1. Introduction
The estimation of the degree of accumulated fatigue damage (damage degree) is necessary for the most common deterministic approach for the fatigue life analysis of structural materials under variable amplitude loading [
1,
2]. In general, the existing damage accumulation rules are based on models that can be classified into two categories—micro-scale and phenomenological [
1,
2]. The modeling of microscale mechanisms is still a challenging issue due to their complexity arising from the diversity of fatigue crack mechanisms and the randomness of material microstructures. The data needed for the calculation process may not be fully available in standard application situations, which, together with their complexity, makes these models difficult to apply in practical engineering. As an alternative, from the phenomenological perspective, fatigue damage is treated as a time- and loading-dependent process. It is assumed that damage gradually accumulates along with the loading cycles and that a developing crack will result in failure when the accumulated damage exceeds its critical value. Estimates of the accumulated damage are obtained mostly on the basis of the widely available stress–life (s-n) curves. The fact that these curves have been identified by the least squares method [
3] leads to the inappropriate assumptions that 50% of the experimental lives are on the left side of the reference s-n curve, and 50% of the experimental lives are on the right side and that the normal distribution is appropriate for modeling fatigue damage. This issue could be overcome by, for example, implying a probabilistic stress–life curve, as was done in Reference [
4], or a non-linear s-n curve, as was done in Reference [
5], with the median interpolated according to the Weibull distribution identified by the maximum likelihood estimation. However, the latter can be viewed as being too complex for engineering approaches, mostly due to the problematic identification of the likelihood function. Nevertheless, in order to maintain the simplicity of the widely used ASTM standard [
3], the modified approach should correspond to the existing s-n curves.
While analyzing the failure probability–stress–number of load cycles (P-s-n) non-normal probability distributions, it can be concluded that the number of cycles corresponding to the mean damage degree does not correspond to the mean fatigue life. At the same time, significant scatter of the computed damage degree versus the mean life can be observed. Although considerable studies have been performed on statistical models of fatigue damage accumulation, they often can be viewed as having a high level of complexity (e.g., References [
6,
7,
8]).
The aim of the present study was to propose a simple way to modify the stress-related damage accumulation rules in order to better reflect the nature of the variance in experimental fatigue life and, as a result, to decrease the scatter between the predicted and experimental life. The approach is based on the probability distribution for a one-stage loading.
2. Relation between Fatigue Damage Accumulation and Fatigue Life
The fatigue life depends on the amount of damage the material can accumulate. Since this measurement cannot be physically obtained, an appropriate deterministic measure is usually introduced. Logically, the damage measure should take into account the applied load (or its result) and the loading duration. The most common measure is the damage degree,
d. In the historically first, phenomenological Miner’s rule, the degree of damage accumulated during one loading stage, can be obtained according to the following equation:
where:
n—the number of cycles in the loading stage;
nf—the s-n curve number of cycles to failure.
In case of a multi-stage loading, it is rewritten as follows:
where:
i—the index of the loading stage;
j—the number of loading stages.
Here, the information on the applied load is hidden in the
nf value, whereas the duration of the applied load is given in the
n value. The value of the number of cycles to failure
nf is derived according to the s-n curve, relating the applied constant amplitude load to the fatigue life (
Figure 1). Most frequently, in such formulations, it is assumed that the fatigue failure occurs when the damage degree reaches the critical value of
d =
dc = 1, which applies also to Miner’s rule.
As can be seen, Miner’s rule is based on a linear model of damage accumulation. Although Miner’s rule is widely used in the case of metallic materials because it is easy to implement and computationally inexpensive, an issue was raised concerning the scatter of the critical damage degree [
1]. Linear damage accumulation rules, under certain circumstances (e.g., sequential loading effects, notches, non-standard environments), can produce significant and unusual scatter, corresponding to the scatter band equal to three or even more [
9,
10]. Non-linear damage accumulation rules offer more accurate or more conservative predictions than the linear rules [
9,
10]. They also seem to be physically justified since the fatigue processes involve crack nucleation, crack propagation, and final failure.
In 1956, Corten and Dolan formulated a power–law relationship for damage accumulation, which allowed for accounting for load effects [
11]. Corten–Dolan’s model is also widely used and, being stress level-dependent, it is considered to be more universal. Nevertheless, the determination of its parameters can be troublesome, as they correspond to microcosmic damage theory and damage propagation. Later, the model was also adopted and developed by other researchers, including Marsh and Mackinnon [
12], Chen [
13], and, more recently, by Zhu et al. [
14]. Their attempts have focused mostly on the re-definition of the exponent interpreted as the inverse slope of a hypothetical s-n curve. Zhu et al. [
14] have introduced a dynamic Corten–Dolan’s model and demonstrated the better consistency of their model, although the determination of the introduced parameters for load interactions can be troublesome (approximates have been given). Furthermore, the definition of the introduced initial strength parameter seems to lack physical justification.
The s-n curve is also being used in models based on isodamage lines (e.g., by Subramanyan in Reference [
15]). This group of approaches seems to be consistent and easy to interpret; however, in the underlying assumptions, it can be found that the s-n curve corresponds to a failure probability equal to 1.
Constant research is also being conducted on the group of damage measures other than the damage degree, for example, evaluations on the plane of maximum normal or shear strains, such as the Smith–Watson–Topper (e.g., Reference [
16]) and Fatemi–Socie [
17] parameters. Both of them utilize the Coffin–Manson equation as the fatigue curve. Difficulties of the practical application of these approaches arise mostly because the algorithm for finding the critical plane orientation is complicated. It also must be noted that not all of the damage can be evaluated on the critical plane [
18].
Apart from those previously mentioned, non-standard approaches are also being developed (e.g., thermodynamic-based theories [
19]), but they are often difficult in practical applications due to the use of non-standard fatigue curves or the introduction of additional material constants (e.g., Reference [
20]).
From a statistical point of view, the reason for the scatter behavior of the critical damage degree lies in, among others, not considering the realistic fatigue life distribution. Another reason lies in the fact that the present formulation does not differentiate the damage degree between the stress levels, as well as it disregards possible material hardening at the first loading stages. It may not matter if the load spectra have similar character, but it can influence the results when the spectra are more diverse [
21,
22]; however, these effects do not arise in all materials [
23].
3. Probability-Modified Fatigue Damage Accumulation Model
The linear formulation of Miner’s rule narrows down the possible solutions for a problem. As an alternative, assume that the damage accumulation rule in the life region before the knee point follows a power–law formulation, which makes it capable of accurately modeling the damage accumulation process in the wide range of loading cycles [
9]:
where:
a, b—parameters corresponding to the loading.
In case of a multi-stage loading, it is rewritten as follows:
where:
ai, bi—values of the parameters for a given loading stage.
Suppose that the parameter
b can be defined as
From the probabilistic point of view, the critical damage degree should be viewed as the mean of the damage degree probability distribution,
, and it should correspond to the mean fatigue life,
. Due to the application of the least squares approach in the s-n curve generation process, the mean fatigue life is equal to the s-n curve number of cycles to failure only when the mean life in the assumed distribution of the fatigue life is the same as the median life. This is because the method of least squares is an approach in regression analysis that assumes the normal distribution of a set of experimental points. In the case of symmetrical probability distributions, such as a normal distribution, the mean life and the median life are the same. Normal, as well as log-normal or Weibull, distributions are usually assumed to be the distributions appropriate for estimating the probability of failure and the mean life. Due to the reasons already stated, the application of a normal distribution remains unjustified, which can be seen in Equation (4). Here, the exponent
a would be equal to 1 due to the equality between the mean and the median life, which automatically implies Miner’s rule with all of its inconveniences. In this way, the critical damage degree corresponds to the number of cycles from the reference s-n curve. In the proposed model, it is assumed that the critical damage degree remains at the same level (i.e.,
dc = 1), but the corresponding number of load cycles is shifted to lower values that result from the probabilistic s-n curve. The probabilistic s-n curve has its source in the appropriate probability distribution for the fatigue life. From the studies performed on P-s-n distributions (e.g., References [
24,
25,
26,
27]), it can be found that, in most cases, researchers have applied different variants of a Weibull distribution. A variant of a Weibull distribution tailored for a logarithmically distributed random variable of the fatigue life of steel welded joints in a one-stage loading (i.e., a constant amplitude) was proposed by the author of Reference [
4]. In general, the probability of failure
Pf in this variant can be described in the following manner [
28]:
where:
hl = log(nf)—scale parameter, s-n curve fatigue life;
gl = p/log(nf)—shape parameter;
p—parameter of the distribution estimation, constant within the entire s-n curve.
In Equation (6), fatigue lives lying on the s-n curve do not correspond to the 50% probability of failure (i.e., the median life), which can be seen in
Figure 2 and compared to
Figure 1.
In order to match the mean damage degree
with the mean fatigue life
, Equation (3) should be rewritten in a logarithmic space, as follows:
where:
dl—logarithm of d, dl = log(d)
nl—logarithm of n, nl = log(n).
It is worth mentioning that Equation (7) can also serve as a linear transfer function between the logarithm distributions of the fatigue life and the damage measure random variable.
Due to the fact that the process of fatigue degradation of the material is non-stationary [
29,
30,
31], the number of load cycles that represent the mean fatigue life corresponds to the mean damage degree only when it accumulates to its failure threshold—the damage degree critical value. Thus, the mean fatigue life directly corresponds to the mean damage degree only at the level of the damage critical value (
Figure 3):
where:
—the mean damage degree critical value, in a one-stage loading.
Based on the above, a relation for the exponent
a in the proposed damage accumulation rule (Equation (3)) can be formulated as follows:
It should be noted that, because the relation for
b is a function of stress (Equation (5)), such a formulation determines parameter
a as stress-dependent. From Equation (9), it can be seen that the exact value of
a also depends on the distribution of the fatigue life. While to some extent, the above approach resembles the one by Sun et al. [
32], the considerable differences can be found in the following underlying assumptions: (i) here, the power–law exponent
a is precisely defined and refers only to the number of cycles at the given loading stage; (ii) due to the above, the damage degree critical value is assumed to be constant on a given stress level, which makes the calculations less time consuming.
Compared to the existing models listed in
Section 2, the proposed model has three advantages—(i) the suggested non-linear formulation is based on the transformed linear transfer function for the probability distribution of a logarithm random variable rather than subjective experiences, so it has rigorous derivation, a theoretical basis, and an additional field of application; (ii) the applied power–law exponent becomes a slope of the linear transfer function, which makes it constant at a given stress level and constitutes this rule as capable of quantifying the dispersion of fatigue damage, which is dynamic during the process of fatigue damage accumulation; and (iii) it can be established based on the basic fatigue data of the material and the cumulative distribution function of the investigated stress level.
5. Discussion
In this study, a non-linear modification of the well-known Miner’s rule was proposed with the aim of reducing the scatter of the damage degree and investigating the relationship between the accumulated damage and the probability of failure. The proposed solution demonstrated superiority over Miner’s rule in two areas—determining the relationship between the failure probability and the damage degree and illustrating the tendency of the fatigue life random variable.
The results were obtained according to the underlying assumption of equivalence in the mean values of distributions of the fatigue life and the damage degree on the failure life level. As a result of this principle, the damage critical value corresponds to the mean fatigue life in any distribution, not only the standard normal distribution. In the case of this distribution, the model was reduced to the original Miner’s rule, which can serve as evidence for the inappropriateness of applying a normal distribution to fatigue analysis and relating the 0.5 probability to the fatigue life of the s-n curve. The mean value of the distributions used in the validation process in each case corresponded to probabilities other than 0.5, whereas, in the case of the distribution of a logarithm random variable, the results were found to have less scatter amongst the different stress levels. Comparing the probabilities that correspond to the damage degree values of 0.8 and 0.9 at different stress levels, it can be found that the probability differs amongst them, which validates the model as capable of reflecting the nature of the load–material interaction, which varies with different stress levels. In the case of the obtained damage degree, the biggest differences arose at the LCF region, as can be seen in
Figure 5, which makes it consistent with the theory that damage in the LCF and HCF does not accrue at the same rate, and nor does it accrue linearly with the applied cycle fraction, as it was underlined by Halford [
36]. This is the main advantage of the present model over the traditional Miner’s rule due to the fact that applying Miner’s rule usually provides unsatisfactory results with the history of loadings corresponding to the LCF region. In the case of such load histories, the predicted fatigue life is non-conservative [
9,
36], whereas the biggest differences between Miner’s rule and the proposed model were demonstrated to be in this region specifically. In the case of the
) random variable at the critical damage degree, the difference at the stress amplitude
MPa was about 1.5 × 10
5 load cycles.
The introduced modified damage accumulation rule reflects the central tendency of the fatigue life random variable, which could help in reducing the scatter of the accumulated damage. At the same time, the approach maintains the simplicity of its traditional version, applying standard design s-n curves into the calculation process. Additional attention may be needed in more complex applications, such as random loading or high–low/low–high multi-stage loadings. The formulation of parameter b remains open and can be reformulated without additional re-definitions of the other parts of the model since the power–law exponent a is a function of not only the mean fatigue life, but it also depends on the value of this parameter.
The adopted form of the proposed relation for the damage degree could help in the future advancement of research undertaken in the field of the probability distribution of the damage degree for multi-stage loading. While the phenomena of damage accumulation cannot be directly quantified, the appropriate distribution should be transformed from the distribution of the fatigue life random variable. The approach requires the probability distribution to be of the same type during the entire loading history, which imposes linear transfer functions as the appropriate relation to adopt in order to transform the fatigue life distribution into the distribution of damage measure, such as the damage degree. At the same time, linear transfer functions impose linear damage accumulation rules, the application of which remains questionable, as was stated previously. A logarithmically distributed random variable may be the solution, which would allow for the linear transfer function to be maintained, while the applied damage accumulation rule is non-linear and described according to the proposed relation. The function proposed for quantifying the fatigue damage becomes linear when transformed into the log-space, which allows for an easy transformation of the distribution of logarithm random variables, for example, the log(N) random variable. Additional future research and fatigue tests are required in order to apply the developed damage accumulation rule and validate it as a probability distribution transfer function.