Study Of Statistical Process Monitoring Method Based On Data Deep Feature Extraction

Real-time monitoring production process is very necessary in modern industrial system.Statistical process monitoring is a method of fault detection.Feature extraction is one of the core problems in process monitoring,thus how to extract the inherent deep features of industrial data is the key to improve the accuracy of fault detection.However,there are still some problems to be solved,such as the nonlinear and robustness problem,the problem of using limited samples,the autoencoder can extracts nonlinear and robust feature through reconstructing input,however,it’s simple structure limits the ability of extracting feature.In this paper,several methods of statistical process monitoring are proposed based on autoencoder aiming at these problems:(1)Considering the manifold characteristic of industrial data,we proposed a statistical process monitoring method based on EmbeddingAE,embedding the original data’s nearest neighbor relationship into the autoencoder through manifold learning algorithm which can change the data’s feature space distribution,imporving model’s discriminant ability,so as to improve the effect of process monitoring.(2)Aiming at the problem of limited data,a new idea that processing monitoring based on distribution is proposed,generative adversarial network is an effective deep generative model and can learn data distribution well,so,we propose a statistical process monitoring method based on adversarial autoencoder(GAN variant),in addition,we targeted put forward two metric to evaluate for generative model.(3)Denoising autoencoder forces the corrupted input to reconstruct original input,which can extract more robust feature,based on the idea,we introduce denosing adversarial autoencoder which keep adversarial training of AAE,so matching the posterior distribution to the prior,meanwhile,representation may be further improved by learning to recover clean input samples from corrupted ones,so,the statistical process monitoring model based on the learned distribution from denoising adversarial autoencoder is more reasonable and effective.