Year 2002
Petteri Kämpjärvi

Online fault diagnosis in an ethylene cracking process

Chemical process plant safety, production specifications, environmental regulations and plant economics are some of the main reasons why process monitoring has been widely studied in recent years. The result of improved process monitoring is a more stable process and better quality of the product.

The literature part focuses on presenting different process monitoring methods. Special attention has been given to neural networks and statistical multivariate methods and their applications in industry.

The purpose of the experimental part was to build a system that monitors measurements of an ethylene cracking process. The main task was to find out when the concentration measurements of continuous analyzers cannot be relied on for automatic control purposes. In the case of erroneous measurements, they are suspended in the control application.

First, the right variables for different methods were chosen using a simulator, and the performance of principal component analysis, Kohonen self-organizing maps and radial basis function network were examined and compared. Based on these results a system that consists of principal component analysis and those two neural networks was built and its performance was tested using online data from simulations and from the real process.

The results were encouraging. The system was able to detect and identify faults both from the simulator and process data at the moment the faults occurred. The system was also able to detect such disturbances in the process for which monitoring had not been planned. The redundancy resulting from combining the principal component analysis and two neural networks gave more reliable results than just using one monitoring method. Especially with process data, when principal component analysis detected unknown process disturbances, they were not wrongly identified as analyzer faults.

This info last modified 27 Apr 2024 - contact webmaster