Artificial Intelligence at the Service of Offshore Pipelaying Operations

Abstract During offshore pipelay the pipeline catenary could be subject to unexpected loading conditions that could potentially trigger a local plastic deformation (e.g. ovalization) exceeding the requirement for piggability which should be carefully addressed. The paper summarizes the research activities performed to evaluate how the Artificial Intelligence (AI), particularly the Machine Learning (ML) techniques, can speed-up the prediction of such type of deformations, in alternative or support to the well-established Finite Element (FE) physic driven approach. A typical methodology to calculate the local plastic deformation of the pipeline foresees to prepare a FE structural model of the pipe and to perform a non-linear Finite Element Analysis (FEA) for each specific load scenario which depends on several parameters: pipelay vessel, pipe mechanical data, lay methodology, sea state, water depth. In some cases, several scenarios should be analyzed with a huge impact on manhours and computational time. ML techniques may reduce the time and manhours by building specific models able to predict the pipe deformations after being trained by a certain number of examples computed through FEA. A summary of the steps performed will be presented. The first activity executed is the data exploration i.e., the study of the available input data and output results from previously performed FEA on some specific pipelay scenarios, trying to find possible correlations between the input data and FEA results, investigating how data are distributed in their domain and the most effective way to pre-process the data for the subsequent application of the ML methodology. The second step performed is the selection, design and implementation of the most adequate ML models to predict the pipe plastic deformation. The most promising ML models have been trained and tested with a synthetic dataset of previously computed FEA simulations for a given set of load scenarios. Finally, their performances have been evaluated by comparing the predictions for new un-seen scenarios, computed by means of new FEA simulations. It is deemed that in some specific cases, like the one here described, the engineering effort and computational time can be optimized and/or reduced by ML technique, if adequately tailored for this purpose