Stationary Feature Extract and Process Monitoring Based on Autoencoder

To monitor non-stationary industrial processes, a stationary feature extract method is proposed, which can be regarded as a nonlinear version of traditional stationary subspace analysis. To enhance feature extraction, the linear transformation in stationary subspace analysis is replaced by an autoencoder. To extract the stationary features, an additional term is introduced into the loss function of the autoencoder, which measures the consistency of the feature distributions over different time periods. The convex hull based on the stationary features is used to construct a monitoring statistic for the alarm of abnormal working conditions. A case study based on an ironmaking process illustrates that the proposed method can effectively achieve abnormality detection.