The prediction of surface integrity on EDMed surface with graphite electrode by using back propagation method

• Main Authors: FENG, SHI-HUANG, 馮世徨
• Other Authors: TAI,TZU-YAO
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/npks7c

碩士 === 南臺科技大學 === 機械工程系 === 105 === Artificial neural network concept has been proposed from the 19th centuries. In the development of Industry 4.0, which play an important role, this research used SKD11 (15mm × 15mm × 10mm) and graphite (10mm × 10mm × 70mm) to perform full factorial experiment. Material removal rate, electrode wear rate, surface roughness, surface topography, crack density and the thickness of the recast layer have been discussed by using artificial neural networks (back propagation method) for prediction. The results show that the material removal rate and the electrode wear rate are the most affected by the discharge current and the discharge-off time. The maximum influence factor of the surface morphology is the discharge current. The crack density increases when the pulse-on duration increases. The factors that have the greatest influence on the surface roughness are the discharge current and the pulse-on duration. The thickness of the recast layer is affected by the pulse-on duration, pulse-off duration and the discharge current. The back propagation neural network method was used to predict the material removal rate, electrode wear rate and surface roughness, among 64 experiments. When neurons was three hidden layers and the numbers of neurons was [23,23,23], the error can be less than 1%. In addition, 16 experiments are used to predict the thickness of the recast layer. The results show that when are hidden layer and the number of neurons [5], the prediction error can be less than 1%. For the crack density, 28 experiments are analyzed. The predicted error can be less than 4% when the two hidden layers and the number of neurons is [21,21]. Finally, the thickness of the recast layer is predicted by the surface roughness. When four hidden layers and the numbers of neurons is [14,14,14,14], the predicted error can be within 6%. This prediction can be applied, for the intelligent manufacturing processes.