AE-SFA model with Bayesian inference for industrial process monitoring

Process monitoring methods based on fusion statistics are widely used in industrial processes. In this paper, a process monitoring model based on the combination of autoencoder and slow feature analysis (AE-SFA) is proposed. First, the defined autoencoder network is pretrained and the features of the training data are obtained and the monitoring statistics are constructed. Second, for the data residuals of the autoencoder, the residual data features are further extracted using multivariate statistical slow feature analysis to compute the slow feature static statistic. The threshold of the statistic is determined by the kernel density estimation method. Finally, the Bayesian inference formula is utilized to construct the integrated monitoring statistic. And the feasibility and validity of the method are verified in the Tennessee-Eastman benchmarking process.