Machine Learning-Based Design Synthesis for Sound Power Minimization Considering a Volume Fraction Constraint
Loading...
Authors
Howes, Jordan
Date
2024-11-21
Type
thesis
Language
eng
Keyword
Acoustics , Machine Learning , Design Optimization , FEA , Sound Power
Alternative Title
Abstract
The design and optimization of structures for acoustics is a challenging task for aerospace engineers, as elevated noise levels in the cabin of an aircraft can have negative effects on passengers. Radiated sound power is a measure of the power radiated from a structure in the form of sound and can be computationally difficult to implement into finite element-based optimization algorithms. Machine learning is a promising approach for sound power optimization as it is effective for finding and exploiting patterns in complex data. Machine Learning has been used to perform regular topology optimization without the need of a sensitivity analysis, making it an appropriate approach to apply to sound power topology optimization as it eliminates the need to compute complex derivatives of the objective function.
The objective of this work is to develop an advanced machine learning algorithm to perform acoustics topology optimization that generates designs with minimal radiated sound power while considering a volume fraction constraint. Supervised learning was used to train a residual network to estimate the broadband sound power of a given design. The trained residual network was then implemented into a reinforcement learning framework that performs design synthesis. A deterministic policy gradient approach was used to train an agent network to generate designs with minimal sound power while adhering to a volume fraction constraint which is incorporated into the loss function.
This methodology was applied to optimize the thickness distribution of a stiffened curved panel, resembling an aircraft fuselage. The inclusion of the volume fraction constraint proved successful, as the agent network generated designs that adhered to the constraint while minimizing sound power. For a volume fraction constraint of 40%, the agent generated designs that achieved a sound power level that was 1.7 dBA lower than the minimum, or 3.5 dBA lower than the average, sound power of designs in the dataset used to train the residual network for that volume fraction. Additionally, a symmetry constraint was introduced, enabling the generation of symmetrical designs, which further reduced sound power. There is significant potential for future work, including enhancing the practicality of the methodology by incorporating additional design variables and considering additional objectives such as buckling.
Description
Citation
Publisher
License
Queen's University's Thesis/Dissertation Non-Exclusive License for Deposit to QSpace and Library and Archives Canada
ProQuest PhD and Master's Theses International Dissemination Agreement
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.
Attribution-NonCommercial-NoDerivatives 4.0 International
ProQuest PhD and Master's Theses International Dissemination Agreement
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.
Attribution-NonCommercial-NoDerivatives 4.0 International