Issue 72

X. Cao et alii, Frattura ed Integrità Strutturale, 72 (2025) 162-178; DOI: 10.3221/IGF-ESIS.72.12

industrial sectors. All experimental materials were tested for fatigue in a stress-controlled mode at room temperature. Additional information regarding the other test data is available in the accompanying references. The distribution of data from five types of experimental data is represented in terms of loading cycle ratios in Fig. 3. Parameter settings In this work, all experiments were made on a personal computer with Windows 10 (64bit) operating system, hardware platform parameters of Intel(R) Core (TM) i5-7300HQ CPU, 2.50GHz main frequency and 16G RAM. The CTGAN model employs the Adam optimization algorithm for training the neural network. The discrete feature embedding dimension is set to 128, while the hidden layer dimensions for both the generator and discriminator are (256, 256). The learning rates for the generator and discriminator are both set to 2e-4 , with a learning rate decay of 1e-6 for both. The batch size during training is 600, and the number of epochs is 2000. 300 fatigue data for each material is generated. The relevant parameters of the machine learning models are presented in Tab. 2:

Machine learning models

Parameters setting

KELM

optimal regularization coefficient C: 20, kernel function parameter S: 1, kernel function: rbf penalty factor c: 4.0, radial basis function parameter g: 0.8, kernel function: rbf

SVM

RF BP

number of decision trees t: 100, minimum number of leaves l: 5

learning rate ： 0.01, error threshold: 1e-6

Table 2: Parameter setting of machine learning models.

In this work, a part of aluminum butt and corner joints, Al-2024-T42, Al-7075-T7451 and Ti-6Al-4V are selected as the test set. A part of the data of the test set is shown in Tab. 3. The rest of the data is the original training set. The generated fatigue dataset is then integrated with the original training set to produce the augmented training set.

n cycles 1 /

n cycles 2 /

f N cycles 1 /

f N cycles 2 /

σ MPa 1 /

σ MPa 2 /

Materials Butt joint Butt joint

104

74 89 73 83

109900 770100 309900 509200

549300 1540100 619800 1546100 150000 430000

1540100 880500 1546100 952300 430000 150000 225800 225800 143633 143633 61400

795800 581400 386120 708200 233400 89000 47400 27100 35700 22300 63800 198600

74 93 73

Corner joint Corner joint Al-2024-T42 Al-2024-T42 Al-7070-T7451 Al-7070-T7451 Al-7070-T7451

200 150 176 176 133 647 595 517

150 200 133

30000

258000

2000 2000 5000

27300 27300 61400 37200 64467

85 85

Ti-6Al-4V Ti-6Al-4V Ti-6Al-4V

517 517 595

18000 40000 30000

143633

64467

Table 3: A part of the test set.

Accuracy evaluation The evaluation indicators in this work are absolute percentage error Error and mean absolute percentage error ME to assess the performance of the model. Its calculation formula is:

− n N N 2 m f n 2

2

f

2

=

×

(18)

Error

100%

n N

2

f

2

171