Perpustakaan Digital ITB

In urban rail systems, metro trains frequently encounter wheel damage specifically due to negotiating the sharp curves, leading to significant wear and fatigue. These deteriorations will degrade the safety, stability, and overall performance of the vehicle. The profile of the wheels and rails is critical, as mismatches in their designs exacerbate wear when rail vehicles run on small radius of curves, thereby reducing vehicle performance and increasing the need for frequent reprofiling to restore wheels to their original standards. Such conditions not only shorten the lifecycle of the wheels but also lead to considerable maintenance costs. Recent research is employing optimization techniques to design new wheel profiles, aiming to enhance vehicle performance by minimizing wear and fatigue, which in turn reduces maintenance demands and costs, ensuring a more sustainable operation of metro rail systems. Several studies related to wheel profile optimization show that optimized wheel profiles can reduce wear and surface fatigue index compared to the original profiles. Additionally, the new profiles also improve the dynamic performance of vehicles and enhance safety, one indicator of which is a lower derailment coefficient. However, the optimization process can yield more accurate results with the use of machine learning. So far, the use of machine learning has been relatively limited in assisting wheel profile optimization. Moreover, only a few studies have used Machine Learning for multi-objective optimization cases on wheel profiles. Therefore, this research aims to optimize the wheel profiles of LRT Jabodebek by minimizing two objective functions, such as wear and fatigue, by building a surrogate model to improve prediction accuracy and data efficiency. This study used numerical simulation models with the Universal Mechanism software to model the railway vehicle and also the wheel wear evolution, which was then validated with field data. The wheel profile design for optimization was created through NURBS parametrization method that requires control points and weight factors based on the initial profile. The sample requirements for training the machine learning model were generated by varying the initial profile into several flange thicknesses from 27.5 mm to 33 mm and also their slopes. The varied wheel ii profiles were simulated with the previously validated dynamic model to obtain vehicle response data. The machine learning model used was a Kriging regression model. Twelve control points defined in each valid sample were selected as model inputs. Vehicle response data in the form of wear index and surface fatigue index served as model outputs. In this study, two separate models were created according to their respective outputs. The training results from the Kriging model showed good accuracy values measured by performance metrics in the form of R2, with the wear index Kriging model R² of 0.91 and the surface fatigue Kriging model R2 of 0.84. The multi-objective model using NSGA-II successfully generated new optimized profiles capable of minimizing objective function values lower than the initial LRT profile. The objective function calculations with the Kriging model also provided valid values against dynamic simulation with error differences below 1%. The optimization of the wheel profile resulted in a considerably improved contact point distribution pattern, contributing to better overall performance. Beyond the targeted objectives, the optimized profiles showed improved performance in contact pressure with the greatest reduction obtained up to 19.04%, decreasing from 2.005 GPa to 1.623 GPa. Regarding safety-related parameters, a reduction in the Nadal criterion and corresponding decreases in lateral forces during curve negotiation were obtained. The optimized profiles also exhibited significant improvements compared to the original profile, with approximately 31% reduction in flange thickness loss after 16,000 km of operation. The results of this wheel profile optimization research can be beneficial for the industry when designing new trains, with the hope of improving vehicle dynamics quality and, of course, minimizing maintenance costs related to railway wheels.