基于声谱图和SVM的煤矿瓦斯和煤尘爆炸识别方法  

Recognition method of coal mine gas and coal dust explosion based on sound spectrogram and SVM

作　　者：孙继平[1] 余星辰 王云泉[2] SUN Jiping;YU Xingchen;WANG Yunquan(School of Mechanical Electronic and Information Engineering,China University of Mining and Technology—Beijing,Beijing 100083,China;China National Coal Group Corporation,Beijing 100011,China)

机构地区：[1]中国矿业大学(北京)机电与信息工程学院,北京100083 [2]中国中煤能源集团有限公司,北京100011

出　　处：《煤炭科学技术》2023年第2期366-376,共11页Coal Science and Technology

基　　金：国家重点研发计划资助项目(2016YFC0801800)。

摘　　要：为提高煤矿瓦斯和煤尘爆炸感知准确率,提出了基于声谱图和SVM的煤矿瓦斯和煤尘爆炸感知方法:在煤矿井下重点监测区域安装矿用拾音器;实时监测煤矿井下设备工作声音及环境音;将采集到的声音提取由MFCC构成的声谱图,通过计算声谱图的灰度共生矩阵得到0°、45°、90°、135°的能量、熵、对比度、相关性,分别求其均值和标准差作为声谱图图像纹理特征,构成该声音的特征量,输入到SVM中建立煤矿瓦斯和煤尘爆炸声音识别模型;对待测声音同样提取其MFCC声谱图图像纹理特征,输入到训练好的识别模型中进行声音识别分类;并进行了试验验证。首先,提取了采掘工作面设备运行、瓦斯和煤尘爆炸等不同声音的MFCC特征值,分析了不同声音的MFCC特征值分布情况;提取不同声音的MFCC声谱图,分析了不同声音的声谱图的特征参数:能量、熵、对比度、相关性的均值和标准差,可见通过提取MFCC声谱图的灰度共生矩阵特征参数构成的特征量可有效表征声音信号;其次,将待测声音输入建立的识别模型中,完成识别分类。结果表明:所提方法的识别率达到95%,整体识别性能高于其他算法;最后,通过贝叶斯参数优化试验结果可知,优化后的SVM识别模型的召回率、识别率分别提高10%、3%,优于优化前的识别模型,能够满足煤矿瓦斯和煤尘爆炸感知和报警需求。In order to improve the detection accuracy of coal mine gas and coal dust explosion,recognition method of coal mine gas and coal dust explosion based on the sound spectrogram and SVM was proposed:installing mining pickups in key monitoring areas of under-ground coal mines,monitoring of working sounds of underground coal mine equipment and environmental sound in real-time.The collec-ted sound was extracted from the sound spectrogram composed of MFCC,and the energy,entropy,contrast,and correlation of 0°,45°,90°,and 135°were obtained by calculating the gray co-generation matrix of the sound spectrogram.Calculated their mean and standard deviation as the texture features of the sound spectrogram image,which constituted the feature quantity of the sound,which brought to SVM to establish the sound recognition model of coal mine gas and coal dust explosion.For the sound to be tested,the texture features were also extracted and input to the trained recognition model for sound recognition and classification,which has passed the test verifica-tion.Firstly,the MFCC feature values of different sounds,such as mining equipment operation,gas and coal dust explosion were extrac-ted,and the distribution of MFCC feature values of different sounds was analyzed;the sound spectrograms of different sounds were extrac-ted,and the mean and standard deviation of the sound spectrograms of different sounds were analyzed by analyzing the feature parameters of energy,entropy,contrast,and correlation,and the grayscale co-generation matrix feature parameters of the sound spectrograms consti-tute the feature quantities that can effectively characterize the sound signals.Secondly,the sound to be measured was input into the recog-nition model established to complete the recognition and classification.The results show that the recognition rate of the proposed method reaches 95%,and the overall recognition performance is higher than others models.Finally,the experimental results of Bayesian paramet-er optimization show that the recall rate and

关 键 词：瓦斯和煤尘爆炸 声谱图 灰度共生矩阵 声音识别 图像识别 支持向量机 

分 类 号：TD76[矿业工程—矿井通风与安全] X936[环境科学与工程—安全科学]