摘要
针对煤矿带式输送机异常声音样本缺少导致训练模型难以进行异常识别的问题,提出一种基于卷积自编码器(CAE)的煤矿带式输送机异常声音检测方法。首先,采集煤矿带式输送机托辊、减速机、电动机正常运行的声音信号,通过WebRTC降噪算法过滤信号中的背景噪声,计算降噪后信号的梅尔频率倒谱系数(MFCC),获得设备正常运行的音频特征并输入到CAE中进行训练,得到训练好的CAE及重构的正常运行音频特征。其次,将正常运行音频特征和重构的正常运行音频特征输入均方损失函数(MSELoss),得到重构误差,并取重构误差最大值作为正常运行音频特征的重构阈值。然后,采集待检测的煤矿带式输送机托辊、减速机、电动机运行的声音信号,经WebRTC降噪、MFCC特征提取后输入到训练好的CAE,获得重构的待检测音频特征,将待检测音频特征与重构的待检测音频特征输入MSELoss,得到待检测音频的重构误差。最后,将待检测音频的重构误差与正常运行音频特征的重构阈值进行比较,若前者大于后者,则判断煤矿带式输送机存在异常。实验结果表明:在没有异常声音样本参与训练的情况下,该方法在带式输送机托辊、减速机、电动机运行声音数据集上的检测精确率分别达92.55%,94.98%,93.60%,单组声音检测时间为1.230 s,实现了检测精度和检测速度之间的平衡。
To address the issue of insufficient abnormal sound samples for coal mine belt conveyors,which makes it difficult for training models to recognize anomalies,an abnormal sound detection method for coal mine belt conveyors based on Convolutional Autoencoder(CAE)is proposed.First,sound signals from the normal operation of the belt conveyor's idlers,reducer,and motor were collected.Background noise in the signals was filtered using the WebRTC noise reduction algorithm,and Mel-Frequency Cepstral Coefficients(MFCC)were calculated from the denoised signals to obtain audio features of normal operation.These features were then input into the CAE for training,resulting in a trained CAE and reconstructed audio features of normal operation.Next,the normal operation audio features and the reconstructed normal operation audio features were input into the Mean Squared Error Loss function(MSELoss)to obtain the reconstruction error,with the maximum reconstruction error set as the reconstruction threshold for normal operation audio features.Then,sound signals from the operation of the coal mine belt conveyor's idlers,reducer,and motor to be inspected were collected.After noise reduction using WebRTC and MFCC feature extraction,they were input into the trained CAE to obtain the reconstructed audio features of the inspected samples.The inspected audio features and the reconstructed audio features were then input into the MSELoss to calculate the reconstruction error of the inspected audios.Finally,the reconstruction error of the test audio was compared with the reconstruction threshold of normal operation audio features.If the former exceeded the latter,the coal mine belt conveyor was identified as abnormal.Experimental results showed that,without abnormal sound samples involved in training,the proposed method achieved detection accuracies of 92.55%,94.98%,and 93.60%for the idlers,reducer,and motor,respectively.The detection time for a single sound sample was 1.230 seconds,achieving a balance between detection accuracy and speed.
作者
申龙
单浩然
裴文良
杨贵翔
王永利
SHEN Long;SHAN Haoran;PEI Wenliang;YANG Guixiang;WANG Yongli(CITIC Heavy Industries Kaicheng Intelligence Equipment Co.,Ltd.,Tangshan 063020,China)
出处
《工矿自动化》
北大核心
2025年第2期100-105,共6页
Journal Of Mine Automation
基金
河北省高端装备制造技术创新专项项目(23311805D)。
作者简介
申龙(1987-),男,河北唐山人,现从事煤矿智能装备方面的工作,E-mail:cediy2088@126.com。
