On-Line Fermentation Process Fault Monitoring Based On JITL Strategy MPLS Algorithm

In the process of modern process industry,the demand for multi specification,multi variety and high quality products is more and more big.The fermentation process with small batch and high added value is of particular importance.It has become the main mode of production in the fields of medicine,food,dyestuff,perfume and biochemical products.In order to ensure the safety and maintainability of fermentation process,it is necessary to monitor the production process.Find out the early failure and get the time for the trouble shooting.Nowadays,the monitoring model of fermentation process is mainly based on data-driven modeling method.By analyzing the data characteristics of fermentation process,the algorithm is improved to achieve the purpose of establishing accurate monitoring model.The main contents and contributions of this paper are as follows:(1)An on line monitoring method based on two dimensional search similar samples JITL-MPLSThe advantages and disadvantages of JITL-MPLS method in the process of fermentation were analyzed.On this basis,we propose a JITL strategy that takes into account the sampling time and similarity.Based on the dynamic problems of batch and batch fermentation process.The sampling time is introduced into the similar sample search process.The dynamic characteristics of the recognition process by using the proportion of similar samples.Adaptive window length adjustment.Combined with the similarity contribution rate,the model is selected to build the JITL-MPLS model.The feasibility and superiority of this method are verified by the normal and fault data generated by Pensim simulation platform.(2)The kernel method is introduced into the JITL strategy to improve the robustness of the MKPLS monitoring methodThe JITL strategy belongs to the local modeling strategy,which can weaken the nonlinearity of the data to a certain extent.But at some stage,the nonlinear characteristics are still obvious,which has a great influence on the JITL-MPLS model.In this paper,the kernel method is introduced into the just in time learning strategy.Since the center of the sample of the JITL strategy is the online sample point,it is not the center of the cluster.Therefore,the local linearization can be better,so as to improve the robustness of the kernel parameters and obtain a better fault monitoring effect.(3)Presents a JITL-MKPLS monitoring method based on residual updating judgment.The traditional JITL strategy belongs to lazy learning".Each collection of data will operate a complete modeling process.This will cause a lot of waste in the process of monitoring more stable phase,affecting the system response to other operations.In this paper,we use the regression characteristic of MKPLS to judge whether the model needs to be updated by comparing the regression residuals.The modeling accuracy of model updating method and traditional JITL method of the same.Greatly reduce the amount of calculation.The application range of this method is extended.(4)Experimental verification of the fermentation process of interleukin-2 in Escherichia coliThe method proposed in this paper is verified by the actual production data of a biochemical pharmaceutical factory in Beijing.In this experiment,the normal batch and the large fault were monitored.And the special experiments on the small faults are carried out.The monitoring results show that the method of monitoring the fermentation process mentioned in this paper,compared with the traditional off-line stage division method can reduce the false alarm rate is sensitive to the small fault,can early warning.Moreover,by introducing the model updating judgment mechanism,the computational complexity is greatly reduced.