The first step is to choose the parameters which influencing the delectability of the defects. Each parameter is treated as a random variable characterized by a probability distribution. These parameters are used as entry variables for the model. We make a random draw of each parameter. After computation we have a signal which corresponds to the initial configuration. We start again the random draw and the computation. When we have made these calculations sufficiently often to have a statistically significant set of results, we can apply the detection procedure to these results. We obtain a probability of detection of the defect. In fact we can use the same principle for the characterization of the defect and build also the probability of good or false sizing.

To illustrate this approach we have took an example on artificial underclad defects in pressure vessel of nuclear power plants. After the choice of the significant parameters (defect orientation, defect size, error on positioning the probe, .... ) and the determination of the statistical law of variation of each parameter for these situations of testing, we have used a model to calculate the response of the defects for each set of parameters and to build a probability of detection. The model used is Mephisto developed by the French Atomic Authority (CEA).

To verify the accuracy of the method and to test the choice of the different laws for the variation of the parameters, we made an experimental validation using a test block which contains various defects. This test block is a cladding block with EDM notches of different lengths and orientations. In the experiment we have also induced some light variations for the probe incidence or for the distance between transducer and the test block and the step between two acquisitions (we use a focussed transducer in the same manner as for the inspection of the nuclear vessels in France).

In this paper we would like to present this approach and the numerical results as well as the comparisons between experimental and numerical results.