Monitoring of complex profiles based on deep stacked denoising autoencoders

Abstract

Profile monitoring remains an interesting issue in statistical process control (SPC). Although there have been considerable researches devoted to analysis of profile data, the challenges concerning the monitoring of complex profiles (e.g., multivariate profiles, nonlinear profile, autocorrelated profiles) is yet to be addressed well. The high-dimension explanatory variables and autocorrelation generally affect effectiveness of those regular profile monitoring models and could cause many false alarms. Recent years have witnessed remarkable successes of deep learning techniques in visual and acoustic studying fields. In this paper, a deep learning model known as stacked denoising autoencoders (SDAE) is developed for complex profiles modeling and monitoring. Three control charts based on the SDAE model are further developed for abnormal detection of complex profiles. Comparison between the proposed method and other typical methods is implemented to illustrate effectiveness of the proposed method in five representative profiles. Finally, a real dataset is further utilized to demonstrate the effectiveness of the proposed method in agriculture fields. The experimental results illustrate the effectiveness of the SDAE-based method on complex profiles monitoring. This paper provides an inspiration for using deep learning techniques to monitor complex profiles.