针对工业现场电机故障样本不足、数据严重不平衡的问题,提出了支持向量数据描述(Support Vector Data Description,SVDD)、自编码器(Auto Encoder,AE)和孤立森林(Isolation Forest)三种分类算法融合的算法,实现了电机声音异常监测.实验数据采自工业现场,根据拾音器采集的声音信息,提取梅尔频率倒谱系数(Mel Frequency Cep-stral Coefficents,MFCC)、梅尔频谱、短时能量、过零率等反映电机运行状况的特征,然后进行特征筛选,去除冗余特征,挑选出最优的特征子集送入单分类器,实现电机的异常监测.实验结果分析表明,提出的方案准确率达到98%,与基于生成对抗单分类网络(Generative Adversarial Single Classification Network,GACN)方案相比提高约 5%.
Motor Sound Anomaly Detection Based on Audio Data Feature Extraction and One-class Classification
In view of insufficient fault samples and serious data imbalance of industrial on-site motors,this study estab-lished an integrated algorithm for motor sound anomaly detection by combining three types of classification algorithms,i.e.,support vector data description(SVDD),auto encoder(AE)and isolation forest. First features about motor opera-tion status,such as Mel frequency cepstral coefficents(MFCC),Mel spectrogram short-term energy,and zero crossing rate,were extracted from sound information collected by factory on-site pickups.Then feature screening was carried out to remove redundant features and select the optimal feature subset which was input into a one-class classifier,thereby re-alizing motor sound anomaly detection.The proposed algorithm was indicated by experiment to achieve an accuracy rate reaching 98%,about 5%higher than the scheme based on generative adversarial single classification network(GACN).