旋转设备数据不平衡问题的数据生成方法

Study of Data Generation Methods for Rotating Equipment Data Imbalance Problem

在旋转设备运行状态监测及故障识别时,采集的样本多为无故障样本,而故障样本较少,这种数据分布的不平衡会严重影响分类器识别的准确性。针对此问题,提出了一种少数样本数据生成方法,即基于傅里叶变换与皮尔逊系数优化的生成对抗神经网络(Fourier-Pearson generative adversarial networks,简称FP-GAN)模型。通过对故障少数样本的扩充,提高故障诊断训练和识别的准确性。首先,使用傅里叶变化得到信号频域的单边谱,使用GAN网络生成信号频域;其次,通过皮尔逊相关系数对生成的数据进行优化;最后,通过傅里叶逆变换获得更接近真实数据的生成数据。仿真和实验数据验证表明,基于FP-GAN生成的数据样本在时域特征、时域统计特征以及分类器分类结果方面都能较好地与已有实际数据融合,可以对小样本数据进行增强,能有效解决数据不平衡问题。

In operation status monitoring as well as fault identification for rotating equipment,collected fault sam-ples are mostly fault-free samples with few fault samples,and this imbalance in data distribution can have a serious impact on the accuracy of classifier identification.To address this problem,a minority sample data generation method is proposed based on Fourier transform and Pearson correlation coefficient optimization of generative ad-versarial neural network(FP-GAN).Such method is able to improve the accuracy of fault diagnosis training and recognition by expanding the fault minority samples.Fourier transform is used for getting the single side spectrum of the signal frequency domain,and GAN network is used for generating the signal frequency domain.After that,the generated data is optimized by Pearson correlation coefficient to obtain data closer to the real data by inverse Fourier transform.The validation of simulation and experimental data show that the data samples generated based on FP-GAN can be better integrated with the existing actual data in terms of time-domain features,time-domain statistical features and classifier classification results,which can effectively solve the data imbalance problem.