Neural Network Modeling for Fiberboard Properties Prediction

摘要

Medium Density Fiberboard (MDF) is an alternative to solid wood used in furniture industry. Although fiberboard is widely preferred for its reasonable price, however, since it is not solid, establishing the level of its strength is the main issue. Therefore, prior to releasing processed fiberboards for manufacturing, samples need to be tested. The test procedures for mechanical and physical properties should conform to BS EN standard. These tests are costly for they involve high amount of resources, especially to research institutions. The aim of this research is to utilize the properties of MDF so that procedures on lengthy tests, namely, 24-hour-thickness swelling, 24-hour-water absorption and 48-hour-moisture content can be reduced. A prediction model was used to make predictions on omitted tests. Utilized properties were fed as input to the multilayer perceptron Neural Network. Several criteria of Hidden Layer (HL) and Output Layer (OL) were involved in search of best prediction. Results have shown low Sum of Squared Error (SSE) for all criteria. Scaled Conjugate Gradient optimizer presents lowest SSE when used with hyperbolic tangent activation function for HL and sigmoid activation function for OL. On the other hand, Gradient Descent optimizer is best with sigmoid activation function in both HL and OL. Prediction models have contributed to the increase in MDF testing efficiency based on BS EN standard.