Study Of Statistical Process Monitoring Method Based On Auto-Encoder

Process monitoring is a very important problem in the industrial process,which plays an important role in ensuring the safe operation of the industrial process.Feature extraction is one of the core problems in process monitoring,thus how to extract the essential characteristics of industrial data is the key to improve the accuracy of fault detection.The feature extraction method of process monitoring has made good progress.However,there are still some problems to be solved,such as the nonlinear and robustness problem,the problem of using fault samples and limited training samples.The Auto-Encoder is a typical unsupervised machine learning method,which can extract nonlinear and robust features automatically.Because of the flexible structure,it's easy to improvement,which is suitable to solve the problem of complex industrial process.In view of the problems of feature extraction in process monitoring,this paper introduces AE model in feature extraction of process monitoring.Mainly include the following:(1)Aiming at the problem of nonlinearity and robustness in feature extraction of data,a process monitoring algorithm based on Auto-Encoder is proposed.The constructing process and principle of the detection statistic is introduced,based on which the kernel density estimation method and the algorithm flow are given.The method is applied in TE process,which can extract more robust features from industrial data to improve the effect of process monitoring and the accuracy of fault detection.(2)In terms of the inherent structure and manifold characteristics of data,manifold learning algorithm is integrated into the Auto-Encoder,thus a feature extraction method based on Embedding-AE is proposed.The method can influence the distribution of the input sample using the nearest neighbor relation between the normal sample and the fault sample,thus improving the clustering ability of feature space and residual space as well as the effect of process monitoring.(3)In term of the problem of limited training samples,this paper presents a statistic process monitoring method based on the Adversarial Auto-Encoder.With the idea of against training,it can learn the distribution characteristics of samples to improve the robustness and generalization ability of the model.Finally,the experimental results show that the feature extraction method based on Auto-Encoder model can improve the effectiveness of process monitoring.