Smart classification method to detect irregular nozzle spray patterns inside carbon black reactor using ensemble transfer learning

The manufacturing industry has been undergoing a paradigm shift toward the concept of a smart factory. To stay abreast of this paradigm shift, extensive research, particularly in the chemical engineering manufacturing field, has been focused on analyzing newly acquired image data. Consequently, this study proposes a novel method to analyze the nozzle spray patterns of feedstock oil inside a carbon black reactor by analyzing images acquired from a machine vision system. To replace conventional methods making use of naked eye measurements, the images inside a reactor were acquired and processed using three different methods. Several models to detect irregular nozzle spray patterns in processed images have been developed through transfer learning. We combined these individual models to develop an ensemble model that exhibited better performance than the individual models. The effect of the ensemble was verified through gradient-weighted class activation mapping analysis. Using the proposed ensemble model, a test dataset accuracy of 98.5% was obtained.