张红英

摘要：为解决以往警报器质检依赖人工听音的问题，研究了音频自动检测方法。通过对比分析正常警报器音频和异常警报器音频的声谱图，分别总结出4类音频异常的频谱特点，提出了以幅频特性曲线峰值点的频率和幅度为主要元素的音频特征向量设计方法。为解决异常音频样本数量较少的问题，提出了三种音频处理方法。基于BP神经网络原理设计并训练了警报器音频识别模型，准确率达99.8%，异常音频召回率达100%。基于Windows系统开发了声光警报器音频自动检测系统，实际使用效果良好。

关键词：声谱图;机器听觉;异音检测;神经网络;特征提取

中图分类号：TP23      文献标识码：A

文章编号：1009-3044（2021）13-0024-04

Abstract： In order to solve the problem that the previous alarm quality inspection relies on manual listening， an automatic audio detection method is studied. By comparing and analyzing the sound spectrums of normal alarm audio and abnormal alarm audio， the spectrum characteristics of four types of abnormal alarm audio are summarized， and the design method of audio feature vector is proposed， which takes the frequency and amplitude of peak points of amplitude-frequency characteristic curve as the main elements. In order to solve the problem of small number of abnormal audio samples， three audio processing methods are proposed. Based on the principle of BP neural network， an alarm audio recognition model is designed and trained， with an accuracy of 99.8% and an abnormal audio recall rate of 100%. Based on Windows system， an automatic audio detection system for acousto-optic alarm is developed， which has a good practical effect.

Key words： sound spectrums; machine hearing; heterophony detection; neural network; feature extraction

1 背景

聲光警报器是一种通过声音和光来向人们发出示警信号的装置，广泛应用在各类场所，是保证生产和人身安全的重要设备。在警报器的制造过程中，需要检查警报声音是否异常，以往主要采取人工听音的方式来进行。由于警报声音的声压级很高，1米内可达到100dBA以上，长期之下会对检验人员的健康造成不利影响，且人工主观判断也存在一定的随意性。因此，研发声光警报器异音检测系统具有现实意义。

受限于应用范围，关于声光警报器音频自动检测技术的研究较少。文献[1-2]中研制了自动测试发声元件蜂鸣器的频率的装置，基本原理是：首先对蜂鸣器音频进行快速傅里叶变换，得到幅频特性曲线，然后提取幅频特性曲线中峰值点对应的频率，最后判断峰值频率是否在正常的频率范围内。……