Extreme response analysis of flexible offshore facilities under current loads is an engineering challenge mainly for deep and ultra-deep waters. In this study, an improved method for defining characteristic current profiles for application in extreme response analysis is presented. It utilises unsupervised dimensionality reduction algorithms including Principle Component Analysis (PCA) and AutoEncoders (AE) followed by application of clustering through K-Means Algorithm (KMA) for two different deep-water locations to identify the current profile corresponding to extreme response of a typical ROV umbilical line. Additionally, a combination of dimensionality reduction and clustering method, known as Embedded Clustering (EC), is also explored alongside various pre-processing techniques. The unsupervised methods presented are demonstrated as an effective approach to scaling the clustering approach to a higher-class resolution.
Knowledge and understanding of ocean current profiles are essential for the design and operation of offshore structures and devices. Umbilicals between surface vessels and Remote Operated Vehicles (ROVs) are examples of flexible structural components exposed to hydrodynamic loads across the vertical expanse of the water column. The effects of ocean current on the umbilicals are becoming ever more prominent in deeper water. As a result, tools that can be used to effectively analyse this phenomenon is becoming increasingly necessary.
In ROV operation, hydrodynamic loads on the umbilical greatly impact the ROV-umbilical motion which results in a time-varying tension load in the umbilical under different current profiles. For design and operation of ROVs, it is important to identify the maximum tension in the umbilical as the extreme response to these current profiles. This is usually carried out through time-domain analysis using the current time history datasets. This is a time-consuming approach with excessive computational efforts.
Within offshore engineering industries, methods such as Current Profile Characterisation (CPC) are used to simplify current profile data. CPC is the process of reducing current time series datasets to smaller characteristic subset while retaining the key information needed for analysis (Prevosto, et al., 2011). This is a useful approach for reducing the computation time required in applications such as extreme response and fatigue analysis. CPC has been implemented in previous studies by clustering vertical current profiles using unsupervised learning (Prevosto et al., 2011; Jeans et al., 2015). Jeans et al. (2015) identified the K-Means Algorithm (KMA) as an effective form of CPC at low profile numbers. However, these studies applied clustering methods to the data in its original feature space only.