PSI - Issue 75

Available online at www.sciencedirect.com Available online at www.sciencedirect.com Available online at www.sciencedirect.com

ScienceDirect

Procedia Structural Integrity 75 (2025) 53–64 Structural Integrity Procedia 00 (2025) 000–000 Structural Integrity Procedia 00 (2025) 000–000

www.elsevier.com / locate / procedia www.elsevier.com / locate / procedia

Fatigue Design 2025 (FatDes 2025) Robustness of a Machine Learning Algorithm for Fatigue Life Estimation Under Data Anomalies Philippe AMUZUGA a, ∗ , Mohamed BENNEBACH b , Jean-Louis IWANIACK b Fatigue Design 2025 (FatDes 2025) Robustness of a Machine Learning Algorithm for Fatigue Life Estimation Under Data Anomalies Philippe AMUZUGA a, ∗ , Mohamed BENNEBACH b , Jean-Louis IWANIACK b

a CETIM, 7 Rue de la Presse, 42000 Saint-E´tienne, France b CETIM, 52 Avenue Fe´lix Louat, 60300 Senlis, France a CETIM, 7 Rue de la Presse, 42000 Saint-E´tienne, France b CETIM, 52 Avenue Fe´lix Louat, 60300 Senlis, France

© 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under the responsibility of Dr Fabien Lefebvre with at least 2 reviewers per paper © 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http: // creativecommons.org / licenses / by-nc-nd / 4.0 / ) Peer-review under responsibility of the scientific committee of the Fatigue Design 2025 organizers. Keywords: Machine Learning; Metamodeling; Feature Selection; Data Anomalies; Welded Joint Fatigue; Finite Element Analysis; Welded Assemblies Abstract Accurate estimation of the fatigue life of welded joints is a key challenge in mechanical engineering. Given the limitations of traditional methods, machine learning-based metamodeling—particularly using the Generalized Linear Model (GLM)—o ff ers a promising alternative, combining predictive accuracy, generalization, and analytical interpretability. However, the operational ro bustness of the GLM in the presence of noise typical of industrial data had yet to be rigorously assessed. This study thoroughly analyzes the robustness of a polynomial GLM optimized via automatic variable selection, developed to predict the fatigue life of T-welded joints. Based on data from finite element and fatigue simulations, controlled Gaussian noise was injected into the target variable across various amplitude and proportion combinations. The impact of these perturbations on structural stability (variable selection, coe ffi cients) and predictive performance (RMSE, MAE, accuracy within specific error margins) is systematically ex amined. The GLM demonstrates remarkable resilience, with a paradoxical improvement in performance for certain moderate to high noise configurations (30% amplitude, 10% proportion), reducing RMSE from 0.075 to 0.041 and increasing accuracy from 56% to 80% within a strict ± 1 % margin. Conversely, lower noise levels (10% amplitude and proportion) significantly degrade performance, suggesting the existence of a beneficial critical threshold linked to noise amplitude. Structurally, the model maintains strong stability in selecting core variables despite a slight increase in the number of retained terms. These findings confirm the operational robustness of the GLM under industrial conditions, providing a reliable tool for preliminary design while highlighting the importance of adequate control over input data quality. © 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http: // creativecommons.org / licenses / by-nc-nd / 4.0 / ) Peer-review under responsibility of the scientific committee of the Fatigue Design 2025 organizers. Keywords: Machine Learning; Metamodeling; Feature Selection; Data Anomalies; Welded Joint Fatigue; Finite Element Analysis; Welded Assemblies Abstract Accurate estimation of the fatigue life of welded joints is a key challenge in mechanical engineering. Given the limitations of traditional methods, machine learning-based metamodeling—particularly using the Generalized Linear Model (GLM)—o ff ers a promising alternative, combining predictive accuracy, generalization, and analytical interpretability. However, the operational ro bustness of the GLM in the presence of noise typical of industrial data had yet to be rigorously assessed. This study thoroughly analyzes the robustness of a polynomial GLM optimized via automatic variable selection, developed to predict the fatigue life of T-welded joints. Based on data from finite element and fatigue simulations, controlled Gaussian noise was injected into the target variable across various amplitude and proportion combinations. The impact of these perturbations on structural stability (variable selection, coe ffi cients) and predictive performance (RMSE, MAE, accuracy within specific error margins) is systematically ex amined. The GLM demonstrates remarkable resilience, with a paradoxical improvement in performance for certain moderate to high noise configurations (30% amplitude, 10% proportion), reducing RMSE from 0.075 to 0.041 and increasing accuracy from 56% to 80% within a strict ± 1 % margin. Conversely, lower noise levels (10% amplitude and proportion) significantly degrade performance, suggesting the existence of a beneficial critical threshold linked to noise amplitude. Structurally, the model maintains strong stability in selecting core variables despite a slight increase in the number of retained terms. These findings confirm the operational robustness of the GLM under industrial conditions, providing a reliable tool for preliminary design while highlighting the importance of adequate control over input data quality.

1. Introduction 1. Introduction

Fatigue durability of welded joints is a critical issue in mechanical structure design, as these regions are partic ularly susceptible to mechanical stress [5, 3, 1]. Although traditional approaches such as Finite Element Analysis (FEA) o ff er rigorous results, they remain complex, time-consuming, and costly to implement [1]. In this context, ma- Fatigue durability of welded joints is a critical issue in mechanical structure design, as these regions are partic ularly susceptible to mechanical stress [5, 3, 1]. Although traditional approaches such as Finite Element Analysis (FEA) o ff er rigorous results, they remain complex, time-consuming, and costly to implement [1]. In this context, ma-

2452-3216 © 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under the responsibility of Dr Fabien Lefebvre with at least 2 reviewers per paper 10.1016/j.prostr.2025.11.007 ∗ Corresponding author. Tel.: + 33677079223 E-mail address: philippe.amuzuga@cetim.fr 2210-7843 © 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http: // creativecommons.org / licenses / by-nc-nd / 4.0 / ) Peer-review under responsibility of the scientific committee of the Fatigue Design 2025 organizers. ∗ Corresponding author. Tel.: + 33677079223 E-mail address: philippe.amuzuga@cetim.fr 2210-7843 © 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http: // creativecommons.org / licenses / by-nc-nd / 4.0 / ) Peer-review under responsibility of the scientific committee of the Fatigue Design 2025 organizers.