Machine Learning Models (SVM, Random Forest,XG Boost) for Ultrasonic Defect Detection in Non-Destructive Testing

Ultrasonic non-destructive testing (NDT) is widely applied in various industries for the early detection of cracks, porosity, and other structural defects. A key challenge in this domain is the automatic interpretation of ultrasonic signals and the reliable classification of defects using machine learning algorithms. In this study, three models Random Forest (RF), XGBoost, and Support Vector Machine (SVM) were implemented to classify ultrasonic testing data.The research workflow included data preprocessing, model training, and performance evaluation. The results showed limited performance for all models, with accuracies close to random guessing. Specifically, Random Forest and XGBoost achieved an accuracy of ۰.۵۰, while SVM reached ۰.۴۰. The maximum F۱-Score reported was approximately ۰.۵۰. Cross-validation results also indicated high variance and poor generalization. These outcomes emphasize the inherent complexity of ultrasonic data and the need for more advanced signal processing and feature extraction methods.The analysis suggests that factors such as data quality, class imbalance, and insufficient feature representation contributed to the suboptimal performance. To improve classification accuracy, advanced signal processing approaches such as Fast Fourier Transform (FFT) and Wavelet Transform, along with hyperparameter optimization, are recommended. Overall, this study highlights the importance of feature engineering and suggests that hybrid or deep learning-based approaches could significantly enhance the performance of machine learning models in ultrasonic defect detection for NDT applications.

کلیدواژه ها:

Ultrasonic Testing ، Non-Destructive Testing (NDT) ، Machine Learning ، Defect Detection