Novel SiameseAE Network for Industrial Process Slow Feature Extraction and Soft Sensing Applications

Abstract

Industrial process data are time series data with strong dynamics and nonlinearities and are based on temporal slowness. For industrial soft sensor modeling, it is critical to extract the slowness features from strongly nonlinear and dynamic process variables. To address these issues, a method based on a Siamese autoencoder and long short-term memory (SiameseAE-LSTM) network is proposed. First, a SiameseAE network that considers both nonlinearities and dynamics is proposed for essential slow feature extraction. Then, a slow feature matrix is extracted and augmented with historical slow features and quality variables, ensuring the validity of slow features. This matrix contains more process information. Finally, an LSTM network is employed to capture the long-term dependencies from the feature matrix to build a soft sensor model. The effectiveness and superiority of the proposed method are demonstrated on a numerical example and two industrial process cases.