Fault Detection And Diagnosis Of Batch Process Based On Multi-scale MICA

Batch process, which is closely related to modern people’s lives, has been extensively used in chemical, pharmaceutical, food, polymer reactions, metal working, etc. So it has a pivotal position and role. It is very necessary to successfully monitor and control batch process, in order to maintain high quality of the final product, and ensure the safe and stable operation of the process, and optimize the production output, and reduce energy consumption effectively.This topic considers batch process as object, and deeply researchs Indpendent Component Analysis(ICA) and wavelet transform for the batch process’ s data is mixed Gaussian distribution, noise, dynamic and multi-scale. In this thesis, mainly the following aspects are researched.(1) Based on the deeply research of Multi-way Independent Component Analysis(MICA), some improvement is proposed on it.MICA has better monitoring results, compared with Multi-way Principle Component Analysis(MPCA), because it can take advantage of higher-order statistics to extract non-gaussian signals from batch process’ s data. Whereas, the 2I and SPE statistics can’t completely get all the features of independent components. Meanwhile, MICA needs to assume the process variables obey non-gaussian distribution. For these two shortcomings, this paper combines One-Class Support Vector Machine(OCSVM) with MICA. Namely, after MICA has extracted the independent components, OCSVM is utilized to train these independent components, then a nonlinear D statistic will be calculated.(2) Considering the batch process’ s data is susceptible to be noised, and variables have strong autocorrelation, and failure often occurs at multiple scales. Multi-scale MICA is researched deeply and some improvement is proposed on it.Multi-scale MICA, which is constructed by combining wavelet transform and MICA, can remove the noise and autocorrelation of the modeling data to some extent. So it is superior to MICA at the aspect of monitoring results. However, the ICA algorithm is used two times to extract features in traditional Multi-scale MICA. First, ICA is utilized to extract information on each scale. Then, ICA is utilized to extract information on the whole scale of the reconstructed data once again. In fact, the statistical significance of use of ICA for the second time is very weak, or even says that it is not necessary. Therefore, combining with previous research, this paper proposes that OCSVM is utilized to build the model for the reconstructed data produced by traditional Multi-scale MICA and then the statistic will be calculated.(3) A fault diagnosis method based on centroid vectors is researched for improved Multi-scale MICA.Process monitoring of batch process generally includes fault detection and diagnosis. A fault detection method based on improved Multi-scale MICA has been researched in our previous work, which obtains better monitoring results. But conventional fault diagnosis method can’t be applied to it. In this respect, referring to the contribution plot method and fault reconstruction method, a new fault diagnosis method based on centroid vectors is researched. Then this new method is applied to the improved Multi-scale MICA and the fault variables can be effectively identified. It is worth noting that the study of fault diagnosis method based on centroid vectors can be applied to all multivariate statistical process monitoring methods, with just a little change. So, it can be said that the new fault diagnosis method has the value of universal application.(4) The above research of fault detection and diagnosis strategy for batch process is compiled to software, and applies it to the actual industrial field.The ultimate goal of the research is to guide practice. Therefore, finally the study of fault detection and diagnosis strategy is compiled to software based on Matlab platform, and applies it to E. coli fermentation scene at a biopharmaceutical company in Beijing Yizhuang. The results indicate that the proposed monitoring strategy can be a good practice guide, and can detect and diagnose the fault timely, and can provide a reference to the operator. So the proposed strategy has high practical value.