Application of machine learning algorithms for the optimization of the fabrication process of steel springs to improve their fatigue performance

Abstract

Machine Learning algorithms are aimed at building generalizable models to provide accurate predictions or to find patterns from noisy data. These characteristics are potentially beneficial for the fabrication of steel products. In this research, 529 rotating bending fatigue tests (R = -1 and σa = 400 MPa) were carried out on steel suspension spring bars fabricated using different combinations of manufacturing parameters. A reliable regression model (R2 = 0.877 on the test dataset) based on the Gradient Boosting algorithm was obtained. The interpretation of the model was carried out through the Permutation Importance algorithm, revealing the relevance of the temperature in the tempering treatment applied after quenching on the fatigue lifespan. This pattern was quantitatively described by means of the Partial Dependence Plot of this feature. Besides, a specific study was carried out to obtain a reliable interpretation of the results derived from the Machine Learning analysis. In this sense, it has been observed that specimens subjected to high temperature tempering display a lower surface hardness that provokes a higher surface roughness after shot peening; this, in turn, facilitates the initiation of surface cracks during the fatigue tests reducing the fatigue lifespan. This study provides a reliable framework to optimize the suspension spring manufacturing conditions to increase their fatigue lifespan as well as an example, generalizable to other manufacturing processes, of the potential benefits of Machine Learning.