MATERIALE PLASTICE ♦ 51♦ No. 3 ♦ 2014 http://www.revmaterialeplastice.ro 343 Artificial Intelligence based System for the Real-time Control of Polymerization Processes TOM SAVU*, BOGDAN FELICIAN ABAZA, PAULINA SPANU University Politehnica of Bucharest, 313 Spl. Independentei, 060042, Bucharest, Romania The paper describes the components of an artificial intelligence based system intended to control the manufacturing processes of composite materials with polymeric matrix. Based on previous results in which neural networks were trained initially with simulated data and then continued to acquire experimental values over Internet for refining their knowledge, the system includes updated state-of-the-art data acquisition components, improved statistical data processing capabilities and the ability of real-time controlling the manufacturing equipment. Keywords: composite materials, polymeric matrix, artificial intelligence The composite materials with polymeric matrix are generally obtained by moulding and rarely by cutting or other processes. Either liquid moulding (specifically Resin Transfer Moulding), injection moulding or compression moulding, each technology is chosen according to the mechanical, geometrical or cost characteristics imposed for the final product. Obtaining the composite materials with polymeric matrix is a continuous process, the critical step being the manufacturing cycle, the composite’s polymerisation, when the temperature variation, expressed by amplitude and time, represents the most important parameter and influences the product quality. Designing the optimum manufacturing cycle for obtaining the necessary material quality during a minimised time is a goal which can not be easily reached due mainly to the intrinsec anisothropy of the composite materials and to the differences between various manufacturing methods or even producers. Actually the “trial-and-error” method is largely used, establishing the manufacturing cycle based on experimental research results which are used for evaluating several process parameters and defining various numerical models. Reducing the process duration is then obtained by applying optimisation techniques on the process models. The Loos-Springer model [1] is able to simulate the time variations for the composite’s temperature in different points and also for the internal pressure, extent of cure, resin viscosity, number of compacted layers, material thickness and residual stress in each layer. The Ciriscioli model [2], which in the case of thin layers is also based on empirical methods, is also providing parameters like temperature, extent of cure, viscosity, compactness and residual stress. Both mathematical models, even confirmed by experiments, have, like many others, a high complexity degree, and using them for simulations require an extended amount of time. Using the above mentioned models, simulations were performed by the authors [3] and a data base containing this type of results was created. Figure 1 is presenting the simulated temperature variation over time for an RTM * email: savu@ctanm.pub.ro; Tel.: 021 316 95 75 obtained vinilester with fiberglass matrix, while in figure 2 the extent of cure’s variation is presented for the same material. The results obtained during the mathematical models simulations were used for training a neural network with two neurons in the input layer and other two neurons in the output one, which basically is receiving data about material thickness and processing time, generating then data about temperature and extent of cure. After establishing its architecture, using the back- propagation algorithm in the Stuttgart Neural Network Simulator (SNNS), the network was trained using the simulated patterns during almost 5000 training cycles, connection weigjhts being thus adjusted and the sum of the squares of the errors minimised. After its validation and testing, the trained neural network was used for simulating RTM manufacturing cycles in the same conditions like those used by the numerical models. Data obtained both for temperature and extent of cure simulations are presented in figure 3 and 4 respectively. The correspondence between the results is obvious, but the time in which the simulation was performed by the neural network is much smaller than that needed for the numerical models, making this method suitable to be used in real-time systems. In previous works [4], the authors designed a data acquisition system which was monitoring the temperature during a manufacturing process and was passing the data, over the Internet, to a neural network for further training it Fig. 1. Simulated temperature variation over time for an RTM obtained vinilester with fiberglass matrix