機械学習による時間変動を伴うエンジン音の特徴量抽出

大島　遥汰（中央大学大学院）・田辺　総一郎・戸井　武司（中央大学）

畳み込みニューラルネットワークと準再帰型ニューラル ネットワークを組み合わせた機械学習モデルに基づき時間周波数解析画像を分類し，ニ ューラルネットワークの中間層抽出手法である勾配加重クラス活性化マッピングを用いることにより，エンジン音の各周波数成分における時間変動を伴う特徴を可視化する．

A Study on Machine Learning and Preprocessing of Parts Operating Sound Data to Establish Quantitative Standards Evaluation for Automotive Parts Noise

Sang Heon Wang･Nak Kyoung Kong･Dong Eun Cha･Ho Wan Jang (Hyundai Motor)

As the electrification of automobiles and the mobility revolution, the noise of the interior space is becoming important. Due to these factors, the standards for operating sound of parts are also becoming stricter. But in most cases, the standards depend on emotional evaluations that human hear and judge. In this paper, we propose a method to establish quantitative standard evaluations for emotional evaluations as a process for performing machine learning after pre-processing frequency analysis of operating sound. Using this method, it is possible to establish quantitative standard evaluations for the operation sound of parts without human intervention.

Prediction of High Frequency Noises of an EV using Machine Learning
-Machine Learning for the Prediction of Both Structure-Borne and Airborne Noises-

Ji Woo Yoo･Yong Dae Kim (Hyundai Motor)･Kwangsoo Yoon･Chanhee Jeong･Hyosik Jung･Dohyeon Oh (Hexagon Korea)

Whine noise of an electric vehicle occurs due to electro-magnetic force or gears in an e-motor, which lies usually at higher frequencies. This paper deals with a proof of concept for machine learning methods predicting the whine and shows the relevant procedure of machine learning. The study investigates two topics: efficient encapsulation on the e-motor to handle acoustic radiation (airborne), and prediction of whine noise in the cabin (structure-borne). Particularly a novel superposition method is proposed to quickly realize various combination of encapsulation. It is shown that machine learning could be a good alternative to predict high frequency whine but also reveals weak points.