Surface roughness prediction based upon experimental design and neural network models

The results presented in this paper are related to the prediction of the surface roughness generated by the grinding process. The main problems associated with the prediction capability of empirical models developed using the design of experiment (DOE) method are given. The first problem is a limited aptitude to calculate an accurate minimal output value as this optimal value was found to be absurdly negative in many cases. The second problem is that these models are not able to detect a particular behavior of the outputs for particular sets of the inputs. This constitutes a serious limitation of the application of this method to ground surface roughness prediction as the surface generation mechanisms differ at low and high work speed. In this study an approach suggesting the combination of DOE method and artificial neural network (ANN) is developed. Data of the DOE were used to train the ANNs and the inputs of the developed ANNs were selected among the factors and interaction between factors of the DOE depending on their significance at different confidence levels, expressed by /spl alpha/%. Results have shown particularly, the existence of a threshold value of /spl alpha/% to which correspond a critical set of inputs up to which no learning improvement could be realized. The built ANNs using these critical sets of inputs have shown /spl ap/ 0% deviation from the training data and low deviation from the testing data. A high prediction accuracy of these ANNs was tested through a good agreement between empirical models constructed using the developed approach and empirical models developed by previous investigations. The obtained results were valid for three kinds of steels having different properties and different hardness.