Surface roughness prediction through internal kernel information and external accelerometers using artificial neural networks

In this paper, the average surface roughness parameter (Ra) is predicted using artificial neural network (ANN) models and internal kernel information and external piezoelectric accelerometer data. Experiments were conducted to Obtain data to develop ANN models to predict surface roughness. A total of 72 samples were used to develop two networks, one based on accelerometer inputs and the other on kernel inputs. The Matlab ANN Toolbox was used for the modeling. The two networks had similar characteristics. Feed-forward back-propagation, 'newff', was the network structure selected, with a Levenberg-Marquardt backpropagation training function, 'trainlm', and a backpropagation weight and bias learning function, 'learngdm'. Samples obtained at the experimental stage were randomly divided into three groups to train (70% of the samples), validate (15% of the samples) and test (15% of the samples) the neural networks with a 'dividerand data division function. The input processing functions used were 'fixunknowns', 'removeconstantrows' and 'mapminmax'. The transfer function was 'tansig' for hidden layers and 'purelin' for the output layer. The output processing functions used were 'removeconstantrows' and 'mapminmax'. The inputs consisted of the process parameters of radial depth of cut (Ae), the axial depth of cut (Ap), the spindle speed (N), the feed rate (f), the feed per tooth (fz), the cutting speed (Vc), the tooth passing frequency (ft), the cutting section (Cs), the material removal rate (MRR) and the cutting tool characteristics of the cutter radius (R), the number of teeth (Z) and the tool shape. The main difference between the two neural networks consisted of data origin: one considered data obtained with accelerometers and the other data collected in the NC kernel. Results showing high correlation factors between outputs and targets confirm that data provided by both internal and external sources can be useful for Ra prediction. However, NC kernel data provide several advantages.