使用麥克風陣列之強健型警報器聲源辨識演算法

Abstract

本文提出結合麥克風陣列、適應性濾波器及聲源辨識演算法的警報器偵測系統，可接收多點不同方位之警報器聲源來進行辨識。近年來，物聯網興起，於居家電器裝置結合網路介面進行連網功能，可達到多系統整合以便家庭遠端智慧監控，在現有警報器上加裝網路介面需要改裝原來的設備，並且每一警報器都將增加額外的能量損耗，將會使整體成本提升。本文提出無須額外改造設備及節省成本的系統，透過聲音的傳遞達成傳輸警報至監控系統的功能，用於家中多警報器之偵測，例如：門鎖防盜警報、瓦斯警報、一氧化碳及煙霧警報器。 文中提出以麥克風陣列技術結合適應性空間濾波器獲得去雜訊聲源，對此聲源頻域與時域做特徵擷取並建立模型，依據警報器週期、主要頻率、頻率分布、過零率及工作週期等特徵，持續判斷環境中是否警報器響起。文章最後，演算法以不同種類的穩態及非穩態噪音於不同訊噪比情況下測試，針對警報音命中率和非警報音命中率進行分析。藉由實際警報器及麥克風陣列之環境配置下進行實驗模擬，利用最小無失真響應的波束形成器達到聲源強化的效果；其中空間前處理所需的角度資訊，事先運用轉移函式比值法求得目標聲源在空間中相對轉移函式做為角度資訊。

This thesis presents an alarm detection system which combines the microphone arrays, adaptive filter and sound source identification algorithms to receive alarms sound from different orientations and then identify the alarm types. In recent years, the rise of Internet of things (IoT) has added the internet connectivity to various home appliances which allows remote intelligent monitoring and control. However, the installation of network interface to existing alarm systems requires the modifications of original device and each alarm system will consume additional energy too. As a result, the overall cost to upgrade the existing system will be high and not worthwhile. Hence, this thesis proposes a system that eliminates the need for additional modifications of equipment and delivers cost savings. The proposed system has utilized the transmission of alert sound from different alarm systems into a monitoring system for detection of multi-alarm devices at home such as door alarms, gas detector alarms, carbon monoxide and smoke detector alarms. In this thesis, a method of using microphones array technology combined with adaptive spatial filter is proposed to obtain the sound source with noise removed. After that, the frequency domain and time domain of this sound source were extracted and modeled. The proposed system can constantly detect the existence of various alarms sound in the environment and identify them based on the characteristics like alarm period, main frequency, frequency distribution, zero-crossing rate and duty cycle. At last, the proposed algorithm was tested under the different kinds of steady-state noise and non-steady state noise with different signal-to-noise ratios. The final results were obtained and analyzed based on the hit rate of alarm sound and non-alarm sound. The experimental setup included the actual alarm systems and microphone arrays where the minimum variance distortionless response (MVDR) beamformer was used for sound source enhancement. The directional information between the sound source and microphones in any given space was required by the spatial preprocessing. This directional information was obtained by calculating the relative transfer function (RTF) using the transfer function ratio (TFR) method.