Supervised and reinforcement learning for the design and optimization of power transmission shafts
Loading...
Date
2026-03-02
Authors
Τζιουφέτας, Αθανάσιος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This diploma thesis focuses on the development and application of machine learning methods for the automated preliminary design of power transmission shafts equipped with mechanical components. Initially, a comprehensive supervised learning framework is developed, based on an extensive dataset generated through analytical mechanical calculations performed in a MATLAB environment. Each dataset entry corresponds to a distinct shaft configuration and includes operating parameters such as transmitted power, rotational speed, and material, as well as the presence of components such as pulleys, gears, sprockets, and bearings.
Within the supervised learning framework, both regression and multi-class classification problems are addressed. Regression models are employed to predict key geometric characteristics of the shaft, namely its diameter and overall length, while classification models aim at the automatic selection of standardized mechanical component codes, including pulley, gear, sprockets and bearing codes, in accordance with catalog-based standards. For this purpose, different algorithmic families are examined and compared, including Random Forest, Support Vector Machine, and Multilayer Perceptron models, with their performance evaluated using established statistical metrics. Particular emphasis is placed on proper data preprocessing, the prevention of information leakage, and the achievement of reliable model generalization.
Subsequently, the scope of the work is extended beyond predictive modeling through the introduction of reinforcement learning methods for active shaft design. Physics-based simulation environments are developed, incorporating fundamental equations of bending, torsion, equivalent stress, and deflection, allowing reinforcement learning agents to interact with mechanically consistent system representations. Through appropriately formulated reward functions, the agents learn to select shaft diameters and materials that simultaneously satisfy strength, stiffness, and efficiency criteria.
The combined utilization of supervised and reinforcement learning highlights the potential for creating intelligent mechanical design tools that not only reproduce classical analytical design methods with high accuracy but also support the exploration of alternative and efficient design solutions across a wide range of operating conditions.
Description
Keywords
Supervised learning, Reinforcement learning, Shaft design, Agents, Optimization