The method based on the modified universal slopes was selected as the best representative of the category of simple approximations, based on the strain amplitude variation diagram, for the material under investigation. Lee & Song (2006) also found that this method provides the best estimation for low-alloy steels and stated that it should be used as the first choice method in the fatigue analysis of this group of steels.
In the category of CDM-based models, only the method proposed by Bhattacharya &
Ellingwood (1999) was chosen because of its simplicity. Other CDM-based models were not investigated as either their validity for the LCF condition is unconfirmed or they require detailed experimental data, making them unsuitable for use by practitioners. The predicted curve using the CDM-based model tends toward the elastic line of the experimental curve as the number of cycles increases. The same trend was also seen in the original study for two other types of steels (Bhattacharya & Ellingwood, 1998). The theory used for the development of this model as an approach for prediction of the early stages of the crack initiation phase (smaller cracks in size compared to those in the conventional strain-life approach) could explain the higher conservatism of the predicted curve compared to the experimental curve and the other estimated curves.
3.3 Discussion on the accuracy of the investigated methods 31 The curve estimated by the ANN-based model also gives acceptable results at a medium to high number of cycles to failure, with its maximum accuracy for reversals higher than 105. Although no mechanical theory supports the application of the ANN method in estimation of the strain-life curve, and subsequently fatigue properties, the results show the potential of this technique to act as an approximation method. It should be mentioned that the current model is implemented mainly on the concept of correlation between monotonic properties and fatigue life proposed by Basan et al. (2010).
Therefore, besides the inherent limitations of the ANN technique for this specific application, uncertainties in the concept itself could act as a source of error in the estimations.
Although non-linear curve fitting (in the form of the Coffin–Manson relationship) could give the fatigue properties, it should be highlighted that these values are not unique, as different networks – even with high levels of correlation in regressions – could yield slightly different life estimations. Therefore, in order to mitigate this variation and obtain more stable estimations for fatigue coefficients, the constant values for fatigue exponents were considered as 𝑏 = -0.08 and 𝑐 = -0.63, which were selected as the mean values for all the low-alloy steels used as the data set. These values differ from the constants that are considered by the modified universal slopes and hardness-based methods for steels. Based on the data set comprising of 60 low-alloy steels, the constant values considered in the current study better define the fatigue exponent values of low-alloy steels, which is also confirmed by the experimentally gained values for S960 MC.
Curve fitting by non-linear regression through estimated lives by using the ANN-based model, along with the consideration of constant values for the fatigue exponents, yielded estimations of σ´f = 1822 MPa and ε´f = 0.8 for S960 MC.
Of all three categories of methods studied here, it can be concluded that the method based on modified universal slopes, requiring only simple monotonic properties, gave the best estimation of the strain-life curve when compared with the experimental curve of the material under investigation.
33
4 Extending the ANN-based models to the estimation of the stress concentration factor of welds
Welding is the most common joining process used in the fabrication of components and structures. Although the high strength steels, benefiting from their strength–weight ratio, are used in very demanding engineering applications, the welded components of these materials could suffer from losing their strength due to the flaws, high residual stresses and other imperfections introduced to the component by improper welding (microstructurally and geometrically). Because of being exposed to in-service cyclic loads, these welded parts are susceptible to fatigue failure and their analysis is an inevitable part of the design process. The term stress concentration and the factor defining it (the SCF), form one of the main parameters which is estimated and used in the fatigue analysis of notched components, such as welded joints. This factor can be obtained experimentally, analytically and numerically; among them the numerical methods, such as the FE method, can be considered the most accurate.
As confirmed by the investigations in the previous chapter, the ANN-based models can be used successfully in order to define the relations (linear to highly non-linear) between related parameters. In this section, it is tried to utilize this technique in the estimation of SCF in welds. Knowing the fact that the implementation of a successful network strongly depends on its input data, special attention was paid to defining the effective profiles (by using an appropriately designed experimental method) and their modelling using the FE method.
4.1
The investigated configurationsIn a successful attempt, the application of ANNs was extended to the calculation of the SCFs of different weld types, such as T-welded and butt-welded joints, schematically shown in Figures 4.1 and 4.2 respectively.
Figure 4.1: The configuration of a T-welded joint and effective parameters for the SCF
Extending the ANN-based models to the estimation of the stress concentration factor of welds 34
Figure 4.2: The configuration of a (a) single-V butt weld and (b) double-V butt weld First of all, a literature review was performed in order to obtain the common empirical equations used for SCF calculation for these welded joints in order to define and select the best and most accurate proposal for comparison. This led to the selection of both the empirical equations proposed by Brennan and Helier (Brennan et al., 2000; Hellier et al., 2014) for T-welded joints and that proposed by Kiyak et al. (2016) for butt-welded joints. Only these equations are listed below and different proposed equations are left out of this section. The reader, however, is referred to P-III and P-IV for an in-depth review of these equations.
The equation used as the reference for comparison is the one proposed by Brennan et al.
(2000) for SCF calculation for T-welded joints in as-welded condition both under an axial load (theta in radian):
𝑆𝐶𝐹𝑎 = 1.1 + 0.067𝜃 − 0.25(𝑟 𝑡⁄ ) − 0.04(𝑤 𝑡⁄ ) + 0.003𝜃2− 12(𝑟 𝑡⁄ )2− 0.014(𝑤 𝑡⁄ )2+ 0.0164𝜃3− 0.0005(𝑤 𝑡⁄ )3+ 0.00004(𝑤 𝑡⁄ )4− 0.3𝜃(𝑟 𝑡⁄ ) − 0.023𝜃(𝑤 𝑡⁄ ) + 0.91(𝑟 𝑡⁄ )(𝑤 𝑡⁄ ) − 8.3𝜃2(𝑟 𝑡⁄ ) + 0.225𝜃2(𝑤 𝑡⁄ ) + 100.5𝜃(𝑟 𝑡⁄ )2− 0.0792𝜃(𝑤 𝑡⁄ )2− 37.5(𝑟 𝑡⁄ )2(𝑤 𝑡⁄ ) + 0.908(𝑟 𝑡⁄ )(𝑤 𝑡⁄ )2+
0.27𝜃0.19(𝑟 𝑡⁄ )−0.47(𝑤 𝑡⁄ )0.25, (4.1) and under a bending load:
𝑆𝐶𝐹𝑏 = 1.14 + 0.13𝜃 − 0.67(𝑟 𝑡⁄ ) − 0.083(𝑤 𝑡⁄ ) + 0.08𝜃2+ 28(𝑟 𝑡⁄ )2− 0.02(𝑤 𝑡⁄ )2+ 0.01𝜃3− 0.0005(𝑤 𝑡⁄ )3+ 0.00002(𝑤 𝑡⁄ )4− 4.3𝜃(𝑟 𝑡⁄ ) −
0.09𝜃(𝑤 𝑡⁄ ) + 1.03(𝑟 𝑡⁄ )(𝑤 𝑡⁄ ) − 13.7𝜃2(𝑟 𝑡⁄ ) + 0.443𝜃2(𝑤 𝑡⁄ ) + 150𝜃(𝑟 𝑡⁄ )2− 0.13𝜃(𝑤 𝑡⁄ )2− 62(𝑟 𝑡⁄ )2(𝑤 𝑡⁄ ) + 1.53(𝑟 𝑡⁄ )(𝑤 𝑡⁄ )2+ 0.005𝜃3(𝑤 𝑡⁄ ) − 30𝜃(𝑟 𝑡⁄ )3+ 3.57𝜃(𝑟 𝑡⁄ )(𝑤 𝑡⁄ ) + 5𝜃(𝑟 𝑡⁄ )2(𝑤 𝑡⁄ ) + 0.35𝜃0.26(𝑟 𝑡⁄ )−0.468(𝑤 𝑡⁄ )0.3. (4.2) The investigation continued by scrutinizing the ability of ANN-based models to be used in the SCF calculation of butt-welded joints. This joint type was studied under axial and bending loads in both single-V and double-V forms. Based on the literature review, the
4.2 Design of the experiments 35