| In recent years,more and more attention has been paid to the clean treatment of papermaking wastewater due to the increasing importance of environmental protection.It is necessary to build a sensor fault detection model to detect the faults in the operation process for the wastewater treatment process.The wastewater treatment process is a continuously operating process,and if something goes wrong during the operation,it can have a severe impact on the environment.With the application of the distributed control system in the wastewater treatment process,the process data collection has increased significantly.In this thesis,to address the problems of high dimensionality,non-Gaussian,and nonlinearity of process data,we adopt the manifold learning method to map the high-dimensional data to the low-dimensional space and extract the key information process data,which can improve the accuracy of the fault detection model.The main content of the researches include:(1)Introducing fault diagnosis in the wastewater treatment process;describing the application of data dimensionality reduction methods in the field of process monitoring,comparing the advantages and disadvantages of different dimensionality reduction methods,and discussing the problems in the application of manifold learning methods in the field of fault detection in the wastewater treatment process.(2)To address the traditional linear dimension reduction methods,which are difficult to identify nonlinear process data and have low fault detection accuracy,an improved t-distributed stochastic neighbor embedding(t-SNE)algorithm is proposed for fault detection of dynamic continuous processes.The improved method can extract potential low-dimensional features of the process data,consider keeping the local structure of the data and use the Mahalanobis distance instead of the Euclidean distance to calculate the conditional probability of similarity between samples point pairs.Results of the simulation show that the improved t-SNE fault detection model has a better fault detection effect.(3)According to the non-Gaussian distribution characteristics of process data,the improved t-SNE algorithm is combined with the Gaussian mixture model(GMM)to construct a fault detection model.The model uses low-dimensional data to train a GMM model and uses an expectation-maximization algorithm to divide the data to obtain a more accurate Gaussian model.Finally,the local probability metric is fused into the global probability metric to realize fault detection.(4)To solve the problem that t-SNE can only preserve the local structure of the process data,it may get stuck in the problem of local optima.A unified manifold approximation and projection(UMAP)is proposed to reduce the dimensionality of process data,and then a fault detection model for the wastewater treatment process is developed combined with the support vector data description(SVDD). |
