基于卷积神经网络的生物式水质监测方法

Biological Water Quality Monitoring Method Based on Convolution Neural Network

生物式水质监测通常是先通过提取水生物在不同环境下的应激反应特征,再进行特征分类,从而识别水质。针对水质监测问题,提出一种使用卷积神经网络(CNN)的方法。鱼类运动轨迹是当前所有文献使用的多种水质分类特征的综合性表现,是生物式水质分类的重要依据。使用Mask-RCNN的图像分割方法,求取鱼体的质心坐标,并绘制出一定时间段内鱼体的运动轨迹图像,制作正常与异常水质下两种轨迹图像数据集。融合Inception-v3网络作为数据集的特征预处理部分,重新建立卷积神经网络对Inception-v3网络提取的特征进行分类。通过设置多组平行实验,在不同的水质环境中对正常水质与异常水质进行分类。结果表明,卷积神经网络模型的水质识别率为99.38%,完全达到水质识别的要求。

Biological water quality monitoring is usually through the extraction of water stress response characteristics in different environments,and then feature classification,so as to identify water quality.Aiming at the problem of water quality monitoring,method of CNN convolution neural network was presented.Fish trajectory is a comprehensive expression of the various water quality classification characteristics used in all the literatures and is an important basis for the classification of biological water quality.Using the image segmentation method of Mask-RCNN to obtain the centroid coordinates of the fish and draw the trajectory image of the fish in a certain period of time.Two sets of trajectory image data sets under normal and abnormal water quality were produced.The Inception-v3 network serves as a feature preprocessing part of the data set,the CNN convolution neural network was reestablished to classify the features extracted by Inception-v3 network.Set up multiple sets of parallel experiments to classify normal and abnormal water quality in different environments.The results showed that the CNN convolution neural network model had a water quality recognition rate of 99.38%,which met the requirements of water quality identification.