Researches On Batch Fermentation Process Monitoring Using Multivariate Statistical Methods

Fermentation is closely related to people's life, which is a nolinear, dynamic and multiphase fed-batch process. Different from a common industrial process, its mechanism and operation is much more complex than that of a continuous process. The quality of fermentation production is liable to deteriorating by uncertainty of materials, equipment conditions and environment. In order to improve production quality and the safety of fermentation process, an efficient monitoring system must be structured to detcect the abnormal events of the process.According to the aprior knowledge, monitoring methods can be classified into three groups, which are analytical model based methods, prior-knowledge based methods and process data-driven based methods. Due to related to the growth and propagation of thalli, the mechanism of fermentation is very complicated, which is the reason we can't find an accurate analytical model. The complexity of fermentation mechanism restricts the applications and developmnet of the first group of monitoring methods. On condition that much production experience and knowledge of experts can be gained, the second group of methods can be used to monitoring the process. However, we can't get sufficient experience and knowledge of fermentation, so prior-knowledge based methods are not fit for its monitoring. With the development of computer technology, lots of process data is saved. Using the process data, a statistical model can be built to detect the running condition of process, identify the abnormal events, instruct the manufacture, and improve the productivity. Considering the characteristic of the fermentation process, many problems of monitoring is systematically studied and new statistical monitoring methods are proposed. The main contributions are described as follows,(1) Difficulties of monitoring and characteristics of batch processes are introduced. Then basic concepts, content and classification of monitoring are addressed. Subsequently, developnent and study conditions of data-diven methods are discussed. Moreover, problems needed to be solved are emphasized. History of fermentation, property and operation modes of fermentation process, and production flow course are explained. Based on Pensim Benchmark, modeling and monitoring status quo of penicillin fermentation process are illustrated.(2) A monitoring method based on multi-MPLS algorithm is proposed to deal with different operation modes of fermentation process. For each operation mode, an individual MPLS model is built. Modeling and on-line monitoring of a fed-batch penicillin fermentation process is presented. The simulation is performed by means of statistical method such as Hotelling T~2, Q and contribution plots based on Pensim. The results confirm that the proposed method can get a better performance, as fewer false alarms are observed and faults can be detected for various operation modes.(3) An on-line recursive PCA algorithm based on rank-one matrix perturbation is proposed to fit with the time-variant characteristics of the fermentation process. The proposed monitoring method is applied to monitoring a fed-batch penicillin fermentation process on line and compared with conventional PCA monitoring methods. The monitoring results clearly illustrate the proposed method can get the time-variant characteristic of the process, with fewer false alarms within normal batch processes and small fault detection delay when faults existed. Moreover, the computational complexity is greatly reduced and the memory saved. (4) A dynamic nonlinear monitoring method that combines an exponentially weighted moving average (EWMA) and a kernel principal component analysis (KPCA), MEWMA-KPCA method, is proposed to treat with the dynamic and nonlinear characteristics. The proposed method is applied to a simple dynamic nonlinear process and a fed-batch penicillin fermentation process. The simulation results confirm that, comparing with other methods, the proposed method results in better performance, with fewer false alarms and missed faults within normal batch processes and small fault detection delay attained when faults existed.(5) Updating the monitoring model duly to fit with the time-variant characteristics has vital significance in detecting abnormalities of chemical process and equipment faults exactly. As traditional ICA-based methods, using to updating the model had a high computational load, rufs-ica algorithm is proposed to model and monitor a fed-batch penicillin fermentation process on line. Compared with conventional ICA methods, the proposed method reduced the false alarm rate significantly and overcame failure reports. Compared with other methods, the algorithm can greatly decrease the computation time and saved the memory.Finally, conclusions and further study aspects are given.