水厂絮凝过程矾花状态的自动识别与应用

Automatic Recognition and Application of Alum State in the Coagulation Process of WTPs

针对净水工艺絮凝剂优化调控过程中面临的非线性、大迟滞和多变量因素问题,文中研究了矾花图像分割和特征提取方法,提出了一种基于深度模糊非参数映射(DFM)模型的矾花状态自动识别技术。该技术将矾花状态细分为密实、中片、大片、不均和稀疏5种类型,通过图像采集系统分别采集相应的图像来构建样本库,并使用样本库训练模型的相关参数。研究结果表明,从矾花图像中提取密度特征和尺寸特征作为模型输入可以获得最佳的识别效果。DFM模型在矾花状态识别方面的准确率可以达到95%以上,明显优于传统的机器学习方法[如支持向量机(SVM)、反向传播(BP)神经网络]和深度学习方法(如ResNet和AlexNet模型)。同时,基于DFM模型的矾花图像识别技术已经成功应用于舟山某水厂,在应用前后,絮凝剂的平均投加量从11 mg/L降低到8.5 mg/L,沉淀池的平均出水浑浊度从0.9 NTU降低到0.4 NTU。

In response to the challenges of nonlinearity,hysteresis,and multivariable factors in the optimization and control process of coagulants for water treatment processes,this paper investigated the method of image segmentation and feature extraction for floc recognition,and proposed an floc image state automatic recognition technique based on the deep nonparametric fuzzy mapping model(DFM).The floc states were categorized into five types:dense,medium-sized,large,uneven,and sparse.Corresponding images were captured using an image acquisition system to build a sample library,and model parameters were trained based on this library.The research results indicated that extracting density and size features from alum flower images as model inputs achieved the best recognition performance.The DFM model achieved an accuracy rate of over 95%in alum flower state recognition,which was significantly superior to traditional machine learning methods such as support vector machine(SVM)and back propagation(BP)neural network,as well as deep learning methods such as ResNet and AlexNet models.Furthermore,the floc image recognition technology based on the DFM model had been successfully applied in a water treatment plant(WTP)in Zhoushan.Before and after the application,the average dosage of coagulant had decreased from 11 mg/L to 8.5 mg/L,and the average effluent turbidity value from the sedimentation tank had decreased from 0.9 NTU to 0.4 NTU.