November 2, 2020
Machine learning for subsea pipeline reeling mechanics
Reel-lay is an attractive method for offshore pipeline installation. It allows to lay a pipeline on the seabed at a very high rate compared to conventional methods. Moreover, onshore controlled welding and inspection conditions are more convenient than in offshore conditions. This pipeline installation method is suitable for relatively flexible pipelines with small diameters, in all water depths.
Figure 1. The reel-lay method is considered to be the fastest pipeline installation method.
Reel-lay vessels have a drum on the deck, which allow to reel and pull-out the pipeline. The pipeline segments, sometimes referred to as stalks, are welded onshore and stored on long racks waiting for the installation vessel. Stalks are welded in long pipe segments, typically several kilometers in length, and reeled onto this large diameter drum, typically of 10 meters radius. When using reeling, local buckling must be properly assessed in terms of probability of occurrence, safety level of the installation, and cost of the reel-lay process. Buckling is a local failure event which is defined as the loss of the round cross-sectional shape of the pipeline on the drum.
Figure 2. Example of pipeline buckling.
This phenomenon has been established and described extensively for different external conditions, loads and pipe shapes and material [1-3]. However, designing a pipeline for reeling still requires the intensive use of numerical simulation.
Numerical modeling of such non-linear system is challenging. Here, it is modeled by Finite Element Analysis (FEA) implemented in Abaqus, with a material model that considers elastic-plastic steel behavior based on Ramberg-Osgood formulation [4, 5].
Figure 3. Key aspects of the Finite Element Model we use to simulate the reeling process.
The interaction between pipeline and reel is defined by hard contact with allowance of separation (lift-off) in the direction normal to the contact surface. The variables of the Finite Element model are the following:
- pipeline nominal outside diameter (discrete variable, nomOD)
- reel radius (discrete variable, reel_radius)
- strong pipe wall thickness (strongWT)
- weak pipe wall thickness (weakWT)
- strong material variation to SMYS (strongMatIncr)
- back tension (tension)
The observed responses are axial strain, peak strain, maximum ovality, and maximum lift-off. These responses allow to assess the local buckling event severity and probability of occurrence. The simulation process is integrated in an automated pSeven workflow to perform a design of experiments and check various configurations of parameters. The results are shown below and analyzed in the next chapter.
Figure 4. Scatter matrix and correlation analysis results for design of experiments sample. Allows for a quick visualization of the design of experiments and assessment of strong correlations.
Reaching a high safety level regarding buckling increases the cost of the pipeline. Hence, designing pipelines at an acceptable cost requires selection of the proper design variables, in order to efficiently effect the safety level behavior. In order to determine which variables have the most effect on the variance of the responses, we perform a sensitivity analysis in pSeven. Among the uncorrelated variables (i.e., variables which can be considered independent between each other), the variations of peak_strain and axial_strain appear to be mostly driven by three variables in the linear trend analysis.
Figure 5. The responsibility for total variance of strain responses is not shared equally among variables.
Other responses are strongly non-linear, hence the simple linear approximation used in the sensitivity analysis tool is not applicable. We need a more complex approximation technique. Moreover, the design methodology requires several expensive FEA simulations to be run each time a new pipeline is designed. The total FEA solving time is on the critical path of a project and can last several hours for each single reel-pipeline configuration. Rather than replicating such a time-consuming process each time a new pipeline is to be designed, an alternative approach is used aiming at replacing the FEA model by an approximation model.
Here we aim to build an approximation for the four responses of the FEA model. To assess the predictive capabilities of the model and ensure that there is no over-fitting, a part of the dataset is extracted before training the model to compute the prediction error in the validation stage. A ratio of 80% is used for the training dataset, representing 380 points.
The technique selected by SmartSelection  is the High Dimensional Approximation (HDA, ) technique. HDA is a non-linear, self-organizing technique for the construction of an approximation model using a linear decomposition in both linear and nonlinear base functions.
Figure 6. Predictive power evaluated on scatter plots (true vs predicted value) for all responses.
An independent test sample is used for this validation phase.
As shown on the plots above, approximation model demonstrates suitable accuracy for all outputs in almost full range of values. However, it diverges from ideal prediction quality for high maximum ovality or high lift-off values. The reasons are twofold. First, those high values represent the extreme limits of the phenomenon we are trying to model, where non-linearities are at their highest and the FEA model might not be reliable. Second, the maximum ovality and lift-off distributions are very skewed towards small values, hence the sparsity of the DoE in such extreme conditions, limiting the generalization capabilities of the model. Overall, the surrogate model has rather good quality, and this process is promising for being used in the design process of pipelines.
- Eric Giry, Lead flowline system, reeling project, Saipem SA
- Julien Ginestet, Lead Engineer, Saipem SA
- Vincent Cocault-Duverger, Flowline & Pipeline System Department Manager, Saipem SA
- Martin Pauthenet, Application Engineer, Datadvance SAS
Saipem is a world leader in engineering and drilling activities and in the development of major project, in the energy and infrastructure sectors.www.saipem.com
 Kyriakides S. and Corona E., Mechanics of Offshore Pipelines, volume 1: buckling and collapse, Elsevier, 2007
 Denniel S. and Tkaczyk T. and Howard B. and Levold E. and Aamlid O., The Development of a Refined Assessment Procedure for the Pipeline Reeling Process Using Reliability and Finite Element Techniques, OMAE2009-79348, Honolulu, United States of America, 2009
 Smith D. and Peters C. and Lahiri S., The application of reliability-based methods in the optimization of reeled rigid pipeline wall thickness requirements, OMAE2018-77865, Madrid, Spain, 2018
 Giry E. and Ginestet J. and Cocault-Duverger V. and Pauthenet M. and Chec L., Machine Learning for Subsea Pipeline Reeling Mechanics, OMAE2020-18685, Fort Lauderdale, United States of America, 2020