china. There existed large errors due to self-correlation between different phenotypic traits,non-linear relationship between some traits and body weight and the collinearity among independent variables,when the linear regression model was used to predict Takifugu rubripes weight. As a solution,a Takifugu rubripes weight prediction RBF neural network model,according to Artificial Neural Networks theory and Radial Basis Function model,was constructed with the phenotypic traits（ including total length,body length,body depth,head length,length between eye and head,snout length,mouth width,eye diameter,space between eye and eye,caudal peduncle length,caudal peduncle depth,caudal peduncle breadth,body width,trunk length,trail length,body girth 1,body girth 2 and body weight） of 72 Takifugu rubripes based on the nearest neighbor clustering algorithm,and the credibility of the neural network model constructed was tested by linear regression techniques. The results showed that the coefficient of determination R2 of RBF neural network prediction model and the linear regression model for Takifugu rubripes weight were 0. 992（ approximately 1） and 0. 949,respectively. Obviously,the coefficient of determination R2 of RBF neural network prediction model was improved by 4. 53% compared with the linear regression model. In addition,the collinearity diagnostics of linear regression,based on tolerance and variance inflation factor as well as maximum condition index and maxi

关 键 词：红鳍东方鲀 体质量预测 径向基函数 神经网络 

分 类 号：S917.4[农业科学—水产科学]