Classification of natural circulation two-phase flow image patterns based on self-organizing maps of full frame DCT coefficients

Abstract

Many of the recent nuclear power plant projects use natural circulation as heat removal mechanism. The accuracy of heat transfer parameters estimation has been improved through models that require precise prediction of two-phase flow pattern transitions. Image patterns of natural circulation instabilities were used to construct an automated classification system based on Self-Organizing Maps (SOMs). The system is used to investigate the more appropriate image features to obtain classification success. An efficient automated classification system based on image features can enable better and faster experimental procedures on two-phase flow phenomena studies. A comparison with a previous fuzzy inference study was foreseen to obtain classification power improvements.

In the present work, frequency domain image features were used to characterize three different natural circulation two-phase flow instability stages to serve as input to a SOM clustering algorithm. Full-Frame Discrete Cosine Transform (FFDCT) coefficients were obtained for 32 image samples for each instability stage and were organized as input database for SOM training. A systematic training/test methodology was used to verify the classification method. Image database was obtained from two-phase flow experiments performed on the Natural Circulation Facility (NCF) at Instituto de Pesquisas Energéticas e Nucleares (IPEN/CNEN), Brazil.

A mean right classification rate of 88.75% was obtained for SOMs trained with 50% of database. A mean right classificationrate of 93.98% was obtained for SOMs trained with 75% of data. These mean rates were obtained through 1000 different randomly sampled training data. FFDCT proved to be a very efficient and compact image feature to improve image-based classification systems. Fuzzy inference showed to be more flexible and able to adapt to simpler statistical features from only one image profile. FFDCT features resulted in more precise results when applied to a SOM neural network, though had to be applied to the full original grayscale matrix for all flow images to be classified.