andyco98 / utile-oxy

We present an automated workflow using deep learning for the analysis of videos containing oxygen bubbles in PEM electrolyzers by: 1. preparing annotated dataset and training models in order to conduct semantic seg- mentation of bubbles and 2. automating the extraction of bubble properties for further distribution analysis.

UTILE-Oxy - Deep Learning to Automate Video Analysis of Bubble Dynamics in Proton Exchange Membrane Electrolyzers

We present an automated workflow using deep learning for the analysis of videos containing oxygen bubbles in PEM electrolyzers by 1. preparing an annotated dataset and training models in order to conduct semantic segmentation of bubbles and 2. automating the extraction of bubble properties for further distribution analysis.

The publication UTILE-Oxy - Deep Learning to Automate Video Analysis of Bubble Dynamics in Proton Exchange Membrane Electrolyzers is available in an open access fashion on the journal PCCP for further information!

Description

This project focuses on the deep learning-based automatic analysis of polymer electrolyte membrane water electrolyzers (PEMWE) oxygen evolution videos. This repository contains the Python implementation of the UTILE-Oxy software for automatic video analysis, feature extraction, and plotting.

The models we present in this work are trained on a specific use-case scenario of interest in oxygen bubble evolution videos of transparent cells. It is possible to fine-tune, re-train or employ another model suitable for your individual case if your data has a strong visual deviation from the presented data here, which was recorded and shown as follows:

Model's benchmark

In our study, we trained several models to compare their prediction performance on unseen data. We trained specifically three different models on the same dataset composed by :

• Standard U-Net 2D
• U-Net 2D with a ResNeXt 101 backbone
• Attention U-Net

We obtained the following performance results:

Model	Precision [%]	Recall [%]	F1-Score [%]
U-Net 2D	81	89	85
U-Net with ResNeXt101 backbone	95	78	86
Attention U-Net	95	75	84

Since the F1-Scores are similar a visual inspection was carried out to find the best-performing model :

But even clearer is the visual comparison of the running videos:

Extracted features

Time-resolved bubble ratio computation and bubble coverage distribution

Bubble position probability density map

Individual bubble shape analysis

Installation

In order to run the actual version of the code, the following steps need to be done:

• Clone the repository

• Create a new environment using Anaconda using Python 3.8 or superior

• Pip install the jupyter notebook library

  pip install notebook
  

• From your Anaconda console open jupyter notebook (just tip "jupyter notebook" and a window will pop up)

• Open the /UTILE-Oxy/UTILE-Oxy_prediction.ipynb file from the jupyter notebook directory

• Further instructions on how to use the tool are attached to the code with examples in the juypter notebook

Dependencies

The following libraries are needed to run the program:

 pip install opencv-python, numpy, patchify, pillow, segmentation_models, keras, tensorflow, matplotlib, scikit-learn, pandas

Notes

The datasets used for training and the trained model are available at Zenodo: https://doi.org/10.5281/zenodo.10184579.