Research On Soft Sensor Method For Anaerobic Fermentation Process Of Kitchen Waste Based On Deep Learning

As a kind of the organic waste with high solid content in urban areas,kitchen waste has caused great harm to people's health and living environment.To this end,the efficient conversion of organic solid waste has been listed as a key research and development plan for the 13 th Five-Year Plan in China.According to the high organic matter content of kitchen waste,anaerobic fermentation has been used as the preferred treatment.The strong nonlinearity,strong time-varying and strong correlation in the fermentation process of kitchen waste make it hard to maximize the utilization of resources in the process of treating kitchen waste by anaerobic fermentation,and even the appearance of acidification leads to the failure of the fermentation process.Therefore,the intelligent detection and intelligent control of the anaerobic fermentation process of kitchen waste is the hotspot and difficulty of research.The paper studies the monitoring of the anaerobic fermentation process of kitchen waste by soft-measurement technology,which is difficult to monitor in real time.For the concentration of volatile fatty acid that cannot be measured in real time,a feature selection algorithm for redundant analysis based on MIC is proposed to determine the auxiliary variables and improve the prediction efficiency and accuracy of the model.The DBN algorithm is improved.On the one hand,the input data is preprocessed by Gaussian mixture model and ensemble empirical mode decomposition;and on the other hand,the extreme learning machine is used to improve the training of deep belief network,and the prediction accuracy of the model is improved.A dynamic soft-measurement model based on IDBN-LSTM is proposed,and the static model is dynamically corrected by LSTM,which further improves the generalization ability of the model.The main research contents are as follows:In this paper,the soft sensor modeling for the concentration of volatile fatty acid in anaerobic fermentation of kitchen waste is studied.Firstly,the MIC-based redundant analysis feature selection algorithm is studied,and the auxiliary variables are selected to improve the prediction efficiency and accuracy of the model.Then the modeling method based on improved DBN is studied,and the Gaussian mixture model and the ensemble empirical mode decomposition are used for further pretreatment of the input data.And the extreme learning machine is used to improve the training of the deep belief network,which improves the performance of the model.Finally the dynamic soft sensor model based on IDBN-LSTM is studied,and the static model is dynamically corrected by the LSTM model to improve the prediction accuracy of the model.The main research contents are as follows:The MIC-based feature selection algorithm for redundant analysis is studied,for the selection of auxiliary variables is often based on the experience of human experience.The necessary features are selected,and the useless features,the harmful features and some of the redundant features are removed based on the analysis of feature correlation.The feature selection algorithm is validated by three common datasets of the UCI machine learning library and the actual dataset of the anaerobic fermentation plant for handling kitchen waste collected from the field.The simulation results show that the MIC-based redundant analysis feature selection algorithm is more accurate and effective compared with the tradition methods such as SFS,FCBF and MRMR.The improved DBN soft measurement model is studied because of the strong nonlinearity and large time lag of the anaerobic fermentation process of kitchen waste and low precision of traditional soft measurement methods.Firstly,the input of the model is classified and decomposed by GMM algorithm and EEMD algorithm.Then the soft measurement prediction model is established by DBN and the DBN model is trained by the deep learning algorithm to get better structural parameter values.Finally,the ELM algorithm is used to improve the model performance.The method is applied to the soft measurement of concentration of VFA in the anaerobic fermentation process of kitchen waste.The experimental results show that the prediction accuracy is higher and the generalization ability is better than the traditional methods such as PLSR,BP neural network and SVM and DBN alone.The dynamic soft sensor model based on IDBN-LSTM is proposed for the dynamic characteristics of industrial processes.Firstly,the static soft-measurement model based on IDBN is established and there is a certain error between the predicted output and the expected output of the model.Then the predicted output and the expected output are subtracted to obtain the error time series.The series are used as the input of the LSTM model for modeling and prediction.Finally,the predicted output value of the IDBN static model and the predicted output value of the LSTM model are superimposed to achieve dynamic correction of the static model.The simulation results show that the prediction accuracy of the dynamic model based on IDBN-LSTM is higher and the generalization ability is better than the IDBN static model without dynamic correction.The performance of the model is improved by dynamic correction.The algorithm proposed in the paper is verified by the data from a kitchen waste fermentation enterprise.The results show that the prediction accuracy and generalization ability of the soft measurement model based on the proposed and improved algorithm is good.It provides the possibility for real-time monitoring of the anaerobic fermentation process of kitchen waste.