采用编码输入的生成对抗网络故障检测方法及应用

Fault detection method and its application using GAN with an encoded input

针对传统基于生成对抗网络的故障检测方法中,生成器输入使用随机噪声,不包含训练集中有效信息造成模型检测效果不够理想的问题,提出一种采用编码输入的生成对抗网络故障检测策略。通过引入自编码器,基于最小化重构误差构建隐变量空间,将降维后的隐变量信息作为生成器输入以提升生成对抗网络的训练效果;进一步考虑故障检测方法中基于生成器的统计量计算成本高和对离群点敏感的问题,计算待测样本经编码后的隐变量到训练集隐变量空间中心点的曼哈顿距离,并作为新统计量进行故障检测。将所提故障检测方法用于田纳西伊斯曼过程及实际的磨煤机过程,本文方法较传统生成对抗网络故障检测在田纳西伊斯曼过程上报警率提升了13%,在磨煤机过程上各统计量报警率均得到了显著提升且本文所提统计量将传统方法中针对生成器的统计量大大降低了检测用时,从而验证了方法的有效性和性能。

In a traditional fault detection method based on a generative adversarial network,random noise is used as the generator input,which does not contain the effective information of training sets and causes unsatisfactory fault detection by the model.This paper proposes a generative adversarial network fault detection strategy using an encoded input.By introducing an autoencoder based on minimizing the reconstruction error,a latent variable space is constructed.The latent variable information after dimensionality reduction is used as the generator input to improve the training effect of the generated confrontation network.Furthermore,a generator-based statistic has drawbacks such as high computational cost and sensitivity to outliers.The Manhattan distance is calculated beginning from the encoded hidden variable of the sample to be tested to the center of the hidden variable space of the training set.This distance is then used as a new statistic for fault detection.The proposed fault detection method is used in the TE and actual coal pulverization processes.Compared with the traditional GAN fault detection,the alarming rate of the TE process increases by 13%.The alarming rate of all statistics in the coal pulverization process also improves considerably.The proposed statistics greatly reduce the detection time for generators in traditional methods,which validates their effectiveness and performance.