标题: 应用声音讯号分类技术于引擎噪音诊断以及音讯前处理
Intelligent Classification of Sound Signals with Applications in Engine Noise Diagnostics and Audio Preprocseeing
作者: 陈孟君
Meng-Chun Chen
白明宪
Ming-Sian Bai
机械工程学系
关键字: 分类;引擎噪音诊断;独立成分分析;类神经网路;模糊类神经网路;隐藏式马可夫模型;Classification;Engine Noise Diagnostics;Independent Component Analysis;Artificial Neural Networks;Fuzzy Neural Networks;Hidden Markov Models
公开日期: 2005
摘要: 本论文乃应用声音讯号分类技术以及独立成分分析于引擎噪音诊断系统以及音讯前处理系统。
传统上,机车修护工作仰赖修护人员的经验来检视及排除故障,然而修护人员养成不易。随着现代科技进步,今日引擎之故障检修皆以电脑诊断方式进行,唯此方式对于因机件异常所产生的噪音讯号无法判断。本研究旨在以智慧型分类技术来建构一套有效率的引擎故障声讯辨识系统,以期在一般引擎故障检修范围之外,能够及早发现其它的机件异常现象,并协助维修人员正确辨识故障来源。
随着网路的发达以及数位音乐的普及,透过网路下载音乐于个人电脑已成为趋势。然而这些大量下载的音乐需要靠人工的方式分类整理,本论文提出的系统能将这些音乐自动分类,歌曲部分以歌手作区分,演奏曲则以乐器来区分。此技术亦可应用于MP3随身听,针对储存在机身中的所有歌曲,自动将歌曲系统化分类。当我们聆听歌曲时,透过此系统,随身听可以搜寻同歌手的专辑清单,或是同类乐器的乐曲,建立个人的音乐资料库。
独立成分分析技术可从混合的资料当中分离出相互之间互为统计独立的成分。本研究将歌曲的左右声道分别当作两个混合声源,经过独立成份分析验证后发现歌曲中音乐成分可被分离,但是无法抽取出人声的部分。若用此技术来萃取背景噪音中之语音讯号,亦无法去除所有的背景噪音,但可以降低背景噪音,突显出语音的部分。
A processor that integrates various intelligent classification and preprocessing algorithms is presented in this thesis. Classification algorithms including the Nearest Neighbor Rule (NNR), the Artificial Neural Networks (ANN), the Fuzzy Neural Networks (FNN), and the Hidden Markov Models (HMM) are employed to classify and identify engine noise, singers and instruments. Audio features in the time and frequency domains are extracted and preprocessed prior to classification. A training phase is required to establish a feature space template. This is followed by a test phase, where the audio features of the test data are calculated and matched with the feature space template. In addition to audio classification, the proposed system provides several Independent Component Analysis (ICA)-based preprocessing functions for blind source separation, voice removal, and noise reduction. The proposed techniques were applied to process various kinds of audio program materials. The results reveal that the performance of methods are satisfactory, but varies with algorithm and program material used in the tests.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009314580
http://hdl.handle.net/11536/78555
显示于类别:Thesis