1Mel倒谱系数的提取.doc

基于MFCC特征提取的故障预测与评价方法* 常 飞1,2，乔 欣1,2，张 申*1,2，许华栋1,2 （1．中国矿业大学 物联网（感知矿山）研究中心，江苏 徐州 221008；2. 矿山互联网应用技术国家地方联合工程实验室，江苏 徐州 221008） failure prediction and evaluation based on MFCC feature extraction CHANG Fei1,2，QIAO Xin1,2，ZHANG Shen*1,2，XU Hua-dong1,2 （1. CUMT-IoT Perception Mine Research Center，Xuzhou Jiangsu 221008 ，China；2. The National and Local Joint Engineering Laboratory of Internet Application Technology on Mine , Xuzhou Jiangsu 221008 ，China） For the accurate prediction problems of gradual failures, the feature extraction method for fault prediction based on Mel Frequency Cepstral Coefficients (MFCC) was put forward. Firstly, the acoustic sensor which was installed on the monitored object get the current machine running sound signal. Then the sound signals was pretreated and transformed by the way of Mel to obtain their MFCC feature vectors. Finally, SVM classifier was used to cluster the machine running test data and the sample data from the normal running of the machine, which was able to give the healthy evaluation of the machine. The healthy degree of a Haima M3 engine was tested when the crash sound, tapping sound and the boom was existed, and the anti-noise performance of collected sound signals was analyzed when the noise was existed. The experimental results indicated that MFCC feature values ??can be the feature vectors for machine health prediction and the method had advantages in forecasting accuracy and robustness. Key words: Failure prediction; MFCC; Cluster analysis; SVM 引言 机器故障预测[1]是指根据机器当前或者历史的运行时的状态预测出未来一段时间内机器的健康状态与是否存在渐变故障，其预测的准确性是机器日常维护、正常运行、安全生产等重要保证，直接关系到经济效益与维护成本，因此是故障诊断的重要课题之一。 与故障诊断相比，故障预测能够事先对可能发生渐变故障的机器做出评估，为有计划的检修与维护提供了前提。声音信号是对多数机器都很敏感，对于正常运行的机器，其每次启动、运行、结束的声音特征变化不大。但是当机器出现渐变故障时，机器的声音特征就会出现剧烈的变化，因而，通过声音特征的提取对机器故障进行预测是可行的。 目前，在机器故障预测的特征提取中，时域振动烈度的特征提取是机器预测的基本参数，通常通过传感器采集到机器的振动烈度，然后经过分析找出符合该机器正常运行的正态分布，当下一次采集的信号特征符合正态分布，且越靠近均值时，就表明机器越正常。频域方面，主要分析频带宽度、或者进行小波变换后的频带能量[2]，统计机器正常运行时各个频带的能量分布作为故障预测的样本特征向量[3]，接着将提取到的机器特征作为测试样本，再通过各种