Fault Monitoring Of Fermentation Process Based On Deep Auto-encoder

Fermentation process is a kind of complex batch production process,which is widely used in pharmaceutical,food,chemical and other high value-added industries,and plays an important role in the process of economic development in China.In recent years,with the continuous expansion of the scale of fermentation industry in China,the safety and reliability of the production process has been paid more and more attention.Therefore,the monitoring technology of fermentation process state is particularly important.It can find all kinds of abnormal states in the production process in time,reduce economic losses and avoid accidents.In this paper,the data of fermentation process and some problems of monitoring implementation are studied in the following aspects:(1)Research on a fault monitoring method based on mutual information and multi-block autoencoderThere are many variables in the fermentation process and each has a complex correlation relationship.When the local fault that only affects individual variables occurs,there will be many redundant variables to interfere with the fault monitoring and reduce the process monitoring effect.Moreover,the data of fermentation process show nonlinear characteristics.The traditional multivariate statistical method uses kernel method to map to high-dimensional space,which has the problem of unknown kernel space.Therefore,a multi-block auto-encoder method based on mutual information is proposed for fault monitoring of fermentation process.The process variables are divided into sub-blocks by using mutual information,and the closely related variables are gathered together to avoid the interference of redundant variables on fault monitoring.Furthermore,the model of auto-encoder is established to extract nonlinear features,and the SPE statistics is constructed by reconstruction error for on-line process monitoring.The simulation results show that the multi-block auto-encoder method based on mutual information reduces the interference of redundant variables on process monitoring,and has better monitoring effect.(2)Research on a fault monitoring method based on stack contractive sparse auto-encoderTraditional multivariate statistical methods monitor according to the normal data distribution information.However,in the actual process,some faults are not easy to be detected because of the setting of the threshold value,which leads to the problem of misjudgment.The traditional neural network has the problems of gradient vanishing and complex calculation,and the shallow neural network has limited ability to express the complex process,so the hidden features cannot be extracted effectively.Therefore,a monitoring method based on stack contractive sparse auto-encoder is studied for fault monitoring of fermentation process.In order to improve the robustness of the network,the contraction term is added,and the sparseness restriction is used to remove redundant information.The greedy layer-by-layer pre-training method is used to overcome the problem of gradient disappearing in deep neural network training and improve the convergence speed of the model.The simulation results show the validity of the method.(3)Research on a deep autoencoder monitoring method based on KNN ruleIn online monitoring,the traditional multivariate statistical method has a high estimation complexity for high-order statistics,which is difficult to use effectively.The process features extracted by stack auto-encoder have high-order correlation,but there is no specific statistical index for its measurement in monitoring.Therefore,a deep autoencoder monitoring method based on KNN rules is studied to make more efficient use of the extracted high-order data information.The deep neural network structure is used to model the process data,and the original data space is transformed into feature space and residual space.The monitoring indexes of(HD)~2 and(RD)~2 are constructed for process monitoring by using the KNN rules in feature space and residual space.(4)Field experiments of Escherichia coli fermentation processIn order to test the effectiveness and practicability of the proposed method in the actual fermentation process,the fermentation data of Escherichia coli from a biopharmaceutical company in Beijing were used for the experiment.The correlation process variables are divided by mutual information,the deep features are extracted by stack contractive sparse auto-encoder,and the monitoring indexes are constructed according to KNN rules for fault monitoring.Through the analysis of the experimental results,it is verified that the research method in this paper can detect the fermentation process faults timely and accurately.Compared with the traditional methods,the monitoring sensitivity is higher and the false alarm phenomenon is less,so it has a better effect of fault monitoring in the actual process.