Proceedings of SerbiaTrib '13

Fractal dimensions were determined using thebox-counting method which has been proven tohave higher calculation speed and more accuracyby Dougan and Shi.the structure of a neural network until it can modelthe problem in the most efficient way. Neuralnetworks are models of biological neural structures.The starting point for most neural networks is amodel neuron, as shown in Fig. 7. This neuronconsists of multiple inputs and a single output. Eachinput is modified by a weight, which multiplieswith the input value.Figure 6. Calculation of fractal dimensions with boxcountingmethodTo analyse the results we used one method ofintelligent system; the neural network. Artificialneural networks (ANN) are simulations ofcollections of model biological neurons. A neuronoperates by receiving signals from other neuronsthrough connections called synapses. Thecombination of these signals, in excess of a certainthreshold or activation level, will result in theneuron firing, i.e., sending a signal to anotherneuron to which it is connected. Some signals act asexcitations and others as inhibitions to a neuronfiring. What we call thinking is believed to be thecollective effect of the presence or absence offirings in the patterns of synaptic connectionsbetween neurons. In this context, neural networksare not simulations of real neurons, in that they donot model the biology, chemistry, or physics of areal neuron. However, they do model severalaspects of the information combination and patternrecognition behaviour of real neurons, in a simpleyet meaningful way. This neural modelling hasshown incredible capability for emulation, analysis,prediction and association. Neural networks can beused in a variety of powerful ways: to learn andreproduce rules or operations from given examples;to analyse and generalise sample facts and to makepredictions from these; or to memorisecharacteristics and features of given data and tomatch or make associations with new data. Neuralnetworks can be used to make strict yes-nodecisions or to produce more critical, finely valuedjudgments. Neural network technology is combinedwith genetic optimisation technology to facilitatethe development of optimal neural networks tosolve modelling problems. Genetic optimisationuses an evolution-like process to refine and enhance3. RESULTFigure 7. A neuron modelGraph [1-2] present relationship betweenroughness R a and hardness in specimens hardenedat different speeds at 1000 °C with both process.Roughness2502001501005004 mm/s3 mm/s5 mm/s2 mm/s55 56 57 58 59 60 61HardnessExperimental dataFitting curve with neural networkGraph 1. Relationship between roughness R a andhardness in specimens hardened at different speeds at1000 °CRoughness1400120010008006004002003 mm/s2 mm/s4 mm/s5 mm/s057,7 57,8 57,9 58 58,1 58,2 58,3HardnessExperimental dataFitting curve with neural networkGraph 2. Relationship between roughness R a andhardness in specimens hardened at different speeds at1000 °C with process of overlapping13 th International Conference on Tribology – Serbiatrib’13 357