基于机器学习方法的电机异音检测研究

发布时间：2018-12-12 13:36

【摘要】：现代工业生产以及家用电器都离不开各式各样的电机,人们在重视电机性能的同时也希望降低电机转动而产生的噪音。目前,工厂对异音电机识别是通过对产线工人进行培训,用人耳听音的方式实现对生产线上大批量小型电机的音质检测,而大量单调、重复听音劳动致使听觉疲劳影响主观判断,导致异音电机混入正常样本流入市场,对公司的经济和声誉会造成不可挽回的损失。因此,实现产线异音电机的自动化检测对电机产业的发展具有十分重要的意义。本文针对电机声音信号的统计特性及其人工质检的特点,利用声传感器技术代替人耳实现对电机声音信号的采集,这种非接触式的测量方式正符合产线测试设备简单、高效等要求。在电机平稳运行情况下采集声信号,根据人耳听觉特性,从电机声音的平稳性来分析电机的频谱。因为人耳对相位不敏感,只需要对幅度谱来分析电机异音的特性,为了突出特征差异的绝对化,本文采用主成分分析法对电机的声信号进行数据压缩、维数来实现电机声信号提取特征。又考虑电机声信号中可能存在非平稳成分,将小波变换引入电机异音检测研究以更精确分析电机的时频特性,利用小波包分解获取电机声信号的各频段系数,根据奇异值分解的特征矢量对特征的贡献量进行降噪、重构并映射到特征矢量所张成的状态空间,实现对电机声信号的特征提取。同时,电机声信号进行小波包分解得到相互正交的频带,其能量没有损耗且蕴含着丰富的特征信息,将电机的特征映射到能量分布的子空间中,并以归一化能量构建特征矩阵实现对异音电机的特征提取。文中根据上述方法分别实现对电机声信号进行特征提取送入分类器进行训练,取得了良好的效果。考虑到产线上异音样本量少、获取困难,个体差异造成异音等问题难以分析,且电机异音形成过程异常复杂,应用支持向量机一类学习这种新的机器学习方法实现对异音电机检测。该方法以正常电机样本为基础建立的质检判别函数,并不需要异音样本,避免了其它分类算法要求训练样本类别全面和覆盖广泛的条件。文章最后通过大量正常电机样本训练并以异音电机样本进行验证,得出对异音电机的识别率能满足工厂的需求,达到了预期目标。
[Abstract]:Modern industrial production and household appliances can not be separated from a variety of motors, people pay attention to the performance of the motor, but also want to reduce the noise generated by motor rotation. At present, the factory to the abnormal sound motor identification is through the production line worker to carry on the training, uses the human ear to listen the sound way to realize to the production line massive batch small electric machine sound quality detection, but a large number of monotonous, Repeated hearing labor causes hearing fatigue to affect subjective judgment and leads to abnormal motor mixed into the normal sample into the market, which will cause irreparable loss to the company's economy and reputation. Therefore, it is very important for the development of the motor industry to realize the automatic detection of the production line abnormal sound motor. In this paper, according to the statistical characteristics of motor sound signal and the characteristics of artificial quality inspection, the sound sensor technology is used to replace the human ear to realize the acquisition of motor sound signal. This non-contact measurement method is in line with the simple testing equipment of the production line. High efficiency, etc. The sound signal is collected under the condition of the motor running smoothly, and the frequency spectrum of the motor is analyzed according to the hearing characteristics of the human ear and the stability of the motor sound. Because the ear is not sensitive to the phase, only the amplitude spectrum is needed to analyze the abnormal sound characteristics of the motor. In order to highlight the absolute characteristic difference, the principal component analysis (PCA) is used to compress the acoustic signal of the motor. Dimension to achieve motor acoustic signal extraction features. Considering that there may be non-stationary components in the motor acoustic signal, wavelet transform is introduced into the research of the abnormal sound detection of the motor to analyze the time-frequency characteristics of the motor more accurately, and the coefficients of each frequency band of the motor acoustic signal are obtained by wavelet packet decomposition. According to the feature vector of singular value decomposition (SVD), the noise is reduced, the feature is reconstructed and mapped to the state space of Zhang Cheng, and the feature extraction of the acoustic signal of the motor is realized. At the same time, the acoustic signal of the motor is decomposed by wavelet packet to obtain the orthogonal frequency band, which has no energy loss and contains abundant characteristic information. The characteristics of the motor are mapped to the subspace of the energy distribution. The feature matrix is constructed with normalized energy to extract the feature of the abnormal sound motor. According to the above methods, the feature extraction of the motor acoustic signal is carried out and sent to the classifier for training, and good results are obtained. Considering that it is difficult to analyze the problems such as the small sample size of abnormal sound on the production line, the difficulty of obtaining the abnormal sound, and the difficulty of analyzing the abnormal sound caused by individual differences, and the abnormal sound formation process of the motor is extremely complex, A new machine learning method, support vector machine (SVM), is applied to detect abnormal sound motors. Based on the normal motor samples, this method establishes the quality inspection discriminant function, and does not need the abnormal sound samples, thus avoiding the condition that other classification algorithms require the training samples to be comprehensive and have extensive coverage. In the end, through the training of a large number of normal motor samples and the verification of the abnormal sound motor samples, it is concluded that the recognition rate of the abnormal sound motor can meet the needs of the factory and achieve the expected goal.
【学位授予单位】：五邑大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TM301.4;TP181

【参考文献】