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

377

6

Although increasing the parameters, to some extent, can improve the accuracy of the ANN prediction, it would complicate the ANN structure and reduce the efficiency of the calculation. Thus, the number of parameters should be limited to ensure the efficiency of the calculation and avoid overlearning. Despite this, the number of parameters required for reliability analysis can't be arbitrarily reduced because the reduction could have an impact on the accuracy of the ANN prediction. To avoid an arbitrary reduction of the parameters and to obtain a reliable simplification of the ANN modeling, a sensitivity analysis is applied to determine the primary parameters that will be retained in the ANN modeling. Since this is a low probability event, pipe failure due to corrosion defects usually doesn't contain sufficient in service inspection data for reliability prediction. Thus, in order to obtain sufficient data for reliability analysis, appropriate statistical methods are needed. MCS is a computerized mathematical algorithm that uses the randomness of primary parameters to obtain numerical results through the process of repeated sampling.

Fig. 1. A flow chart of the proposed methodology.

A lot of previous research has made it possible to verify the reliability and feasibility of MCS in predicting failures. The ANN procedure calculating the reliability of the IG-SCC pipe is illustrated in Figure 1. Based on the probability distributions of the parameters, the data set necessary for estimating the limit state can be generated by random sampling. Finally, the entire database generated through the MCS will be separated into three groups. The data groups are for ANN training, testing, and validation, respectively. The percentage of the entire database for each group should be determined based on specific requirements. The ANN inputs are the sampled parameters, while the outputs of the ANN are the reliability estimated by the MCS. In looking for the best ANN model, one has to determine the appropriate number of hidden layers and the number of neurons in each one. This is done though training and testing of different network structures and the appropriate