Un enfoque de machine learning para la predicción de la calidad de tableros contrachapados

Abstract

38 Because of the impact on productivity and cost reduction, decision making in industrial processes is one of 39 the most required aspects in the industry. Specifically in the panel industries, product quality depends on 40 multiple variables, especially wood variability. Among other factors, quality depends on the adhesion of 41 veneers or perpendicular tensile strength. The main objective of this study was to evaluate a Machine 42 Learning approach to predict the adhesion under industrial conditions in the gluing and pre-pressing stage. 43 The control variables that determine this adhesion are mainly: operational times, amount of adhesive, 44 environmental conditions, and veneer temperature. Using Knowledge Discovery in Databases data analytics 45 methodology, Artificial Neural Networks and Support Vector Machine were evaluated. Sigmoid activation 46 function was used with 3 hidden layers and 245 neurons. In addition to the Adam optimizer, Multi-Layer 47 Perceptron, Artificial Neural Networks delivered the best accuracy levels of over 66 %. Best result with Relu 48 and Sigmoid functions were obtained. Sigmoid showed accuracy over 66 %, precision fit good to find 49 positive results (70 %). Relu function obtained the best recall (over 74 %) showing a good capacity to identify 50 reality. Results show that it is not sufficient to generate a data set using the averages of each process variable, 51 since it is difficult to obtain better results with the algorithms evaluated. This work contributes to defining a 52 methodology to be used in plywood plants using industrial data to train and validate Machine Learning 53 models. 54