| Brake friction material of the friction performance prediction can to the brake friction material research and development of new products play an important role. Artificial neural network(ANN) is a kind of network which simulation from brain, To be external environment for learning process by set up a method of pattern recognition. Artificial intelligence has self-learning, self-organizing, self-adaptive. It is suit for nonlinear function approximation, and has excellent capacity of tolerant imperfection, especially well done in nonlinear system. Which also means artificial neural networks using in complex non-linear system, will be approach to real or raw date, no matter what kind of functions setting in there. However, the more complex system in the study which we observe by artificial neural networks, the more well performance that network will produce.In this paper we gather the large group raw data form the method’GB5763-2008 in class 4 disc brake’. Then use ANOVA to classification data groups. Grouping data study by the artificial neural network, and to compare predict performance of 3 artificial neural networks, which is feedback Elman neural network, BP feedforward network, radial network RBF. In the study mainly compare friction material heating and cooling coefficient form 3 networks predicted, BP and Elman network training function trainlm and traingdx and trainbr (Bayesian regularization training)will be cautious compare by prediction error, finally we chose the Bayesian regularization by using in Elman network. Single-layer network experiment in the experiment to choose Elman network to predict the experimental data of single layer 6[10]11, This structure that single-outlayer network usually show good prediction precision. Using 10 neurons in hidden layer and transfer function is tansig, single-outlayer output function is logsig, the Elman network show accurately predict, when the lower brake material are used. Elman network transform use multi-layer network structure, by using double and triple layer of network show different to original 10 neurons single hidden layer network, the middle layer is still the tansig multi-layer transfer, the network structure Elman6[5,5]21、Elman 6[6,5]21、Elman 6[6,4]21、Elman 6[4,6]21、 Elman 6[7,3]21、Elman 6[4,3,3]31、Elman 6[4,3,2]31、Elman 6[3,3,2]31、Elman 6[5,4,3]31、such as increased from 10 neurons to 12 or more neurons to end up with the total average error smallest Elman 6 [5,3] 31, hidden layer structure 5,3-6 input, three layer structure, single output of he network. So I conclusions something as follows:(1) More hidden layer neural network can be optimized single hidden layer neural network in the friction material performance prediction accuracy.(2) The Elman network forecasting precision of the experimental data of this study is highest, single-layer network chose 6 [10] 1, use the s-shaped tangent transfer function and s-shaped logarithmic output function, more accurately predict the compounds containing lower abrasive friction material of warming and cooling friction coefficient of friction coefficient.(3) Multi-layer network higher selectivity, and can predict the experimental results on a small scale better.(4) Considering the Elman network is often used to predict the dynamic, nonlinear data, and feed forward BP, RBF network is often used to predict the static data and radial function fitting work, the results of this study support to some extent in the friction material according to the GB5763-2008 class 4 disc brake for the friction lining performance testing methods of data presentation dynamic and nonlinear relationship.(5) To optimize Elman network transform multi-layer network structure, respectively, to differentiate the original 10 neurons use double deck and three layer of network are single hidden layer network, from a total of 10 neurons increase to 12 or more neurons to forecast, combined average error using Elman6 [5,3] 31 structure minimum error. |
PDF Full Text Request