Using Feature Variables Integrated With Multilayer SOM For Online Monitoring Visualization

As the society developing, process monitoring and fault diagnosis for complex chemical industry process has been attached more and more importance by researchers from home and abroad. What’s more, how to visualize the fault diagnosis results efficiently becomes an important issue in this field. Self-Organization Map (SOM) network can preserve the topological structure and the density distribution of the input dataset and map these data onto the output neurons. Samples belonging to identical class get together on the map which makes the visualization of fault detection and diagnosis. However, the relation of the variables of chemical process data nowadays becomes more and more complicated and the simple SOM algorithm cannot meet the requirements. In order to solve this problem, two feature variables integrated with multiple-layer SOM methods are proposed in this thesis and applied to the fault diagnosis and process monitoring of Tennessee Eastman (TE) process which improves the accuracy rate of fault diagnosis and visualization results. The specific content is studied as follows:(1) Fault diagnosis based on multiple self-organizing maps integrated with canonical variate analysis (CVA) is proposed. CVA does its work as a feature extraction technique by maximizing the correlation degree between two sets of variables. Based on the canonical variates, the complex process data with high dimension are reduced to low dimension, which greatly improved the results of the SOM maps.(2) Fault diagnosis based on multiple self-organizing maps integrated with statistic pattern analysis (SPA) is proposed. SPA is a novel method for data transformation which calculates the different orders statistics of the original data set. The introduction of the high order statistics information makes it describe the status of the process much more exactly. Then multi-SOM is involved to map the cluster results.(3) Case studies of the Tennessee Eastman (TE) process are employed to illustrate the effectiveness of the proposed methods for fault diagnosis and monitoring based on multi-layer SOM. The results show that the feature variables integrated with multilayer SOM maps and growing hierarchical self-organizing map (GHSOM) have the same diagnosis result, but performance of the former methods is much clearer. Both of the two methods proposed in this work are efficient to improve the performance of visualization. The two methods can not only increase the accuracy rate of fault diagnosis, but also be suitable for more fault types, which can make real time monitoring of complex chemical process true.