Pressure drop through textile fabrics - experimental data modelling using classical models and neural networks

This work studies pressure drops through several textile fabrics. A preliminary study of cloth characteristics, including scanning electron micrographs, shows their specificities towards particular media. For 20 different cloths, in terms of weave and raw material (rayon or activated carbon fibers), an experimental study is carried out using a pilot-unit, in order to measure air and water pressure drops through one layer of each cloth. Fluid Reynolds numbers range from 0 up to 2500 for both fluids. This experimental study shows the influence of specific parameters of cloths (like weave) on their dynamic behavior. Furthermore, the swelling phenomenon of fibers in water is considered. Goodings' model is set up for woven structures and it enables the fabric opening diameter to be calculated around 10 mu m. Experimental data are then modelled, firstly using classical models set up for particular porous media (Ergun, Carman's dimensionless model, Comiti-Renaud), and then using a statistical tool, neural networks. These models are tested using three different definitions for the specific surface area, on the fabric, yarn, and opening scale, respectively. Whichever the definition used, they are not suitable to describe the flow through woven structures. However, they enable the swelling phenomenon of fibers in water to be confirmed, and the flow into the fabric yarn to be located. The experimental study, coupled with these modelling results, leads to the choice of input neurons in the neural network (fluid properties - mu, rho, Re - and fabric characteristics - thickness, density, number of openings N-o, S-o, and raw material) in order to predict pressure drops as the output neuron. The statistical results obtained with this architecture are satisfactory and a variable analysis carried out with connection weight values enables the influence of specific parameters of cloths (like N-o) on pressure drops to be quantified. (C) 2000 Elsevier Science Ltd. All rights reserved.