Research And Application Of Semi-supervised Soft Sensor Modeling Methods For Imbalance Data In Wastewater Treatment Processes

In wastewater treatment processes,it is a key for the safe and stable operation of wastewater treatment plants to effectively monitor the significant variables.Due to the harsh working environment,high equipment maintenance cost,strong coupling and large lag of data,the traditional instrument detection methods cannot meet the requirements.Soft sensor technology provides a reasonable and effective approach to solve such problems.However,with the development of wastewater treatment technology,wastewater treatment processes are becoming more and more complex,and the distribution of sample data is seriously imbalanced.The traditional soft sensors based on supervised learning cannot meet the prediction requirements.Therefore,depending on the research status of soft sensor at domestic and overseas,this paper carries out a series of soft sensor modeling methods based on semisupervised learning algorithms for the variable prediction problem of various imbalanced data in different wastewater treatment processes.The main innovative research works of this paper are as follows:1.Aiming at the multi-variable prediction problem of static imbalance data in wastewater treatment processes,Co-training and Tri-training based semi-supervised multi-output soft sensors are proposed.Firstly,the formula of confidence under multi-output systems is defined and used for Co-training and Tri-training algorithms,which solves the problem of imbalanced distribution of static data.At the same time,the single-output regression algorithms are multivariate extended,and then combined with the Co-training and Tri-training algorithms for the first time to establish semi-supervised multi-output soft sensors based on Co-training and Tritraining,which achieves the simultaneous prediction of multiple output variables.The proposed soft sensors can solve the multi-variable prediction problem for static imbalance data under different conditions,and their prediction performance and prediction efficiency are validated in a numerical simulation case and the Benchmark Simulation Model No.1(BSM1)wastewater treatment simulation platform.2.Facing the multi-variable prediction problem of dynamic imbalance data in wastewater treatment processes,a heterogeneous Co-training based semi-supervised adaptive multi-output soft sensors are proposed.To better simulate complex and variable wastewater treatment processes,an odd-even grouping method is proposed to group the labeled data and obtain two sets of labeled subsets containing global information.Then,two different types of multi-output regression algorithms are used to train and establish the models in two sets of labeled subsets respectively.Finally,a more comprehensive online update of soft sensors is performed to improve the adaptive ability by using the moving window(MW),a long-short term memory neural network(LSTM)with "memory function" and the gain coefficient of Kalman filter(KF).In the case studies of the BSM1 wastewater treatment simulation platform and University of California at Irvine(UCI)wastewater treatment plants,it is verified that the proposed soft sensors have better prediction performance when solving the multi-variable prediction problem of dynamic imbalance data in the wastewater treatment processes.3.For the multi-variable prediction problem of large-scale imbalance data in the wastewater treatment processes,a deep neural networks Co-training based semi-supervised multi-output soft sensor is proposed.The soft sensor constructs a stacked autoencoders(SAE)based deep neural network by stacking multiple AEs and adding a multivariate output layer at the final stage.The deep neural network not only extracts high-quality hidden features,but also links the input variables with multiple output variables to achieve simultaneous prediction for them.Finally,the formula of the confidence of the Co-training algorithm is redefined,and an improved Co-training algorithm for deep neural network training and learning is obtained,which can solve the problem of large-scale imbalance data distribution more efficiently.The prediction performance and prediction efficiency of the proposed soft sensor are validated in the large-scale data collected from the Benchmark Simulation Model No.2(BSM2)and the oxidation ditch wastewater treatment plant.4.To address the variable prediction problem of multi-modal imbalance data in wastewater treatment processes,a structure entropy clustering based hybrid semi-supervised adaptive soft sensor are proposed.Depending on the newly defined dissimilarity measurement strategy,a structure entropy clustering(SEC)method is derived.By analyzing the modal characteristics of sample data and clustering them,the difficulty of modeling and prediction is reduced.To address the problem of multi-modal imbalance data distribution,a hybrid semi-supervised learning algorithm is used to supplement the output variables with appropriate methods for all the unlabeled data which are then used to enrich the labeled dataset.Finally,a clustered subset with highly similar modal characteristics to the test data is selected through the dissimilarity measurement strategy based just-in-time learning(JITL),and a local adaptive soft sensor is established to ensure the prediction performance.In the case studies of UCI and the oxidation trench wastewater treatment plant,it is verified that the proposed soft sensor cannot only reduce the difficulty of modeling prediction but also improve the prediction quality.5.For the variable prediction problem of time-space imbalance data in wastewater treatment processes,an adversarial transfer learning based semi-supervised adaptive soft sensor is proposed.The soft sensor first uses a hybrid transfer learning algorithm that integrates feature-based and model-based learning algorithms to transfer historical data,which solves the problem of imbalanced time-space distribution of data.Then,an adversarial learning(AL)network is designed to guarantee the quality of transfer data and the prediction performance by continuous adversarial learning which happens through feedback information passed between the generator and the adjudicator.Finally,a grounded theory analyzer based on an improved Granger causality analysis(GCA)is added to analyze the causal relationships between the input and output variables,which visualizes the interpretation of the soft sensor running and prediction results.The prediction performance of the proposed soft sensor and the interpretability of the analyzer are verified in the data collected from the BSM2 wastewater treatment simulation platform and the oxidation trench wastewater treatment plant.Finally,in the conclusion and prospect part,the main contributions of this paper are summarized and a brief plan for the future research direction and research focus is provided.