Mutual Information-Weighted Principle Components Identified From the Depth Features of Stacked Autoencoders and Original Variables for Oil Dry Point Soft Sensor

Abstract

In modern chemical process control, the application of data-driven soft sensor has become increasingly extensive. Feature extraction is an important step in soft sensor. A novel feature extraction and integration method based on stacked autoencoders (SAE) and mutual information (MI)-weighted principle component analysis (PCA) was proposed to solve the loss of information on shallow depth features and original variables in neural network models. First, an SAE model was trained to extract the features of the original variables with varying depths. Second, through an MI indicator, the original variables and features with strong dependency on the outputs were selected. Then, MI was used to assign varied weights to the features and original variables, and the PCA method was used to remove any possible redundancy between the original variables and features of varying depths to obtain the principle components. Finally, the principle components were used to construct a regressor, such as a neural network. The model was first tested using the Boston housing dataset as a benchmark and then applied to the soft sensor of a constant top oil dry point. The proposed model achieved optimal results in terms of the root mean squared error and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$r$ </tex-math></inline-formula> indicators in the experiments and was thus proved feasible and useful.