Faster R-CNN-based Detection and Tracking of Hydrogen and Oxygen Bubbles in Alkaline Water Electrolysis

In this study, a method for detecting and tracking hydrogen and oxygen bubbles during alkaline water electrolysis was developed using Faster R-CNN and DeepSORT. The images required for CNN training were automatically generated by a pseudo-bubble image generation algorithm specifically developed for the purpose of this study. The method was applied to the results of observations on alkaline water electrolysis obtained under various current densities and wire electrode diameters. Evaluation of detection performance using a confusion matrix showed that for the hydrogen evolution reaction (HER) at a current density of 1.0 A cm−2 and a wire electrode diameter of 200 µm, the method achieved a precision of 1.00, recall of 0.840, and F1 score of 0.940, indicating very high detection performance. For the oxygen evolution reaction (OER), bubbles were detected almost perfectly under all conditions, with all detection metrics exceeding 1.00. The proposed method was approximately 20000 times faster than humans. Bubble diameter distribution, total volume, total number, and Sauter mean diameter were obtained and quantitatively assessed, and the relationships between current density and electrode diameter for both HER and OER have been discussed. This method enables accurate, rapid, and automatic quantitative evaluation of visualization results from various alkaline water electrolysis observations, which were previously difficult and labor-intensive to perform manually.