Other
Study of the variability in processes of measurement of surface States through algorithms of neural networks
P. j. Núñez López, e. Morales Tower, Seville Hurtado (1), M. a. Sebastián Pérez (2).
University of Castilla - La Mancha. ETS for industrial engineers. Avda Camilo José Cela s/n, 13071-Ciudad Real. Tlfno: 926295218. E-mail: pjnunez@ind-cr.uclm.es(1) University of Málaga. Dept. of Civil Engineering, materials and workmanship. Plaza El Ejido s/n, 20013-Málaga.
(2) National University of distance education. Dept. engineering construction and manufacturing. C / City University s / n. 28080-Madrid. PO box 60149.
01/12/2002 1. Introduction
The objective of this paper is the study of the variability generated in the measurement of surface quality in machining processes through the use of artificial neural networks. This variability is changing and dynamic, reason by which his study by means of linear models to enable a proper analysis of the same arises. Artificial neural networks are perfectly adapted to the characteristics mentioned, being systems of information processing specially trained to capture the underlying non-linear structure in the data submitted and perfectly to adapt to changing environments.
2 Experimental
The development of the pilot phase have been assessed the surfaces of parts machined in similar conditions of court, through a roughness with probe of contact. Data processing has been used the multilayer Perceptron composed of three layers: input, output and hidden. Different architectures of networks through the programming of subroutines in Matlab (version 5.3) have been trained.
3. Results and discussion
Figure 1 shows that from the use of three neurons in the hidden layer, regardless of the measured sectors, as well as the number of measures carried out (1-5 steps), the error quadratic means is significantly reduced. To a lower number of nodes, peaks of error may be checked when the number of measured sections of the profile is less than four. If focuses the Studio to architectures of three neurons onwards, is checked to be measured at least four areas of the eight forming the profile, to achieve errors of the order of 10-3. The abrupt step that appears in Figure 1 is a clear demonstration of the said fact.
4 Conclusions
Algorithms of neural networks have made it possible to analyze the variability of the different strategies of measurement of the surface States. As a result, has managed to establish an optimal measurement method for each of the surfaces that make up the machined parts.
Figure 1. Results of the training.
5 References
[1] P.J. Núñez, "Analysis Experimental de la quality surface in processes of disposal of Material", Doctoral dissertation, UNED, Madrid, 1998.
[2] D.T. Pham, Pham P.T.N., "Artificial Intelligence in Engineering", Int. J. of Machine Tools & Manufacture, 39 (1999) p. 937.
[3] Mathworks, "Matlab User's Guide", The Mathworks, Inc., 1996