| Resumo : |
Surface roughness is one crucial quality element in machining production, and its prediction has been the case of study for many years. Generally, three approaches are performed to model it: empirical methods, theoretical/simulation methods, and soft computing methods. In this work, artificial neural network models (ANN) were trained with two training datasets generated by machining experiments-turning of an AISI H13 steel with cutting fluid. The first experiment with theoretically new-tool conditions produced 324 samples, and the second with cutting tool flank wear varying in three levels produced 288 samples. The ANN models were formed with three hidden layers (28-56-14) and Ra as the predicted output. One model used the depth of cut, feed rate, and cutting speed as inputs, and another model used those features and the machining force components-cutting force (Fc), feed force (Ff), passive force (Fp), and the resultant force. A strategy was proposed to augment the available data six times by increasing the measurements without increasing the experiments. The models trained with a smaller dataset were prone to overfitting and achieved less performance than those with larger datasets. The model trained with the machined forces performed 75% better than the model without them, producing a prediction error (MAPE) of less than 5%. Moreover, the models trained with the first dataset (without tool wear) could not generalize to the second dataset (with tool wear), indicating that tool wear is an essential factor when modeling surface roughness in turning processes using ANNs. Also, the ANN models performed better than a classical theoretical surface roughness equation. |