Process monitoring is a crucial part of industrial manufacture, which plays a significant role in ensuring the equipment operation safety and improving the product quality. However, conventional algorithms tend to have low robustness and cannot solve the problem of nonlinearity and extract the multimodality information at the same time. This paper proposes an improved process monitoring model based on the existing algorithms to solve the problems above. To be specific, denoising autoencoder (DAE) is utilized to decompose and reconstruct the data with artificially added Gaussian noise, which can analyze the nonlinear process and improve the robustness of the model. Gaussian mixture model (GMM) then is applied to accurately describe the multivariate normal distribution of the low dimensional features, which can solve the problem of the inability to extract multimodal information. Finally, the Bayesian inference distance (BID) statistic is calculated, and then its control limit is obtained by introducing kernel density estimation (KDE). The case studies on the Three-phase Flow Facility process verify the effectiveness of the proposed model.
An improved denoising autoencoder for multimodal industrial process monitoring combined with Gaussian mixture mode
12.10.2022
1332838 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch