PSI - Issue 75

Philippe AMUZUGA et al. / Procedia Structural Integrity 75 (2025) 53–64 Author name / Structural Integrity Procedia 00 (2025) 000–000

54

2

chine learning (ML) methods are emerging as e ff ective alternatives to streamline analysis and accelerate fatigue life prediction, while integrating geometric parameters, loading conditions, and weld quality classes [1]. The e ff ectiveness of ML in various mechanical applications, such as springback prediction in forming or residual stress estimation from manufacturing processes, is already well established [7, 6]. However, its application to welded joint fatigue remains relatively unexplored [1]. Previous studies have compared several regression models (GLM, SVR, RFR, MLP) within a complete metamodel ing framework (FEA and fatigue), highlighting the superior balance of the polynomial GLM with automated variable selection in terms of accuracy, generalization, and analytical interpretability [1]. Its ability to produce an explicit formula is particularly advantageous for rapid integration into engineering design tools. Nevertheless, the practical reliability of ML-based metamodels strongly depends on their robustness to real-world data perturbations such as random noise, measurement errors, and numerical anomalies [2]. Such disturbances can introduce significant bias and a ff ect the stability of feature selection methods like Recursive Feature Elimination with Cross-Validation (RFECV) [4]. This study specifically investigates the robustness of the polynomial GLM with optimized variable selection for fatigue life estimation of T-welded joints, under controlled Gaussian noise injected into the target variable. We sys tematically assess the impact of these perturbations on predictive performance, model structural stability (selected variables, coe ffi cients), and critical thresholds where noise compromises model reliability. Section 2 describes the dataset, experimental pipeline, and Gaussian noise injection protocol. Section 3 presents and analyzes the results. Section 4 interprets them in light of regression theory and engineering implications. Finally, Section 5 summarizes the main findings and outlines future directions.

2. Materials and Methods

This article extends the work presented in Model reduction for fatigue life estimation of a welded joint driven by machine learning [1], which focused on model reduction for fatigue life estimation of welded joints using machine learning algorithms.

2.1. Overview of the Metamodeling Methodology from [1]

Several estimators were evaluated: the Generalized Linear Model (GLM), Support Vector Regression (SVR), Ran dom Forest Regression (RFR), and Multilayer Perceptron (MLP). The training and validation of these algorithms were based on data generated from a parametric finite element (FE) model, with a focus on approaches accessible to design engineers without advanced expertise in data science.

Fig. 1: Summary of the fatigue life estimation method [1]

The modeling workflow is summarized as follows: