Automatic assessment of band carbide in bearing steels based on machine learning

Abstract A machine learning framework was proposed for objective and efficient band carbide assessment. The rating process framework contained three core parts: band carbide recognition, quantization, and rating. First, an improved DeepLabv3+ deep learning architecture was designed to extract the positional information of band carbides. The encoder part of the modified DeepLabv3+ utilized the pre-trained Mobilenetv2 as the backbone feature extraction network with an embedded coordinate attention mechanism; the original bilinear interpolation upsampling was replaced by the decoder part with the CARAFE upsampling operator; and a sum of the cross-entropy loss and the Dice loss was used as the loss function of the model. Compared to DeepLabv3+, the modified DeepLabv3+ achieved high recognition accuracy with improved computational efficiency. Subsequently, the band carbide morphology recognized by the segmentation model was quantitatively characterized. Finally, six classifiers, i.e., Random Forest, Catboost, KNN, SVM, Decision Tree, and Gradient Boosting Decision Tree, were selected to rate the carbide bands. The quantitative features of the band carbides and the manual rating results were taken as model inputs and outputs, respectively. The results showed that the Catboost classifier achieved the best performance with Accuracy, Precision, Recall, and F1-Score of 94.00%, 94.60%, 94.00%, and 94.21%, respectively. This study provides an automated process to evaluate the band carbides, demonstrating the high potential of machine learning in bearing steel microstructure analysis.