Abstract
Nowadays, multivariate batch process has been widely adopted in chemical, pharmaceutical and semiconductor high-tech industries. Under such type of process, how to utilize the process data effectively to diagnose abnormal phenomenon and execute fault classification becomes a critical issue. With the progress of advanced automatic data collection and inspection techniques, more and more artificial intelligent data analysis methods are now being emphasized and implemented in the real manufacturing environment. Through the quick computation capability of computer, learning and pattern recognition of data can be realized. In this research, some neural network architectures are provided as the process monitoring and fault classification tools. First, collect the data sets under normal process condition. These data sets are used to train Self-Organizing Map (SOM). By the algorithm of SOM, high dimensional data are automatically clustered into the predefined number of groups and also the natural distribution of the data can be observed via topology map. The trained SOM can remember the structure of data under normal process condition. After that when a new observation comes, through the quantitative information shown by MQE chart, we could know how far it differs from the normal process state. While the process signals something wrong in the previous describe monitoring stage, next step is using the trained back propagation network (BPN) to find out which process variables shift and its shift magnitude. In the real case study, we see neural networks do help the engineer point out the possible abnormal observations and provide the different point of view of the process data analysis.