Fault Monitoring Of Fermentation Process Based On Kernel Entropy Component Analysis

The fermentation industry is one of the important pillar industries of China's national economy,and its prosperity is directly related to the national economy and the people's livelihood.According to statistics,there are more than 5,000 production enterprises in China's fermentation industry.With the continuous improvement of science and technology,the modern industrial process is developing towards integration and diversification.In the process of industry,it is necessary to monitor the quality and safety of production.In case of failure,light causes waste of raw materials,and heavy loss of life and environment.Intermittent production is different from continuous production mode.The batch characteristics and dynamic characteristics carried by itself show different process characteristics and need to be further studied.Therefore,it is very practical to excavate the data information in the intermittent process and establish a real-time monitoring model for the whole fermentation process.Fault monitoring means when fault occurs,can find abnormal timely troubleshooting and inform the relevant operators.This topic in-depth analysis of the fermentation process of multi-stage features and dynamic features,and aiming at the defects of traditional methods used for process monitoring,aims to study a new method of fermentation process fault monitoring,the monitoring of the false alarm rate and leakage alarm rate.(1)Implementation of the failure monitoring of fermentation process based on nuclear entropy component analysis.According to the correlation between the fermentation process variables,the kernel entropy component analysis method was introduced.The entropy value information of the data is fully considered,and it is effective to overcome the shortcoming of the main component based on the characteristic value.The MKECA monitoring model was established by using the data in normal working conditions,and the faults in the intermittent process were detected by two statistics of the monitoring process.(2)Research on the stage attribution of the fermentation process based on the extended kernel entropy load matrix.Aiming at the multi-phase characteristics of the fermentation process,a phase partitioning strategy based on extended kernel entropy load matrix is proposed.The principal component and load matrix are obtained by the kernel entropy component analysis(KECA)of the expanded 2-d time slice data matrix,and the first step is divided according to the number of winners.In addition,the time slice matrix is added to the kernel entropy load matrix to obtain the extended kernel entropy load matrix,and the similarity between each extended load matrix is calculated.The second phase is divided by fuzzy c-means method.By increasing the time index that reflects the change of production process,the insufficiency of sclerosing is overcome effectively,and the jumping point error is avoided.(3)An online updating algorithm based on statistical and mechanism model.A new method combining statistical model and mechanism model is proposed to solve the slow time-varying characteristics of fermentation process.Firstly,the statistical monitoring model is established by using the data under normal working conditions,and the variable mechanism model is established by using the variable correlation between data.If the statistical monitoring model is reported,the relevant variables of the current time are brought into the mechanism model to see whether the threshold is met,so as to decide whether to update the current monitoring model.Thus,not only the shortcomings of the fixed model monitoring time-varying process are overcome,but also the monitoring precision of the model is improved,and false alarm and false alarm are reduced.(4)Field experiment study on the fermentation of Escherichia coli and amino acid.Applying the proposed method in this article to the actual production process,and the effectiveness of the proposed method is proved by e.coli and amino acid fermentation.Experimental results show that the proposed method can effectively extract data information and quickly monitor the occurrence of faults.Compared with traditional methods,false positives and omissions are significantly reduced,with more reliable monitoring performance,which can guide operators to detect and effectively eliminate faults in time.