Premium
Multi‐manifold NIRS modelling via stacked contractive auto‐encoders
Author(s) -
Zhang Jin,
Luan Xiaoli,
Liu Fei
Publication year - 2021
Publication title -
the canadian journal of chemical engineering
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.404
H-Index - 67
eISSN - 1939-019X
pISSN - 0008-4034
DOI - 10.1002/cjce.23934
Subject(s) - manifold (fluid mechanics) , heuristic , computer science , encoder , dimension (graph theory) , sensitivity (control systems) , manifold alignment , nonlinear dimensionality reduction , nonlinear system , algorithm , mathematical optimization , mathematics , topology (electrical circuits) , pattern recognition (psychology) , artificial intelligence , dimensionality reduction , engineering , combinatorics , mechanical engineering , electronic engineering , physics , quantum mechanics , operating system
Abstract Considering different operation statuses of industrial processes, a multi‐manifold learning method that incorporates multi‐manifold assumption into near‐infrared spectroscopy (NIRS) modelling is proposed in this paper. Due to the nonlinearity, high dimension and high sensitivity problems of spectral data, the stacked contractive auto‐encoder (SCAE) is introduced to extract the multi‐manifold information from the spectral data. Then, in order to detect and evaluate the current operation status from the extracted manifold information, a heuristic criterion that compares reconstruction errors of the SCAE is proposed to evaluate the positional relation between the data point and sub‐manifold surface. By this means, ordinary least squares models are used to make appropriate predictions based on the low‐dimensional space derived from the SCAE. Finally, the NIRS data of the desalination and dehydration of crude oil are investigated to demonstrate the effectiveness and the practical application of the proposed method.