Wood moisture content prediction using feature selection techniques and a kernel method

• Published in: Neurocomputing (Amsterdam) Vol. 237; pp. 79 - 91
• Main Authors: Daassi-Gnaba, Hela, Oussar, Yacine, Merlan, Maria, Ditchi, Thierry, Géron, Emmanuel, Holé, Stéphane
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 10.05.2017; Elsevier
• Subjects: Condensed Matter; Feature selection; LS-SVM; Materials Science; Moisture content; Nonlinear regression; Physics; Feature selection; LS-SVM; Nonlinear regression; Moisture content; Moisture Content
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2016.09.005

Wood is a renewable, abundant bio-energy and environment friendly resource. Woody biomass Moisture Content (MC) is a key parameter for controlling the biofuel product qualities and properties. In this paper, we are interested in predicting MC from data. The input impedance of half-wave dipole antenna when buried in the wood pile varies according to the permittivity of wood. Hence, the measurement of reflection coefficient, that gives information about the input impedance, depends directly on the MC of wood. The relationship between the reflection coefficient measurements and the MC is studied. Based upon this relationship, MC predictive models that use machine learning techniques and feature selection methods are proposed. Numerical experiments using real world data show the relevance of the proposed approach that requires a limited computational power. Therefore, a real-time implementation for industrial processes is feasible. •The prediction of moisture content for two wood chips species using the wood dielectric property is studied.•Nonlinear models are built to predict the reflection coefficient values from frequencies.•Those reflection coefficients are used as input variables of a moisture content predictive model designed using Least Squares Support Vector Machines (LS-SVM) technique and feature selection methods.•Numerical experiments using real world data show the effectiveness of the proposed methodology that requires a limited computational power.