Machine Learning–Based Optimization of Frictional Power Losses in Spur-Geared Transmissions

Abstract

Abstract. Frictional power losses in spur-geared transmission systems significantly affect the efficiency and performance of mechanical systems. Understanding and minimizing these PVZP are crucial for enhancing gear system efficiency. This study presents a comprehensive analysis and optimization of PVZP using statistical methods and advanced machine learning (ML) techniques. The advanced optimization techniques utilized the multidisciplinary design optimization (MDO) platform mode FRONTIER, which incorporated machine learning (ML) algorithms to further refine the models. An initial dataset of 2000 design points was generated using uniform Latin hypercube (ULHC) sampling. Subsequently, the adaptive space filler (ASF) technique was employed to expand the dataset to 2162 points. Sensitivity analysis was conducted using polynomial chaos expansion (PCE) and distributed random forest (DRF) methods to identify the most influential variables. Among various algorithms, the logarithmic artificial neural network (Log. ANN) model most accurately predicted PVZP. Optimization using the multi-objective genetic algorithm (MOGA) led to significant efficiency improvements. The optimized models were validated against experimental data based on Forschungsstelle für Zahnräder und Getriebebau (FZG)-A10 spur gears evaluated at various load stages and rotational speeds, demonstrating strong alignment with empirical results. The study highlights the effectiveness of integrating statistical and ML techniques in optimizing gear performance, providing valuable insights into the critical factors affecting PVZP. These findings underscore the potential for significant enhancements in gear efficiency and reliability through the use of advanced optimization methodologies.