Machine Learning Based Optimization of Failure Parameters in Highperformance Steel Fastening Screws after Post-processing
Loading...
Date
2025
Journal Title
Journal ISSN
Volume Title
Publisher
ECOLE NATIONALE SUPERIEURE DE TECHNOLOGIE ET D’INGENIERIE – ANNABA
Abstract
The main objective of this work is to develop a machine learning model capable of estimating, from fracture surface images obtained via Scanning Electron Microscopy (SEM), the percentage of martensite present in the fractured zone. This estimation would indirectly help identify metallurgical failures related to prolonged in-service aging, considered here as a form of post-treatment. The adopted approach is based on a pipeline combining: (i) an unsupervised segmentation algorithm (K-means) to extract martensitic regions, (ii) a pretrained convolutional neural network model (EfficientNetB1) adapted for supervised regression, and (iii) fine-tuning training using a dataset of automatically annotated images. Model evaluation on a validation set shows satisfactory performance, with a mean absolute error below 2% in most cases, supported by residual analysis, a discretized confusion matrix, and a t-SNE projection of internal representations.