Monitoring the process of curing of epoxy/graphite fiber composites with a recurrent neural network as a soft sensor

Controlling the curing of fiber-reinforced composites involves an on-line evaluation of their properties such as viscosity, resin content, and degree of cure (DOC). Infrared spectroscopic and dielectric sensors have commonly been considered for monitoring these properties. Nevertheless, they are expensive, and yet do not yield a precise cure history during the entire process. Artificial neural networks have successfully been adopted for the dynamic modeling of nonlinear systems. Inasmuch as the actual DOC of a composite cannot readily be measured in situ during the cure, long-term prediction of the DOC is critical. In the present study, a unique integrated sensor has been constructed that comprises a dual heat-flux sensor serving as a hard sensor for determining the Damkohler number (Da) and a recurrent neural network (RNN) serving as a soft sensor for evaluating and predicting the DOC on the basis of the Da obtained. At the outset, the prototype soft sensor, i.e., RNN, was configured through a series of repeated and rapid simulations of an analogous model system with known performance equations for learning and testing. Subsequently, this prototype RNN was tuned and validated through a minimum number of laborious experiments,so that the resultant soft sensor is capable of effectively monitoring on-line the DOC of the prepreg of a commercial epoxy/graphite fiber composite in a bag-molding process. (C) 1998 Elsevier Science Ltd. All rights reserved.