A novel generative design method for an optimal silencer using an artificial neural network(ANN) model is proposed. The noise attenuation performance of the silencer is evaluated by its transmission loss(TL) curve. Because the frequency-dependent characteristics of the TL curve strongly depend on the partition layout in the silencer, partitions are optimally placed in the silencer according to the design goal. In this work, the backpropagation method of ANN model is used to develop a generative design method for optimal partition layout. To the end, an ANN model is developed to take a design variable constituting partition layout as input and predict its TL curve as output. A partition generation algorithm is suggested, and a finite element analysis is performed for generating training data. The objective function value is determined by the output of the ANN, and the sensitivity of the objective function value with respect to the design variable is calculated utilizing the backpropagation method of the pre-trained ANN. To ensure that the new partition can be connected to existing partitions, the region where design variables can be updated is restricted, and one design variable within this limited region is updated per iteration during the design process. The proposed design method is applied to the formulated silencer design problem and derives various design proposals satisfying the design goal. The designer evaluates the complexity and performance of the design proposals and selects the optimal partition layout among them. The characteristics and advantages of the proposed design method are discussed and compared with the conventional silencer design method.|인공신경망을 이용한 새로운 소음기의 생성적 설계 방법을 제시한다. 소음기의 소음 저감 성능은 주파수에 따라 달라지는 특성인 투과 손실로 평가한다. 투과 손실의 주파수 의존적 특성은 소음기 내부의 격벽 구조에 강하게 영향을 받기 때문에 설계 목표에 따라 격벽을 최적의 위치에 배치해야만 한다. 본 연구에서는 인공신경망의 역전파법을 이용하여 최적 격벽 구조의 생성적 설계 방법을 개발한다. 이를 위해, 격벽 구조를 구성하는 설계변수를 입력으로 하여 투과 손실 곡선을 예측하는 인공신경망을 개발한다. 학습 데이터의 생성을 위해 격벽 구조 생성 알고리즘을 제안하고 유한 요소 해석을 수행한다. 인공신경망의 출력 값을 이용하여 목적 함수 값을 계산하며 학습된 인공신경망의 역전파법을 이용해 설계변수에 대한 목적 함수 값의 민감도를 계산한다. 새로운 격벽이 기존 격벽 구조와 연결될 수 있도록 설계 변수가 업데이트 될 수 있는 영역을 제한하고 설계 과정에서 제한된 영역 내의 한 개의 설계 변수를 매 반복마다 업데이트한다. 제안한 설계 방법을 정식화 된 소음기 설계 문제에 적용하여 설계 목표를 만족하는 다양한 설계안을 도출한다. 설계안들의 복잡도와 성능을 평가하고 설계자는 그들 중 최적 격벽 구조를 선택한다. 제안한 소음기 설계 방법의 특장점에 대해 논의하고 기존 소음기 설계 방법과 비교한다.
Alternative Abstract
A novel generative design method for an optimal silencer using an artificial neural network(ANN) model is proposed. The noise attenuation performance of the silencer is evaluated by its transmission loss(TL) curve. Because the frequency-dependent characteristics of the TL curve strongly depend on the partition layout in the silencer, partitions are optimally placed in the silencer according to the design goal. In this work, the backpropagation method of ANN model is used to develop a generative design method for optimal partition layout. To the end, an ANN model is developed to take a design variable constituting partition layout as input and predict its TL curve as output. A partition generation algorithm is suggested, and a finite element analysis is performed for generating training data. The objective function value is determined by the output of the ANN, and the sensitivity of the objective function value with respect to the design variable is calculated utilizing the backpropagation method of the pre-trained ANN. To ensure that the new partition can be connected to existing partitions, the region where design variables can be updated is restricted, and one design variable within this limited region is updated per iteration during the design process. The proposed design method is applied to the formulated silencer design problem and derives various design proposals satisfying the design goal. The designer evaluates the complexity and performance of the design proposals and selects the optimal partition layout among them. The characteristics and advantages of the proposed design method are discussed and compared with the conventional silencer design method.
