PSI - Issue 52

Wu Zonghui et al. / Procedia Structural Integrity 52 (2024) 203–213 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

206

4

Fig. 2. Imitation effect of MSE.

Fig. 3. Imitation effect of MAPE.

2.2. Sample generation The cost, time for finite element analysis, and resources for fieldwork are pretty high with the increasing complexity of the model. So there is a high demand to reduce the samples for ANN's training procedure. The LHS and the UED are used respectively for large and small sample generation. These two methods are both effective techniques to reduce the sample size, and points sampled by the UED are more well-distributed in the whole space while the LHS is stochastic and hard to get fail points with a small sample size. Hence the UED was taken as the method for small size, and the LHS technique was used to fill the vacancy from the small size to the large one. The sampling interval obeys the 6- σ criterion: ( μ - 6σ , μ+6σ ) is taken as the interval of normal distributed variables, which contains 99.99966% proportion of its whole distribution. The interval of other distributed variables also contains a 99.99966% proportion of its whole distribution. For example, (0, α ) is the sampling interval of a lognormal variable, and it needs to fit this inequation: ( ) 0 d 99.99966% f x x    . The diagrammatic drawing is shown in Fig.4.

(a) Normal distribution.

(b) Lognormal distribution.

Fig. 4. Sampling interval.