Modeling Wastewater Treatment Processes Using Neighborhood Component Analysis Optimized By Deep Learning

With the development of industry,environmental problems caused by the discharge of industrial wastewater are also increasing.The establishment of an intelligent system for the entire wastewater treatment process can more effectively purify wastewater through intelligent decision-making.However,because the wastewater treatment process involves a variety of physical,chemical and biological reactions,and the process is sensitive to the environment,it is easily disturbed by many factors.In addition,the use of instruments and electrical equipment in the wastewater treatment process will gradually cause problems such as accuracy decline.These problems lead to a variety of potential failures in the wastewater treatment process.Meanwhile,measuring quality indices in the wastewater treatment process by sensors often has time lag,and the cost of measurement and maintenance is usually high.To realize the establishment of wastewater treatment intelligent system,this thesis conducts the following researches on the prediction of effluent quality indices and data-driven fault diagnosis in wastewater treatment process.1.A method for predicting effluent quality indices based on neighborhood component analysis optimized by kernel principal component analysis is proposed.Neighborhood component analysis is used as the basic model for the uncertainty of operating and inflow conditions in the wastewater treatment process.During the establishment of the regression model,a reference point is set.According to the distance between samples,the prediction of the preference point is converted into the probability of the point by neighbor points,thus the prediction of the model for a point is only affected by neighbor points in a local range,which makes the model less affected by the distribution of data.On this basis,kernel principal component analysis is used to extract nonlinear information in original data.Therefore,the prediction performance of the model for nonlinear processes can be optimized,and the accuracy of the model can be improved.2.A method for predicting effluent quality indices based on neighborhood component analysis optimized by spatio-temporal convolutional latent variables is proposed.Different from the neighborhood component analysis prediction model which optimized by kernel principal component analysis,a 2-D convolutional neural network is used as the feature extractor for the correlation between variables and time delay of the wastewater treatment process.The model takes variable and sample as two input dimensions of convolutional neural network to obtain the potential spatio-temporal information in original data.In addition,partial least squares is used to reduce the dimensionality of deep features extracted by convolutional neural network,and latent variables are used as input variables of neighborhood component analysis to retain key information in the data and optimize model’s efficiency.Compared with the unsupervised feature extraction process of kernel principal component analysis,the supervised feature extraction process using spatio-temporal convolutional latent variables can further improve the model’s performance.3.A multi-scale convolutional neural network optimized by proxy neighborhood component analysis for fault diagnosis in wastewater treatment process is proposed.Firstly,use convolution kernels of different scales to perform feature extraction on the original data,and obtain variation features of the data in large range and small range,respectively.Then,by fusing features obtained by convolution at different scales,the error feedback can act on two sub-models of different scales at the same time.Finally,the proxy neighborhood component analysis is used as the loss function to replace conventional cross-entropy loss function,so that the model continuously narrows the distance between similar data and expands the distance between different types of data during the optimization process.Therefore,the model could expands the distance between different types of samples located near the decision boundary in the iterative process,and optimized the accuracy of fault diagnosis.Finally,main research results of this thesis are summarized.Deficiencies in this research and further research directions are analyzed and discussed.