PSI - Issue 75

Jan Schubnell et al. / Procedia Structural Integrity 75 (2025) 94–101 Schubnell/ Structural Integrity Procedia (2025)

99

6

5. Results The results of the investigations from Table 3 are illustrated in Figure 5. As shown similar accuracy is reached than former investigations on SN-curve level (Fliegener et. al., 2024) based on the new evaluation according to DIN 50100:2022. On specimen level lower scores could be achieved ( 2 <0.3). The authors assume this is based on inconsistency of the data (for example by incomplete documentation followed by estimation of single parameter by (Fliegener et. al., 2024)). In comparison of the two given approaches in Figure 1, a direct determination seems much more meaningful because the prediction of the fatigue life of single specimen based in predicted SN-curve parameter and , see approach in Figure 1 (a), fail in this work, see Figure 3 (c). While, the prediction of the fatigue life by ML based on the approach displayed in Figure 1 (b) leads to clear but not very accurate results. However, even small and relative consistent data (Data Space 6), that is limited to a single material, similar specimen geometry and supported by a low number of labs no significant increase of 2 is determined, see Figure 2 (b), but an reduction of the RSME from data space 3 to data space 6 by a factor of 20, shown in Figure 4. In comparison between the Random Forrest approach with a shallow ANN, the ANN leads to sightly but not significant better results. Comparisons in this study show also that Random Forrest may lead to smaller RSME if the data size is low (300 data points and less). Different techniques were investigated to improve the results by increase the consistency of the training data space (from Data Space 1 to Data Space 6), see Table 2. For example, exclude runouts, limit the number of cycles or limit the training data space to single materials. The only significant effect was determined by average the fatigue life of the single specimen for the same stress amplitude (data space 3 to data space 4), according to equation 2. This leads to an significant improvement, see Figure 4. The feature importance score represents the significance of the feature regarding the prediction accuracy in random forest and is displayed in Figure 5. As shown most important feature on SN-curve level is the SCF, see Figure 5 (a), and the stress amplitude on specimen level, see Figure 5 (b).

(a)

(b)

Data Space 1 1151 SN-curves Random Forrest Target: [MPa]

Data Space 6 1095 specimen ANN Target: [-]

2 = 0.29 RSME = 202404

2 = 0.71 RSME = 89 MPa

(d)

(c)

Data Space 4 3092 specimen ANN – Specimen level Target: [-]

Data Space 4 3092 specimen ANN – SN-curve level Target: [-]

2 = 0.23 RSME = 393551

Prediction fail ( < 0)

Fig. 3 . Illustration of results depending on target parameter, ML-approach and data space