Multi-model And Just-in-time Learning Based Soft Sensor Development For Penicillin Fermentation Processes

As the scale of the biological fermentation processes continue to expand,it is significant to improve the product quality and yield while reducing production costs through advanced control and optimization of fermentation processes,and these advanced control theories and methods rely on accurate measurements of important process parameters such as product concentration and cell concentration in real time.Traditional measurement methods have long time lags,which is difficult to meet the requirements of advanced control for real-time measurement of key parameters.Soft sensor technology is an effective means to address the problem of online estimation for important process parameters in complex processes,and can provide timely feedback information for online monitoring and control of processes,thus to achieve the purpose of improving the quality and efficiency.Therefore,this thesis analyzes the characteristics of penicillin fermentation processes and establishes the corresponding soft sensor models by combining the multi-model and the just-in-time learning strategies.(1)To solve the problem of high dimension of variables and correlation between variables which affect the performance of soft sensor model,a soft sensor modeling method based on sparse auto-encoder with mutual information is proposed by using sparse auto-encoder to extract features from input data and introducing mutual information to weight at the same time.First,the reconstruction error term in the loss function of the sparse auto-encoder is weighted by calculating the mutual information between the input and output variables,so that the input variables with different correlations with the output variables have different reconstruction accuracy and the more relevant features with the output can be extracted.Then,a least squares support vector machine model based on the extracted features is established.Simulation experiments of the penicillin fermentation processes show that the accuracy of online prediction of penicillin product concentration can be improved by introducing the mutual information weighting method to extract more relevant data features with the output to establish a soft sensor model.(2)To solve the problem that the fermentation processes have multi-phase and strong nonlinearity,and the estimation accuracy of a single soft sensor model is difficult to meet the requirements,a multi-model soft sensor modeling method based on mutual information sparse auto-encoder and improved density peak clustering is proposed.First,the similarity function is introduced to calculate the k-nearest neighbors of each sample point and the shared neighbors with its k-nearest neighbors to redefine the local density of the sample point.Then,the k-nearest neighbor relationship between the sample points is used to redefine the allocation strategies of the sample points.Finally,the improved density peak clustering algorithm is used to cluster the training samples and the features of subsets can be extracted based on the mutual information sparse auto-encoder to develop the corresponding least squares support vector machine models,the outputs of sub-models are further fused by using the switching method.The simulation results show that the modified local density calculation and sample point allocation strategy by using k-nearest neighbors and shared neighbors can effectively improve the clustering quality and the estimation accuracy of the soft sensor model.(3)Aiming at the strong time-varying characteristics of the fermentation processes,the soft sensor model established offline will age with time and environment and its predictive performance is difficult to guarantee,so a new soft sensor modeling method based on multi-stage error compensation and just-in-time learning is proposed.First,based on the just-in-time learning strategy,for each sample point in the training sample set,a similar sample set in the training set is selected for modeling with the criterion of minimum weighted distance,and the estimated output of the built model is obtained and the error between which with the true value is calculated.Then,the training set is divided into four stages by the improved density peak clustering algorithm,the error model between the input variables and the calculated errors in the training set is established for different stages.For the new test sample,a soft sensor model is established based on the just-in-time learning strategy to obtain the estimated value,and the error compensation value calculated by the error model of the corresponding stage is added as the final estimated value of the soft sensor model.The estimation results of the product concentration in the penicillin fermentation processes show that the multi-stage error compensation model can effectively compensate the estimated output value.