仿聽覺處理之聲音異常偵測

Abstract

本論文針對居家機器人之應用，提出一個仿聽覺處理之模型，以解決居家異常聲音偵測之問題。首先，以類神經網路模型模擬大腦聽覺皮質區，取得語意抽象特徵，接著，在仿大腦前額葉模型中，利用語意符號的運算，在基底核區域辨別出正異常聲音。此外，本論文亦比較不同異常偵測方法之數值結果，所提出之方法能夠在測試集中取得最佳的結果(0.972 AUC)，且僅需極少的運算時間即可完成偵測，相當適合作為居家機器人之聽覺應用模組。

This thesis focuses on the application of household robot and provides a humanoid auditory processing model to resolve the problem of acoustic anomaly detection in household environment. First, the proposed model uses deep neural networks to imitate auditory cortex in human brains, in order to extract abstract semantic features. Then, in the SPA module which imitates prefrontal cortex, the anomaly is detected at basal ganglia area by using symbolic computations. In addition, different anomaly detection methods are compared in this thesis. Our proposed method gets the best result on test set, with 0.972 AUC. Meanwhile, it takes less computational time to detect the anomaly, so it is very suitable for the application to a household robot.