Soft Sensing Based On Data-Knowledge-Driven Fuzzy Neural Network For Ammonia Nitrogen

Water is the source of life.With the progress of society,the acceleration of industrialization and urbanization,the water consumption is increasing day by day,and the amount of sewage discharge is also increasing,thus aggravating the water shortage and pollution.Sewage treatment is one of the important ways to alleviate water shortage and pollution.Ammonia nitrogen is one of the key indicators to evaluate water quality,and its concentration value reflects the removal effect of nitrogen pollutants in sewage treatment.Excessive ammonia nitrogen leads to eutrophication of water bodies and endangers human health.Therefore,the detection for effluent ammonia nitrogen in the sewage treatment process is of great significance.Due to the shortcomings of long-time detection and low accuracy of chemical detection technology in sewage treatment plants,it is difficult to measure effluent ammonia nitrogen accurately and in time.Data-driven soft sensing technology is widely adopted for its real-time,fast,and accurate advantages.However,in the process of sewage treatment,the collected data are often incomplete due to the time-consuming for ammonia nitrogen data collection,sensor failure,environmental interference or human operation,which affects the prediction performance of the established model.Aiming at this problem,a data-knowledge-driven fuzzy neural network(DAK-FNN)model is proposed,which can realize real-time and accurate prediction for effluent ammonia nitrogen.In order to realize the self-organization of DAK-FNN structure,a target model structure adjustment method based on long-short-term memory mechanism is proposed to improve the prediction accuracy and generalization ability of the model.Additionally,to solve the problem of data limited and abnormality,a bounded DAK-FNN model is proposed.The main research work in this paper includes:(1)The auxiliary variables for the effluent ammonia nitrogen soft sensing model are selected.Firstly,12 auxiliary variables with high correlation with effluent ammonia nitrogen are preliminarily selected by analyzing the mechanism,and then the data are collected and preprocessed.The dimensionality of auxiliary variables is reduced by the principal component analysis method,and finally 5 auxiliary variables are determined,which are used as the input variables of soft sensing model.(2)Aiming at the problem of poor performance of models established with incomplete real-time data,a data-knowledge-driven fuzzy neural network(DAK-FNN)is proposed,which not only takes full advantage of the existing knowledge in the source scene,but also utilizes the data of the current scene.In order to effectively combine the existing knowledge of the source scene and the data of the current scene,a knowledge transfer method based on the idea of transfer learning is proposed.This method uses the parameters obtained after pre-training the reference model with lots of historical data as knowledge to guide the target model,adopts the idea of transfer learning,transfers the acquired knowledge to the target model,and uses real-time data to fine-tune the parameters of the target model online.The experimental results show that this method has higher online prediction accuracy and better real-time performance than other methods,which provides a theoretical basis for the establishment of soft-sensing model for effluent ammonia nitrogen.(3)A target model structure adjustment method based on the long-short-term memory mechanism is proposed to realize the self-organization of the DAK-FNN structure.The neurons obtained from the pre-trained reference model are regarded as core neurons,and the acquired knowledge is displayed in the form of core neurons.The structure of the target model is fine-tuned based on the knowledge acquired from the reference model,and according to the principles of biological neuroscience,neurons are divided into core neurons and non-core neurons,which correspond to the long-term and short-term memory characteristics of the model respectively,and the addition and deletion thresholds need to meet different conditions.The experimental results show that this method not only improves the prediction accuracy and generalization ability of the model,but also saves the training time.(4)A soft-sensing model for effluent ammonia nitrogen is established.The 5auxiliary variables are used as the input variables of the soft sensing model,which are determined by the principal component analysis method for effluent ammonia nitrogen,and the effluent ammonia nitrogen is predicted by DAK-FNN model,and its evaluation index is compared with other models.Ammonia nitrogen experiments show that the model can effectively achieve real-time and accurate prediction for effluent ammonia nitrogen;in addition,for the data limited and abnormality problems,the bounded knowledge is embedded in DAK-FNN,and a bounded DAK-FNN method is proposed that determines the bounded knowledge based on system output analysis,and forms constraints on the consequence parameters of the model to ensure that the model output is bounded.The experimental results show that this method has higher prediction accuracy and less training time than other methods.