1. Introduction
Deepwater risers are important equipment connecting the floating platform with subsea wellhead during offshore oil and gas development. Their main function is to provide fluid passage between the surface platform and subsea wellhead in drilling and production processes [
1]. Deepwater risers are sensitive to dynamic loads, and fatigue damage due to ocean waves and movements of the floating platform is one of their main failure modes. The long term accumulation of fatigue damage leads to the fatigue fracture of risers and even leads to major accidents [
2]. Several accidents, including oil leaks and subsea wellhead breakage, have occurred due to riser fatigue [
3,
4]. Consequently, studying the fatigue reliability of risers during drilling and production processes is essential.
Fatigue analyses are the basis of fatigue reliability analysis for deepwater risers. Physical-based models are widely used in riser fatigue analyses at present. The fatigue analysis schemes used by these models can be divided into frequency and time domain methods. The frequency domain method is rarely used because of possible linearization errors in the calculation. By contrast, the time domain method is often employed in riser fatigue analyses due to its high computational accuracy. In the time domain method, the dynamic response of the riser system is first analyzed using the finite element method, the finite difference method, and other approaches [
5,
6]. Second, the dynamic response is processed through the rain flow counting method. Lastly, fatigue damage is calculated based on the S–N curve or fracture mechanics [
7,
8]. The time domain fatigue of deepwater risers, including drilling, steel catenary, and flexible risers, has been widely investigated in previous studies.
Research on riser fatigue reliability has mainly focused on catenary and top-tensioned risers. The process of fatigue reliability analysis is relatively complex. Generally, the assessment of the fatigue capacity and fatigue failure of risers is conducted by applying two different methodologies, namely, crack growth rate curves using the fracture mechanics model and S–N curves using the Palmgren–Miner damage model [
9,
10]. During the fatigue reliability analysis of risers, the randomness of general dynamic model parameters, such as wall thickness, the additional mass coefficient, and the drag force coefficient, is often considered [
11,
12,
13,
14]. The uncertainty of the fatigue performance parameters of the S–N curve method and that of the Paris parameters of the fracture mechanics method are also considered [
15,
16,
17,
18]. In particular, the uncertainty of geotechnical model parameters and the nonlinear behavior of the pipe–seabed interaction are also considered when evaluating the fatigue reliability of catenary risers in the contact area [
19,
20,
21]. The uncertainty of these parameters results in the uncertainty of riser fatigue. Fatigue assessment using certain parameters only cannot meet the needs of safety design and riser assessment. The fatigue reliability of risers must be assessed reasonably based on statistical theory [
22]. Several practical reliability analysis approaches, including numerical simulation and analytical algorithms, have been developed for riser fatigue reliability assessment. The Monte Carlo method belongs to the former type of reliability analysis, whereas the first order and second order reliability methods belong to the latter type. The three methods are widely used, and each has its own advantages and disadvantages. The Monte Carlo method has the highest solving accuracy but the lowest solving efficiency, and the first order reliability method has a solving performance that is opposite to that of the Monte Carlo method; the second order reliability method is relatively mediocre when compared with the two other methods [
23,
24].
In previous research, physical-based models were widely used for fatigue and reliability analyses of risers. However, the calculation for single fatigue analysis using physical-based models is time consuming [
25]. The calculation cost of fatigue reliability analysis using physical-based models increases significantly with the increase in the number of random parameters and the degree of random distribution of the parameters. These reasons account for the low efficiency of the fatigue reliability analysis of risers. Therefore, some scholars have attempted to improve the efficiency of fatigue reliability analysis by reducing the time of single fatigue analysis or the scale of fatigue analysis [
26,
27]. Data-driven models have been introduced as substitutes for time consuming physical-based models [
25,
28,
29,
30]. Notably, risers need to be inspected in a fixed time to guarantee their safety, and using the inspection results to update the fatigue reliability is essential to the safety of risers in the process of drilling and production. A problem at present is the difficulty of introducing inspection results into riser fatigue reliability analysis. The dynamic Bayesian network (DBN) is an optional method for solving this difficult problem. The DBN has increasingly been applied to probability analysis because of its functions of forward and backward analyses [
31]. Fatigue reliability can be calculated based on the prior probabilities of root nodes in forward analysis, and the posterior reliability can be updated based on a few observations. Thus, the DBN was selected for riser fatigue reliability analysis and updating based on the inspection results in this study.
In this study, data-driven models based on the RSM were established to substitute for physical-based models in the fatigue analysis of risers. The direct mapping between riser dynamic analysis parameters and fatigue damage was established by the data-driven models, which can greatly improve the efficiency of fatigue analysis. The annual fatigue crack growth model was then established based on fracture mechanics theory and data-driven models that consider the crack inspection data as a factor. The crack growth DBN model was then established to obtain the fatigue reliability of the risers. The crack inspection data of the risers were introduced into the DBN model to update the fatigue reliability of the risers.
The rest of the paper is organized in the following manner. The riser fatigue reliability analysis methodology is introduced in
Section 2. The physical-based models are introduced in
Section 3. The data-driven models established by RSM are introduced in
Section 4, and the fatigue reliability analysis method based on DBN is introduced in
Section 5. A case is introduced for fatigue reliability analysis and updating in
Section 6. The conclusions are summarized in
Section 7.
3. Physical-Based Models
Figure 2 shows a deepwater drilling platform/riser coupling system that comprises a platform, a drilling riser system, and tensioners. The top end of the riser system is hung on the drilling platform via the tensioners and upper flex joint, and the bottom end of the riser system is connected to the Blowout preventer (BOP), the subsea wellhead, and the conductor via the lower marine riser package (LMRP). The riser system vibrates transversely and axially under environment loads and platform motions.
The transverse vibration equation of the riser system can be written as [
32,
33]
where
m is the mass per unit length of the riser,
c is the damping coefficient,
E is the elastic modulus of the riser,
I(
x) is the second moment of area,
T(
x) is the effective axial tension, and
Fsea(
x,
t) is the marine load applied per unit length of the riser.
The axial vibration equation of the riser system can be written as [
32,
33,
34]
where
u(
x,
t) is the axial vibration,
F is the axial hydrodynamic load on the riser system, and
A(
x) is the area of the riser cross-section.
The effective axial tension,
T(
x), can be expressed as
where
Ttop is the tension at the top of the riser system, and
L is the length of the riser system.
The marine load applied per unit length of the riser,
Fsea(
x,
t), can be written as
where
ρ is the density of seawater,
CM is the coefficient of inertia force,
Ca is the coefficient of additional mass (
Ca =
CM-1),
Dh is the hydrodynamic outer diameter,
Cd is the coefficient of drag force,
uw is the water point velocity,
is the water point acceleration,
uc is the current velocity,
is the velocity of the riser, and
is the acceleration of the riser.
The top end of the risers is connected to the platform by the upper flex joint, and the bottom end is connected to the LMRP via the lower flex joint, as is shown in
Figure 2. Therefore, the top and bottom boundary conditions of the riser system are expressed as [
35]
where
Kb is the rotation stiffness of the lower flexible joint,
Ku is the rotation stiffness of the upper flexible joint,
L is the length of the risers,
Spx(
t) is the floating platform’s heave displacement, and
Spy(
t) is the floating platform’s drift distance.
Spx(
t) and
Spy(
t) can be calculated by Equation (7) [
36].
where
Sn is a single wave’s amplitude,
Nw is the number of random waves,
kn is the number of constituent waves,
ωn is the wave frequency,
φn is the initial phase angle of the wave, and
αn is the phase angle between the platform movement and wave.
On the basis of the motion equation of the riser system, the finite element method was used to establish the finite element analysis model of the riser system [
37,
38]. The whole dynamical equation of the riser system is written as
where {
u} is the displacement vector of the riser system; [
M], [
C], and [
K] are the mass, damping, and stiffness matrices of the riser system, respectively; and {
F} is the load vector of the riser system.
The dynamic response of the risers was obtained by solving the whole motion finite element model of the riser system. The rain flow counting method was then used to count the fatigue stress time history and obtain the stress cycle and amplitude [
7,
8].
where
DL is the long term fatigue damage,
nf is the fatigue life, and
C and
mf are the S–N curve method parameters.
5. Dynamic Bayesian Network
The Bayesian network (BN) is utilized to analyze the fatigue reliability of risers due to its forward and backward inference capability. The BN is a combination of probability and graph theories and is widely used in data analysis [
43]. The structure of a BN model is clearly represented by graph theory, and the problems to be solved are analyzed by probability theory [
44,
45,
46]. A basic BN is shown in
Figure 3.
Each node is attached to a probability distribution. Root nodes
A and
B are attached to their marginal probability distributions
P(
A) and
P(
B), respectively. The non-root node
X (
C,
D,
E) is attached to the conditional probability distribution
P(
X|pa(
X)), and
pa(X) is the father node of node
X [
47]. A logical analysis is performed by analyzing the relationship between nodes in the Bayesian model.
Prior and posterior probabilities are relative to a certain set of evidence in Bayesian theory. In the presence of two different random variables
A and
D,
A =
a is an event, and
D =
d is evidence. The probability estimation
P(
A =
a) of event
A =
a before considering evidence is called prior probability. Conversely, the probability estimation
P(
A =
a|
D =
d) of event
D =
d after considering evidence is called posterior probability. The relation between prior and posterior probabilities is described as
where
P (
D =
d A =
a) is the likelihood of event
A =
a and can be called
L (
A =
a|
D =
d). In reality, the likelihood of the event is easier to obtain than the posterior probability.
The DBN is an extension of static Bayesian network modeling in the time dimension. The state of the DBN is updated over time to analyze dynamic stochastic processes. Some assumptions need to be made before establishing a DBN model:
The conditional probability is uniformly stable for all variable t over a finite period of time.
The DBN is a Markov process. Future probabilities are determined by present probabilities and not by past probabilities, namely, P(Y[t+1]|Y[1],Y[2],…,Y[t]) = P(Y[t+1]|Y[t]).
The conditional probability of adjacent time slices is stable. P(Y[t+1]|Y[t]) is independent of time.
The DBN is composed of a prior network and a transfer network. The prior network is a joint probability distribution defined at the initial state. The transfer network is the conditional transition probability
P(
Y[
t+1]|
Y[
t]) defined by the variables
Y[
t] and
Y[
t+1]. The DBN is shown in
Figure 4.
A DBN model of riser fatigue crack was established in this study to analyze the fatigue reliability of risers, as is shown in
Figure 5. The crack depth in Equation (21) is the parent node, and the other variables are the child nodes. The distribution of fatigue crack size can be calculated through forward Bayesian inference. The inspection results can also be introduced to the DBN model, and the influence of the inspection results on riser failure probability is examined using backward Bayesian inference.
7. Conclusions
This study attempts to solve a challenging problem: the fatigue reliability analysis of deepwater risers. An improved fatigue reliability analysis method based on RSM and the DBN for deepwater risers was proposed. A more efficient data-driven model was established based on RSM compared with physical-based models. A crack growth model of the riser was established by considering the crack inspection data as a factor. The crack DBN model was employed to conduct a fatigue reliability analysis and updating. The combination of data-driven models and the crack growth model could solve the problems of traditional methods in the fatigue reliability analysis of risers, such as low computational efficiency and the inability to introduce inspection data.
The whole fatigue reliability analysis process includes fatigue damage analysis, fatigue damage modeling, and fatigue reliability analysis. First, the fatigue damage of the riser system was analyzed based on physical-based models. Second, a data-driven model for riser fatigue damage was established based on the fatigue damage analysis results, and, further, a riser fatigue crack growth model was established. Lastly, a fatigue crack DBN model was established for fatigue reliability analysis based on the fatigue crack growth model, and the fatigue reliability of the riser was updated by taking the crack inspection data as evidence.
The proposed method was applied to the fatigue reliability analysis and updating of deepwater risers. The results showed that fatigue reliability could be calculated efficiently through dynamic Bayesian forward inference. Moreover, the distribution of random variables and fatigue reliability could be updated based on the inspection results via the backward inference of the DBN model. The proposed method could analyze and update the fatigue reliability of deepwater risers accurately and efficiently.