NEURAL NETWORK FOR THE MFL INSPECTION OF STEEL TUBES
E. Altschuler**, H. Gavarini*, R. P. J. Perazzo, A. Pignotti**, S. L. Reich** Centro de Estudios Avanzados, University of Buenos Aires *Depto. de Fisica, Lab. TANDAR, CNEA, FUDETEC **Center for Industrial Research, L. Alem 1067, (1001) Buenos Aires, Argentina
The ability of neural networks to improve the assessment of the severity of cracks in steel tubes is analyzed. A Monte Carlo calculation is performed using a numerical model of crack detection by magnetic flux leakage [1, 2]. The analysis is carried out with a population of signals in which the depth of the crack, its width, shape, location, and other parameters of the detection process, are allowed to change randomly. The simulated signal is sampled under different realistic measurement conditions, and is input to the network, after a reduction process in which the main features of the signal are extracted. The output of the network indicates whether the crack is an external or internal one, and provides a numerical assessment of the crack depth that is compared with a threshold. The probability of detection of flaws that exceed a critical size and the occurrence of false alarms are studied as functions of this threshold, and are compared to the corresponding values obtained via the normal procedure, which only takes into account the peak value of the signal. The analysis based on the neural network is shown to be quite superior, decreasing significantly the occurrence of false alarms for the same probability of detection. REFERENCES
[1] E. Altschuler and A. Pignotti, "Non linear model of flaw detection in steel pipes by magnetic flux leakage", NDT E International, Vol. 28, NO. 1, pp. 35-40, 1995.
[2] E. Altschuler, A. Pignoti and J. Paiuk, "Monte Carlo simulation of false alarms and detection reliability in MFL inspection of steel tubes", to be published in Materials Evaluation.
Publication Source: Trends in NDE Science & Technology; Proceedings of the 14th World Conference on Non-Destructive Testing, New Delhi, 8-13 December 1996.Vol. 3, pages 1841 - 1844 Publisher:Ashgate Publishing Company
NDTnet
