PSI - Issue 41

Mohamed Amine Belyamna et al. / Procedia Structural Integrity 41 (2022) 372–383 Mohamed Amine Belyamna et al. /Structural Integrity Procedia 00 (2022) 000–000

378

7

one should ultimately be determined by evaluating tolerance between predicted and actual data. In this work, the input variables are IG SCC D  Values for various degrees of sensitization, different levels of applied stress (both the applied service-induced pressure and thermal and residual stresses), different steady state temperature, different O 2 content and the periods of plant operation. The output variable is the failure probability. MATLAB based the Levenberg-Marquardt algorithm, which is known to be highly efficient in solving problems of nonlinear optimization, was used to train the neural network. The training and testing exercise as indicated in the previous paragraph resulted in a network of one hidden layer with 5 neurons shown in Figure 2.

Fig. 2. The architecture of the optimal ANN model.

4. Results and Discussion In this section the calculations were focused on leak probabilities because M-PRAISE calculates probabilities of leaks more accurately and economically than probabilities of double-ended pipe breaks. The parametric calculations concentrated on welds in 304 stainless steel piping and on the cumulative leak probabilities over a 40-year time frame of plant operation. Table 1 summarizes the matrix of calculations along with the input parameters for the calculations. These calculations assumed realistic ranges for the various input variables that govern the initiation and growth of IG-SCC cracks. The following variables were addressed: O 2 content, temperature, coolant conductivity, applied stress, and frequency of heat-up and cooldown. In this paper, the ANN established, trained, and validated presented in Figure 2 is a powerful tool to estimate the reliability of the pipe effectively and quickly. This section presents a collection of plots that show trends for pipe initiation, leak and break probabilities and the result of the second ANN used to generalize the PFM model based IG SCC D  . 4.1. Calibration of Model Figure 3 compares the field cumulative leak probabilities for two size pipes given by (Khaleel et al., 2009) and results obtained by (Guedri, 2013b). In this case, the adjusted residual stress level used to limit the disagreement between predicted and observed leak probabilities were set at 75 % of their original values. The resulting predictions had a much more rational basis and were in very good agreement with operational data for periods beyond six years. The less satisfactory level of agreement for periods less than six years can be attributed in a large measure to lack of observed failure events for the early periods of plant operation.