@techreport{oai:ipsj.ixsq.nii.ac.jp:00216627,
 author = {重見, 和秀 and 小山, 翔一 and 中村, 友彦 and 猿渡, 洋 and Kazuhide, Shigemi and Shoichi, Koyama and Tomohiko, Nakamura and Hiroshi, Saruwatari},
 issue = {26},
 month = {Feb},
 note = {少数の観測点による単一周波数音場推定問題に対し,Helmholtz 方程式を考慮した損失関数を用いた畳み込みニューラルネットワーク(CNN)ベースの音場推定手法を提案する.CNN に基づく音場推定手法は,事前の測定データを利用することで推定対象の音環境を学習し,より高精度に音場を推定できる.しかし,従来手法では  Helmholtz 方程式を満たさない物理的に実現不可能な出力が許容されていた.これに対し,提案法では Helmholtz 方程式の差分近似に基づく損失関数を設計することで,物理的な制約を反映した CNN の学習法を可能にする.数値実験により,提案法は従来法と同程度の音場推定精度を保ちつつ,Helmholtz 方程式からの逸脱がより少ない音場を推定できることを確認した., We propose a single-frequency sound field estimation method from a small number of observations that uses a loss function based on the Helmholtz equation for training a convolutional neural network (CNN). Conventional CNN-based sound field estimation methods can enhance the estimation accuracy by using the measurements of the target sound environment. However, since they treat the sound field as a two-dimensional array, their estimated results may be physically infeasible, i.e., those results do not always satisfy the Helmholtz equation. To overcome this problem, we propose a loss function using the difference approximation of the Helmholtz equation, which enables us to encompass the physical constraint of the sound field in the CNN training. Results of numerical experiments show that the proposed method can estimate the sound fields less deviated from the Helmholtz equation while maintaining the accuracy of the sound field estimation.},
 title = {差分近似型Helmholtz 方程式に基づく損失関数を用いた深層学習による少数観測点からの音場推定},
 year = {2022}
}