Neural modelling of friction material cold performance

Abstract

The complex and highly non-linear phenomena involved during braking are primarily caused by friction materials’ characteristics. The final friction materials' characteristics are determined by their compositions, manufacturing, and the brake's operating conditions. Analytical models of friction materials' behaviour are difficult, even impossible, to obtain for the case of different brakes' operating conditions. That is why, in this paper, all relevant influences on the friction materials' cold performance have been integrated by means of artificial neural networks. The influences of 26 input parameters, defined by the friction materials' composition (18 ingredients), manufacturing (five parameters), and brake's operating conditions (three parameters), have been modelled versus changes of the brake factor C. Based on training and testing of 18 different architectures of neural networks with five learning algorithms, a total of 90 neural models have been investigated. The neural model (BR26841) trained by the two-layered neural network, with a Bayesian regulation algorithm, was found to reach the best prediction results. This neural model was able to generalize the friction materials' cold performance, for temperatures in the contact of the friction pair T>100°C, in the range of application pressure changes between 20 and 100bar, and for initial speed changes between 20 and 100km/h.