^{1}

^{2}

^{3}

^{4}

^{4}

^{4}

The purpose of this research is to develop a model, with emphasis on compatibility conditions and model building, valid for high cycle fatigue design components such as wind turbines, automobiles, high speed railways and aeronautical material. In this work, we have added the frequency as one more variable to an existing fatigue model that already includes maximum stress, stress ratio and lifetime. As a result, a model and estimation method has been proposed and a random variable V has been identified, which, allows the accumulated damage and the probability of failure to be assessed for any load history in terms of stress levels, stress ranges and frequencies. Finally, the mod el is validated using a large set of real experimental data.

The current structural integrity concept in the fatigue design in aeronautics, electronics, railway material, rotating machines and many other fields demands the probabilistic fatigue characterization of the materials and the mechanical and structural components they are made off, so that a reliable safety level is ensured during their service life. With this aim, high performance test machines, from middle-high frequencies (200 to 1000Hz) [^{9}, or even 10^{10} cycles, where new fatigue mechanisms may arise other than those observed in the high cycle fatigue domain. Second, a remarkable reduction of test duration and cost implied when conventional servo-hydraulic or electric fatigue test machines are used. In addition, the correspondence among the fatigue S-N fields for different test frequencies, ranging from low till ultrasonic frequencies must be established and validated if a high quality congruent fatigue design is desired.

Traditionally, most fatigue models have been considered for low frequencies and they are two-dimensional, that is, involving a stress, for example, the stress range, Δ σ , and lifetime as the main variables. However, this was proved as not sufficient.

Fatigue has been considered from a long time ago, as one of the most important causes of failure, and being responsible for important accidents in machines, structures and bridges, for example. However, the appearance of models able to analyze the problem in detail has been slow and almost always below the expectations and, mainly, the needs. Today, it is known that a precise analysis of fatigue requires a random model including at least four variables: the stress variables, stress level and stress range, the frequency and the lifetime.

A simple revision of the existing literature shows that there is no unanimity in the selection of these two variables. Common models choose them from a set of six different candidate stress related variables: { σ M , σ m , Δ σ , σ a , R , σ m e a n } , which are the maximum stress, the minimum stress, the stress range, the stress amplitude, the stress ratio, and the mean stress, respectively. However, though in theory different choices are equivalent, as indicated in [

Recently a three dimensional model has been given, see [

Fatigue models started with the up-and-down method, that was designed for determining the endurance limit as its only variable. However, most fatigue models are bidimensional and include only one of the first two stress variables already mentioned and the lifetime, which are the models dealt with in this section. However, one must keep in mind that these models are valid only when the remaining two variables, the second stress and the frequency variables, are considered constant.

A significative advance in the development of fatigue models took place with the appearance of bidimensional S-N regression models, initially linear, later bilinear and finally of hyperbolic type. The selection of the mathematical structure of these models was initially based on convenience and simplicity of the models. This is the case of the linear models, which, when they were perceived as not adequate, a log transformation of one or the two axes was used to improve them. When this was not sufficient, the bilinear models were used as a simple and practical solution.

These regression curves, in the sense of being conditional means, were used extensively, but unfortunately its random character and the corresponding percentile curves were not fully introduced as a necessity to deal with the actual random nature of the problem.

Fortunately, curved trend models of random type were soon accepted and other models appeared. This was the case of the model in [

In this context, a special mention deserves the use of functional equations in the process of building models. These equations, as differential or integral equations, express properties that must be satisfied by the models being looked for. However, the unknown functions neither appear as derivatives, nor inside integrals, what makes the interpretation and handling of the problem more convenient, but without losses in its power.

As a reminder, in

One first comment on it consists in observing that these blue curves do not

intersect, as it occurs in the plot. This means that one specimen is more resistant than other no matter the value of Δ σ , what is a simplifying, but a reasonable assumption.

The first compatibility functional equation, used in [

F Δ σ ( Δ σ | N ) = F N ( N | Δ σ ) , (1)

where F Δ σ ( Δ σ | N ) and F N ( N | Δ σ ) are the cdf of Δ σ given N and the cdf of N given Δ σ , respectively. The variability field of both variables in a correct model must fulfil this strong condition. Otherwise, the fatigue model would become not valid.

Next, we assume that the random variables implied in the fatigue behavior definition belong to the same family of distributions, such as one of the extreme value distributions (Weibull, Gumbel, Fréchet) or any other location-scale family. This is a weak and natural assumption, because it would have no sense if for given units the distribution belongs to the family but it fails to belong when changing units of measure. Similarly, when asymptotes are involved it is natural to admit location changes so that changes in the asymptotes are included in the selected family of distributions.

If the cdfs involved in (1) are assumed strictly increasing, the functional equation

( Δ σ − λ 1 ( N ) δ 1 ( N ) ) β 1 ( N ) = ( N − λ 2 ( Δ σ ) δ 2 ( Δ σ ) ) β 2 ( Δ σ ) , (2)

with six unknown functions: λ 1 ( N ) , δ 1 ( N ) , β 1 ( N ) , λ 2 ( Δ σ ) , δ 2 ( Δ σ ) and β 2 ( Δ σ ) results.

A very general solution of this equation was found in [

H ( Δ σ , N ) = ( N − B ) ( Δ σ − C ) + E D = constant . (3)

Next, assuming validity of the weakest link principle for the lifetimes, the fatigue lifetime distribution represents a minimum problem asymptotically defined by the generalized extreme value family, see [

Consequently, these models were selected as the only appropriate ones for fatigue analysis when σ M is kept constant and when R is kept constant during the test, see [

Since most bidimensional fatigue models used in practice are written in terms of the lifetimes supplied for the stress range, for a fix value of the stress level, selected from the set of secondary variables { σ M , σ m , σ m e a n , R } , to reduce the effect of this limitation, a hypothetical equivalence among the different S-N fields for different stress levels is assumed, even though this is only a strong simplification. Some examples of this approach are the models of [

Unfortunately, these empirical type proposals are supported only by their simplicity and the associated fitting quality of the associated experimental results, but with no theoretical justification. A revision of some of these methods used in fatigue analysis can be seen in [

The final conclusion is that bidimensional models are not satisfactory for fatigue analysis, but the compatibility condition must hold for all bidimensional models coming from higher dimension models when the remaining variables are kept constant.

Since the same considerations used for the Δ σ − N models for constant σ M are also valid for constant R, this implies that we have two possible compatible bidimensional models to be used:

H ( Δ σ , N ; σ M ) = ( N − B M ( σ M ) ) ( Δ σ − C M ( σ M ) ) + E M ( σ M ) D M ( σ M ) (4)

and

H ( Δ σ , N ; R ) = ( N − B R ( R ) ) ( Δ σ − C R ( R ) ) + E R ( R ) D R ( R ) . (5)

Even though these two similar models can be used, if not carefully chosen they will be incompatible. In the following section the compatibility problem of both bidimensional models is stated and solved. This permits identifying the only compatible three-dimensional models of the form H ( σ M , R , N ) .

In [

H ( N , σ M , R ) = ( N − B M ( σ M ) ) ( σ M ( 1 − R ) − C M ( σ M ) ) + E M ( σ M ) D M ( σ M ) = ( N − B R ( R ) ) ( σ M ( 1 − R ) − C R ( R ) ) + E R ( R ) D R ( R ) , (6)

which depends on five unknown functions B M ( σ M ) , C M ( σ M ) , D M ( σ M ) , E M ( σ M ) , B R ( R ) , C R ( R ) , D R ( R ) and E R ( R ) . Even though its non-consideration implies incompatibility, most new fatigue models proposed in the last thirty years do not pay attention to this property, as indicated in [

In addition, some of the functions appearing in (6), which define model asymptotes, as indicated in [

C M ( σ M ) = 0 ; C R ( R ) = ( 1 − R ) σ M 0 . (7)

The functional Equation (6) and the asymptotes in (7) lead to the following Weibull model:

p = 1 − exp { − [ ( log e N N 0 − δ 1 R 1 − R ) ( 1 − R ) ( σ M σ M 0 − 1 ) − β 1 ( 1 − R ) − γ 1 a ] K } (8)

which depends on seven parameters: a , σ M 0 , N 0 , β 1 , δ 1 , γ 1 and K.

From (8), the following V Weibull random variable, which will play a key role in the cumulative damage evaluation, arises:

V ( N ; σ M , R ) = ( log e N N 0 − δ 1 R 1 − R ) ( 1 − R ) ( σ M σ M 0 − 1 ) − β 1 ( 1 − R ) (9)

The physical meaning of these parameters is as follows: a is the scale parameter of the V Weibull distribution; σ M 0 is the endurance limit for σ M ; N 0 is the N asymptote for R = 0 ; β 1 is the increment of the Weibull location parameter of the V variable for R = 0 with respect to R = 1 ; γ 1 is the Weibull location parameter of the V variable; δ 1 is the displacement of the log e N * asymptote for R = 1 / 2 with respect to R = 0 and K is the Weibull shape parameter of the V variable.

In addition, the associated regression models are:

σ M = σ M 0 { 1 + β 1 ( 1 − R ) + h ( γ 1 , a , p , K ) ( log e N N 0 − δ 1 R 1 − R ) ( 1 − R ) } ; N ≥ N 0 exp ( δ 1 R 1 − R ) , (10)

where

h ( γ 1 , a , p , K ) = γ 1 + a ( − log e ( 1 − p ) ) 1 / K (11)

is the random part of the model, and letting Δ σ = σ M ( 1 − R ) one gets:

Δ σ = σ M 0 { 1 − R + β 1 ( 1 − R ) + h ( γ 1 , a , p , K ) ( log e N N 0 − δ 1 R 1 − R ) } ; N ≥ N 0 exp ( δ 1 R 1 − R ) , (12)

which allows us to plot σ M and Δ σ versus log e N N 0 for constant R values, and

Δ σ = σ M δ 1 ( σ M σ M 0 − 1 ) + h ( γ 1 , a , p , K ) ( log e N N 0 + δ 1 ) ( σ M σ M 0 − 1 ) − β 1 ; N ≥ N 0 exp ( β 1 σ M σ M 0 − 1 − δ 1 ) , (13)

which allows us to plot Δ σ versus log e N N 0 for constant σ M values.

The asymptotes of the curves in (10), (12) and (13) are given in

Next, a physical interpretation of the V variable defined in Expression (9) is given.

The V variable, which has arisen in a natural form, when evaluating the probability of failure of this model and imposing the compatibility constraints, can be interpreted as damage level associated with an effective load, the pair ( σ M , R ) and the lifetime N. In other words, it can be associated with the damage caused to the material when subjected to N cycles under a load ( σ M , R ) for constant frequency. Consequently, using the V variable in (9), a damage comparison can be made of two specimens subject to different load histories, when the frequency is constant. This is relevant, because later, based on this variable, the damage accumulation associated with any varying load history, will be calculated.

Equation | Horizontal asymptote | Vertical asymptote |
---|---|---|

(10) | σ M = σ M 0 | N = N 0 exp ( δ 1 R 1 − R ) |

(12) | Δ σ = σ M 0 ( 1 − R ) | N = N 0 exp ( δ 1 R 1 − R ) |

(13) | Δ σ = 0 | N = N 0 exp ( β 1 σ M σ M 0 − 1 − δ 1 ) |

In this section the basic ideas and formulas used to reproduce the effect of frequency on fatigue tests are discussed. It is assumed that the influence of frequency on fatigue is manifested in the variation of the elastic modulus tending to an effective elastic modulus, E e f . The observation is consistent with experimental experience that the speed in a static test influences the apparent modulus of elasticity.

The higher the test speed, the more rigid the material behaves. Therefore, the higher the frequency, the higher the effective elastic modulus, E e f . Asymptotically, when the frequency is practically zero, the behavior of the material, i.e. the effective modulus of elasticity, is the nominal one, so E e f = E 0 . When the frequency is very high, the effective modulus of elasticity tends to an asymptotic value, E ∞ .

The effect of frequency on the S-N fields has been studied in [

All this is in accordance with a viscoelastic simile of the material, in the sense that for high values of frequency, the behavior of the material tends to reach an asymptotic value according to a sigmoidal function, such as a biparametric Weibull cdf. If we observe a cycle in fatigue under the condition of test in load control, this assumes that the induced deformation in a load cycle decreases as the frequency increases, as it can be seen in

Since deformation is the determining mechanism in the fatigue process, the range of deformations deduced from the range of stresses, assuming that the static modulus of elasticity is the key factor. This implies that in reality, a lower deformation is being applied according to the relationship E 0 / E e f and, accepting the linear relationship between strain and stress given by E 0 , for the representation in the S − N field, a shift to the right of the S − N curves will be observed as the frequency increases.

According to this, we can normalize the Young modulus and, since the normalized range becomes the closed interval [ 0,1 ] we can assume a Weibull cdf for the logarithm of the frequency, as follows:

E ¯ = E e f − E 0 E ∞ − E 0 = 1 − exp [ − ( log e ( f / f 0 ) τ ) β ] , (14)

where f 0 is a reference frequency and τ is a dimensionless scale parameter. Note that the Weibull distribution has been justified for the logarithm of f / f 0 , a dimensionless variable, as it must always be for logarithm and trigonometric functions, a condition frequently ignored. Then, we have

E e f = E 0 + ( 1 − exp [ − ( log e ( f / f 0 ) τ ) β ] ) ( E ∞ − E 0 ) . (15)

From now on, we call

a e f = E 0 E e f ≤ 1. (16)

which, from (15) gives

a e f = 1 1 + ( 1 − exp [ − ( log e ( f / f 0 ) τ ) β ] ) ( E r a t i o − 1 ) (17)

where

E r a t i o = E ∞ E 0 ≥ 1. (18)

Accordingly, the value of σ M and the stress amplitude are affected by the factor a e f , while the mean stress is not affected. This is due to the way of carrying out the test, in which the load is placed in the medium stress and then it is expanded until it reaches the desired amplitude. With which, it turns out that:

σ M , e f = σ m e a n , e f + σ a , e f = σ m e a n , n o m + a e f σ a , n o m (19)

= σ M , n o m + σ m , n o m 2 + a e f σ M , n o m − σ m , n o m 2 (20)

= ( 1 + a e f ) σ M , n o m + ( 1 − a e f ) σ m , n o m 2 (21)

= [ ( 1 + a e f ) + ( 1 − a e f ) R n o m ] σ M , n o m 2 (22)

and then,

R e f = σ m , e f σ M , e f = σ m e a n , n o m − σ a . e f σ m e a n , n o m + σ a . e f (23)

= σ m e a n , n o m − a e f σ a . n o m σ m e a n , n o m + a e f σ a . n o m = σ M + σ m 2 − a e f σ M − σ m 2 σ M + σ m 2 + a e f σ M − σ m 2 (24)

= ( 1 − a e f ) σ M + ( 1 + a e f ) σ m ( 1 + a e f ) σ M + ( 1 − a e f ) σ m = ( 1 − a e f ) + ( 1 + a e f ) R n o m ( 1 + a e f ) + ( 1 − a e f ) R n o m . (25)

Referring now to the nominal values, those intended in the tests, the following efective σ M , e f and R e f values must replace the nominal ones, σ M , n o m and R n o m in (10), (12) and (13):

σ M , e f = [ ( 1 + a e f ) + ( 1 − a e f ) R n o m ] σ M , n o m 2 ; (26)

R e f = ( 1 − a e f ) + ( 1 + a e f ) R n o m ( 1 + a e f ) + ( 1 − a e f ) R n o m (27)

or their inverses:

σ M , n o m = ( ( 1 + a e f ) − ( 1 − a e f ) R e f ) σ M , e f 2 a e f ; (28)

R n o m = ( 1 − a e f ) − ( 1 + a e f ) R e f ( 1 − a e f ) R e f − ( 1 + a e f ) . (29)

This means that the influence of the frequency on fatigue is equivalent to a change in the σ M and R values by their equivalent nominal ones σ M , e f and R e f and that there exists a unique fatigue limit, independently of the frequency and R.

If the model turns out to be correct, the frequency characterization of a material, that is, the determination of the curve ϕ = ϕ ( f ) would be easily determinable. It would only be a matter of applying a relatively high stress at different frequencies and measuring the deformation due to that stress, which would allow us to observe the evolution of the modulus of elasticity as a function of frequency.

The model proposed in this paper starts from the model developed in [

It is clear that this model and consequently, the following equations must be applied to the effective values of σ M and R, because, even though the material is apparently subject to given nominal values of σ M and R, the frequencies modify these actual values to the effective ones. Thus, the following equations refer to effective values.

In the previous model it was demonstrated that the variable in (9) plays a crucial role, because all the remaining variables can be considered as deterministically dependent on this one. In other words, the model can be written in a form such that its deterministic and its random parts are identified, that is, they become separable. In addition, since the variable V is unidimensional, only the random parameters associated with it rule the random behavior. Obviously, the remaining model parameters are deterministic.

Replacing σ M , R and Δ σ by σ M e f , R e f and Δ σ e f , respectively, in (10), (12) and (13) one gets:

1) σ M e f for given R e f :

σ M e f = σ M 0 { 1 + β 1 ( 1 − R e f ) + h ( γ 1 , a , p , K ) ( log e N N 0 − δ 1 R e f 1 − R e f ) ( 1 − v R e f ) } ; N ≥ N 0 exp ( δ 1 R e f 1 − R e f ) . (30)

2) Δ σ e f = σ M e f ( 1 − R e f ) for given R e f :

Δ σ e f = σ M 0 { 1 − R e f + β 1 ( 1 − R e f ) + h ( γ 1 , a , p , K ) ( log e N N 0 − δ 1 R e f 1 − R e f ) } ; N ≥ N 0 exp ( δ 1 R e f 1 − R e f ) , (31)

3) Δ σ e f for given σ M e f :

Δ σ e f = σ M , e f δ 1 ( σ M e f σ M 0 − 1 ) + h ( γ 1 , a , p , K ) ( log e N N 0 + δ 1 ) ( σ M e f σ M 0 − 1 ) − β 1 ; N ≥ N 0 exp ( β 1 σ M e f σ M 0 − 1 − δ 1 ) , (32)

where δ 1 , β 1 ≥ 0 and σ M e f > σ M 0 for the model to be valid.

These expressions, allow us to plot σ M e f and Δ σ e f versus log e N N 0 for constant R e f values and Δ σ e f versus log e N N 0 for constant σ M e f values (some examples can be seen in the same paper [

It is important to realize that the common term h ( γ , a , p , K ) in (11) appearing in expressions (30) (31) and (32) refers to the Weibull distribution of the V in Expression (9), where γ , a and K are its location, scale and shape parameters, respectively. Alternatively, other family of distributions can be selected to reproduce the random behavior of V.

The asymptotes of the curves in (30), (31) and (32) are given in

Note finally, that the role of the V damage variable remains valid, when frequencies are considered if the effective stresses are used instead of the nominal ones.

To estimate the ten model parameters, a , N 0 , σ M 0 , β 1 , δ 1 , τ , γ 1 , f 0 , β and K, a two-step method is used. In the first step the seven deterministic parameters, { N 0 , σ M 0 , δ 1 , τ , β 1 , f 0 , β } , are estimated using a regression estimate based on

Equation | Horizontal asymptote | Vertical asymptote |
---|---|---|

(30) | σ M e f = σ M 0 | N = N 0 exp ( δ 1 R e f 1 − R e f ) |

(31) | Δ σ e f = σ M 0 ( 1 − R e f ) | N = N 0 exp ( δ 1 R e f 1 − R e f ) |

(32) | Δ σ e f = 0 | N = N 0 exp ( β 1 σ M e f σ M 0 − 1 − δ 1 ) |

least squares, which minimizes the sum of error terms, that is, the differences between the observed values, σ M f i , and those given by the regression curves of Expression (10), which correspond to the V mean value V ¯ = γ 1 + a Γ ( 1 + 1 K ) . In other words, the following optimization problem is solved:

Minimize N 0 , σ M 0 , δ 1 , τ , β 1 , V ¯ , f 0 , β Z = ∑ i = 1 n [ σ M e f i − σ M 0 ( 1 + β 1 ( 1 − R e f i ) + V ^ ( log e N i N 0 − δ 1 R e f i 1 − R e f i ) ( 1 − R e f i ) ) ] 2 (33)

where M e f i and R e f i depend on f 0 , τ and β , subject to

N i ≥ N 0 exp ( δ 1 R e f i 1 − R e f i ) , (34)

δ 1 , β 1 , N 0 ≥ 0, (35)

σ M e l l i ≥ σ M 0 , (36)

where n is the sample size.

In the second step, and keeping V ¯ = γ + a Γ ( 1 + 1 K ) , that is, the mean of V,

constant, the random parameters, { γ , a , K } are obtained by restricted maximum likelihood. This restriction guaranties that the previously estimated regression curves are not modified and also avoids overfitting. This means that the following optimization problem is solved:

Minimize γ , a , K Z = ∑ i = 1 n ( − log e K + log e a + W i K − ( K − 1 ) log e W i ) (37)

subject to K ≥ 0 , where the expression between parentheses in (37) is the log-likelihood, changed sign, and

W i = v i − γ a ; i = 1 , 2 , ⋯ , n . (38)

This optimization problem has been implemented in matlab, using the function “fmincon”. One example of application is given in Section 9.

The damage interpretation of the V variable in (9) offers us the possibility of deriving from it a damage accumulation conversion law, which will allow us to calculate the probability of failure derived from any varying load history.

It has already been indicated that the V expression in (9) gives the cumulative damage associated with N cycles when the load σ M e f and R e f is applied assuming that the initial damage is zero. In addition, its cdf gives the probability of failure, associated with N cycles when that load is applied assuming a threshold value, V 0 , associated with a threshold number of load cycles, for which the probability of failure is zero. Thus, the damage level associated with N load cycles and zero initial damage, is:

V ( N , σ M e f , R e f ) = Q ( 0, N , σ M e f , R e f ) . (39)

However, we look for a more general formula

V = Q ( V 0 , N , σ M e f , R e f ) , (40)

allowing us to consider how the damage level evolves after a given damage, V 0 , is reached. This is one of the aims of this section.

We start by saying that the Expression in (40) cannot be given arbitrarily. Otherwise, using different threshold damages will lead to different final damage levels even though the loads and the number of cycles coincide (see, for example, [

To avoid this problem, the formula to obtain V in (40) must be subjected to a compatibility condition, which states that the same damage level must be obtained if (a) we start from a V 0 threshold damage and apply N 0 + N cycles of the load or (b) we start from the damage level associated with N 0 cycles and afterwards we apply N cycles of the same load. This, compatibility condition can be written as:

Q ( V 0 , N 0 + N , σ M e f , R e f ) = Q ( Q ( V 0 , N 0 , σ M e f , R e f ) , N , σ M e f , R e f ) , (41)

which is a functional equation, with its only unknown function being Q ( V 0 , N , σ M e f , R e f ) (see [

Q ( 0, N 0 + N , σ M e f , R e f ) = Q ( Q ( 0, N 0 , σ M e f , R e f ) , N , σ M e f , R e f ) . (42)

and using (39) twice one obtains:

V ( N 0 + N , σ M e f , R e f ) = Q ( V ( N 0 , σ M e f , R e f ) , N , σ M e f , R e f ) . (43)

and letting V ( N 0 , σ M e f , R e f ) = V 0 in (43) the solution of Equation (41) is obtained:

Q ( V 0 , N , σ M e f , R e f ) = V ( V − 1 ( V 0 , σ M e f , R e f ) + N , σ M e f , R e f ) , (44)

where the inverse of the V function refers to its first argument and it has been assumed that it exists.

If the load history is defined as a set of loads { ( σ M e f ( j ) , R e f ( j ) ) ; j = 1 , 2 , ⋯ , m } , for given frequencies, repeated for N ( j ) cycles, the partial and the final damage states can be calculated as follows:

V 0 ( j ) = Q ( V 0 ( j − 1 ) , N ( j ) , σ M e f ( j ) , R e f ( j ) ) = V ( V − 1 ( V 0 ( j − 1 ) , σ M e f ( j ) , R e f ( j ) ) + N ( j ) , σ M e f ( j ) , R e f ( j ) ) ; j = 1 , 2 , ⋯ , m (45)

where V 0 ( j − 1 ) is the threshold damage for load { σ M e f ( j ) , R e f ( j ) } , and the initial and final damage levels are V 0 = V 0 ( 0 ) and V 0 ( m ) , respectively.

It is obvious that these two formulas become crucial when one needs to evaluate damage states caused by different load histories, as it happens in real practice.

In this example we have a set of 243 fatigue test data. The samples consisted of commercial high purity Al wires (type Al-H11 from Heraeus in as-received condition with a diameter of 400 μm and a gauge length of 10 mm. A TA ElectroForce DMA 3200 testing machine with a load capacity of 500 N, a frequency range of 0.00001 - 300 Hz and a displacement range of ±6 mm was used for the experiments. The load controlled fatigue tests were carried out in tension-tension mode at a stress ratio of R ≈ 0.1 by using 6 - 10 samples at each stress level. The S-N curves were obtained by collecting the data in a broad range of 10^{2} up to about 10^{8} loading cycles at three testing frequencies of 2, 20 and 200 Hz, and the model parameters have been estimated using the methods described in Section 7 and commented below. E_{ratio} is assumed equal 3.

The resulting lifetimes are plotted in

A simple look to the data in

advance that these five outliers, which are only 2% of the 243 sample data points, were selected using both, probability papers and analyzing to percentile curves of the fitted model including all data points (see Figures 4-8).

The two-step method described in detail in Section 6 has been used to estimate the parameters with all data points and the resulting fitted model plotted to analyze the possibility of some outliers, as revealed by our first look to the data points. The outliers were confirmed based on Figures 4-8, but this can also be done using well known methods, such as, for example, those given in [

Once this has been done and the outliers identified, they have been removed from the sample and the estimation and possible detection of more outliers repeated to select the final model for its use in practical applications, such as an analysis of safety against fatigue in high speed railway bridges.

For the sake of facilitating the comparison of both results, with and without outliers, and to analyze how them affect the final results, the plots corresponding to both estimation processes have been grouped, such that the left plots refer to the whole data and the right plots to the removed outliers case, respectively.

The first estimation step, in which the regression curves are obtained using least squares, provides the seven deterministic parameter estimates and the second step gives the three random ones, obtained by restricted maximum likelihood, assuming a Weibull distribution for the residuals. They are shown on the left and right parts of

Data point i | Lifetime (cycles) | σ M MPa. | Data point i | Lifetime (cycles) | σ M MPa. |
---|---|---|---|---|---|

131 | 31,517 | 40.00 | 150 | 94 825 | 37.80 |

209 | 92,666 | 39.79 | 242 | 2 293 106 | 33.00 |

243 | 33,714,952 | 31.99 |

Outliers | Deterministic parameters | Random parameters | ||||||||
---|---|---|---|---|---|---|---|---|---|---|

σ M 0 | N 0 | β 1 | δ 1 | f 0 | β | τ | a | γ | K | |

Included | 18.96 | 0.388 | 0.174 | 0 | 1.33 × 10^{−}^{5} | 2.63 | 12.1 | 1.40 | 0.91 | 9.60 |

Removed | 18.92 | 0.347 | 0.104 | 0.143 | 1.56 × 10^{−}^{5} | 2.63 | 11.8 | 0.76 | 1.59 | 6.66 |

a close to linear general trend but showing some outliers at both ends. As already indicated, the five outliers have been selected using this plot and the following percentile plots. They clearly show that we are in front of outliers, some of them confirming our first impression after looking to the data in

The right plot in

The plots in

It is clear that the Weibull distribution of V is justified, theoretically and physically, by extreme value theory, because fatigue failure is a process that takes place using the weakest link or weakest region principle, stating that failures occur at the weakest locations.

Alternatively, in the second step, the errors or residuals, V i , are also calculated and plotted, in

The plots in

respectively. It is interesting to see how the top plots allow identifying the outliers on their left and right zones.

If the Weibull model is used, the resulting regression curves and the associated percentiles are shown in

Finally,

selected model. As it can be seen, this also reveals the presence of some outliers, corresponding to the two sides of the regression curve.

Note also that the outliers also appear in

The regression curves σ M versus N and σ M e f versus N corresponding to the kernel estimates together with the associated percentiles, as those in

Finally, the resulting diagram of E e f ∗ versus frequency is shown in

In summary, it can be concluded that, at least in this example including a

large size data set, the model behaves well and as expected. Future analysis will be required to confirm the use of this model. However, the possibility of a four dimensional analysis, that is, including frequencies as the fourth variable is now possible.

The most important conclusions from this paper are:

1) The necessity of replacing or improving existing models, broadly acknowledged and recommend by standards and guidelines, needs to be emphasized to avoid the consequences of application of invalid models as reported in the literature.

2) A new method for considering the joint effect of maximum stress, stress ratio and frequency on the lifetime is presented in Sections 0 to 1. The frequency is included as a new and relevant variable to be considered in the design of important high-risk components such as wind turbines and automobiles, high speed railways and aeronautical material.

3) It has been shown that compatibility conditions are important tools to build models without arbitrary assumptions. These conditions were easily stated as functional equations, in Sections 2 to 4, which when solved provide the unique models satisfying the corresponding conditions. To our knowledge, no other method exists satisfying these conditions.

4) The model shows, in Section 5, that the frequency effect is equivalent to a concurrent modification of the maximum stress-stress ratio pairs from each test, what has immediate consequences in how to design the tests conducting to a precise parameter estimation.

5) The model parameters are classified, in Section 7, into two distinct classes: those defining the deterministic model, which are estimated by least-squares regression, and those reproducing the random ones, estimated by a restricted maximum-likelihood method. An estimation method is proposed and applied independently to both classes to avoid overfitting, in Section 7, and tested later in Section 9.

6) A model damage measure variable, V, is identified in Section 4, which together with its cdf, permits the damage accumulation to be evaluated and the associated probability of failure to be determined subject to any varying load history. A general formula is given in Section 8, which allows the cumulative damage to be analytically calculated from the joint effects of the three changing variables, maximum stress, stress ratio and frequency.

7) Due to its integrated character, because it includes all variables, the use of the model, will certainly allow us to reduce the sample size, and also to optimize the test sequences when used in designing test studies.

8) A more general formula that provides the damage of any load when the initial damage is known is obtained in Section 8 as the result of a functional equation that guaranties the compatibility, that is, damage to become independent on the way the loads are considered.

9) An unusually, rarely available large sample data, from an experimental campaign in which 243 fatigue specimens were tested, was used to validate the model, in Section 9, resulting very good results, as shown in the figure examples.

G. Khatibi, B. Czerny, and M. Zareghomsheh gratefully acknowledge the financial support by the Austrian Federal Ministry for Digital and Economic Affairs and the National Foundation for Research, Technology and Development for the support provided to the experimental program in this work.

The authors declare no conflicts of interest regarding the publication of this paper.

Castillo, E., Fer- nández-Canteli, A., Blasón, S., Khatibi, G., Czerny, B. and Zareghomsheh, M. (2021) Step-by-Step Building of a Four Dimensional Fatigue Compatible Regression Model including Frequencies. Open Journal of Statistics, 11, 1072-1096. https://doi.org/10.4236/ojs.2021.116064