PSI - Issue 75

Target parameter (specimen level)

Target parameter (specimen level)

0,35

(a)

(b)

100000000

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

100

7 Random Forrest ANN

Random Forrest ANN

0,3

10000000

0,25

0,2

1000000

2 [-]

0,15

RSME [-]

0,1

100000

0,05

0

10000

DS1

DS2

DS3

DS4

DS5

DS6

DS1

DS2

DS3

DS4

DS5

DS6

Fig. 4. (a) RSME and 2 score of the results from the investigations on specimen level (see Table 3) Data spaces

Data spaces

(b)

(a)

Specimen level

SN-curve level

Feature

Score 0.417 0.157 0.104 0.069 0.039 0.027 0.025 0.024 0.023 0.018 0.016

Feature

Score 0.512 0.228 0.042 0.039 0.033 0.027 0.021 0.019 0.018 0.014 0.011

6. Conclusions In this work different approaches compared regarding the prediction of fatigue life (number of cycles) of steels based on an extensive database by (Fliegener et. al., 2024) by ML. The fatigue life is evaluated based on estimated SN-curves and (SN-curve level) or directly predicted (specimen level), see Figure 1. The data spaces for the training of the ML approach were used and different ML approaches were applied. Following conclusions are made: • Higher 2 scores are achieved for the prediction of the SN-curve parameter and compared to the prediction of the fatigue life by the applied ML approaches. •

Stress amplitude Surface hardness

Stress concentrationfactor

Surface hardness

Stress concentrationfactor

Cr

Ni

Loadtype

Load type

P

Temperature

C

C

Roughness Rz

Roughness Rz

Ni

Cr

Mn

P

Si

Stress ratio R

Stress ratio R

Fig. 5. Feature importance score

However, the direct prediction of based on the given input parameter, see Figure 1 (b), seem more meaningful than predicting SN-curve parameter and and then the fatigue life from based of these parameters because in that case a much higher prediction accuracy is reached. • The feature importance score of the random forest approach using all data shows that SCF is the most important factor regarding prediction accuracy on SN-curve level and on specimen level it is the stress amplitude. • In the given database are data sets (SN-curves) that have a large scatter (different values of for the specimen on the same stress amplitude). At the same time no parameter is documented how this can be considered for further analysis (for example individual roughness, initial crack size, SCF … of the single specimen). This limits the applicability of the data and results of a low accuracy in the training of ML algorithm. • Averaging the fatigue life of single specimen at the same stress amplitude can decrease the RSME by ML based fatigue life prediction. Increasing the consistency of the data by decreasing the variation of materials, specimens, … had only a minor effect in this work. In general, this work shows the limitation of ML-based fatigue life prediction if the training data space is large and not always consistent. A lack of standardization in fatigue testing and often incomplete documentation of fatigue test results and the material conditions limit the applicability of older data for current used ML approaches. Indeed, this crucial factor needs to be improvement until open-source databases can be used for fatigue life prediction by ML.