Prediction of Inclusion Types From BSE Images: RF vs. CNN

Abstract

The analysis of non-metallic inclusions is crucial for the assessment of steel properties. Scanning electron microscopy (SEM) coupled with energy dispersive spectroscopy (EDS) is one of the most prominent methods for inclusion analysis. This study utilizes the output generated from SEM/EDS analysis to predict inclusion types from BSE images. Prediction models were generated using two different algorithms, Random Forest (RF) and convolutional neural networks (CNN), for comparison. For each method, three separate models were developed. Starting with a simple binary model to differentiate between inclusions and non-inclusions, then developing to more complex four and five class models. For the 4-class model, inclusions were split into oxides, sulfides, and oxy-sulfides, in addition to the non-inclusion class. The 5-class model included specific types of inclusions only, namely alumina, calcium aluminates, calcium sulfides, complex calcium-manganese sulfides, and oxy-sulfide inclusions. CNN achieved better accuracy for the binary (92%) and 4-class (78%) models, compared to RF (binary 87%, 4-class 75%). For the 5-class model, the results were similar, 60% accuracy for RF and 59% for CNN.

Classification of Nonmetallic Inclusions in Steel by Data‐Driven Machine Learning Methods

“…It was also noted that the prediction models strongly relied on the reliability of the training dataset. [19] The aim of this work was to attempt classification of nonmetallic inclusions based on their morphological and mean gray value (MGV) tabular dataset. The dataset was obtained from the Aztec Feature module.…”

“…This holds it back to serve as an online tool for monitoring steel production. [ 19 ] Therefore, there is a scope to introduce new and efficient techniques to drastically reduce the time and cost factors.…”

“…[ 23–26 ] Machine learning has been implemented in the context of classifying inclusion types from BSE images obtained from the SEM. [ 19,27 ] The primary aim was to determine the chemical composition of the particles only based on the gray value data from the images without subsequent EDS analysis. The work consisted of training a supervised model algorithm with a training data set and then validating it with a test dataset.…”

“…It was also noted that the prediction models strongly relied on the reliability of the training dataset. [ 19 ]…”