Search

Machine learning has affected the way in which many phenomena for various domains are modelled, one of these domains being that of structural dynamics. However, because machine-learning algorithms are problem-specific, they often fail to perform efficiently in cases of data scarcity. To deal with such issues, combination of physics-based approaches and machine learning algorithms have been developed. Although such methods are effective, they also require the analyser's understanding of the underlying physics of the problem. The current work is aimed at motivating the use of models which learn such relationships from a population of phenomena, whose underlying physics are similar. The development of such models is motivated by the way that physics-based models, and more specifically finite element models, work. Such models are considered transferable, explainable and trustworthy, attributes which are not trivially imposed or achieved for machine-learning models. For this reason, machine-learning approaches are less trusted by industry and often considered more difficult to form validated models. To achieve such data-driven models, a population-based scheme is followed here and two different machine-learning algorithms from the meta-learning domain are used. The two algorithms are the model-agnostic meta-learning (MAML) algorithm and the conditional neural processes (CNP) model. The two approaches have been developed to perform within a population of tasks and, herein, they are tested on a simulated dataset of a population of structures, with data available from a small subset of the population. Such situations are considered to be similar to having data available from existing structures or structures in a laboratory environment or even from a model and needing to model a new structure with only a few available data samples. The algorithms seem to perform as intended and outperform a traditional machine-learning algorithm at approximating the quantities of interest. Moreover, they exhibit behaviour similar to traditional machine learning algorithms (e.g. neural networks or Gaussian processes), concerning their performance as a function of the available structures in the training population, i.e. the more training structures, the better and more robustly the algorithms learn the underlying relationships. [ABSTRACT FROM AUTHOR]

Digital twins seek to replicate a physical structure in a digital domain. For a digital twin to have close correspondence to its physical twin, data are required. However, it is not always possible, or cost-effective, to collect a complete set of data for a structure in all configurations of interest. It is nonetheless useful to repurpose data to help validate predictions for different configurations and scenarios. This statement is true in drilling applications, where, for example, the length of the drill string is altered throughout operation. This paper demonstrates how transfer learning, in the form of three domain-adaptation methods, — transfer component analysis (TCA), maximum independence domain adaptation (MIDA) and geodesic flow kernel (GFK) — can be used to construct a digital twin for localising torsional friction in deviated wells under structural changes (e.g., when the drill column gets longer). The method uses a physics-based torsional model to train a machine-learning classifier that can localise torsional friction for a given drill string length and diameter, where friction localisation labels are known (source). As the length or diameter of the drill string are altered in the field, transfer learning is utilised to map the classifier from the labelled (source) scenario onto these unlabelled (target) scenarios. As a result, transfer learning improves the performance of the classifier when applied to the target data, and increases the domain of validity for the classifier. The performance of the classifier, and therefore its suitability to new drill-string configurations, is estimated by utilising two different distance metrics between the source and a proposed target dataset. [ABSTRACT FROM AUTHOR]

Linear modal analysis is a useful and effective tool for the design and analysis of structures. However, a comprehensive basis for nonlinear modal analysis remains to be developed. In the current work, a machine learning scheme is proposed with a view to performing nonlinear modal analysis. The scheme is focussed on defining a one-to-one mapping from a latent 'modal' space to the natural coordinate space, whilst also imposing orthogonality of the mode shapes. The mapping is achieved via the use of the recently-developed cycle-consistent generative adversarial network (cycle-GAN) and an assembly of neural networks targeted on maintaining the desired orthogonality. The method is tested on simulated data from structures with cubic nonlinearities and different numbers of degrees of freedom, and also on data from an experimental three-degree-of-freedom set-up with a column-bumper nonlinearity. The results reveal the method's efficiency in separating the 'modes'. The method also provides a nonlinear superposition function, which in most cases has very good accuracy. • Cycle-GANs are used to define mappings between modal and physical coordinates. • An assembly of neural networks is used to enforce orthogonality of the mode shapes. • Statistical independence of modal coordinates is imposed by predefining the modal space. • Orthogonality in the frequency domain is used to pick the most efficient decomposition. • The effectiveness of the method is demonstrated on simulated and experimental systems. [ABSTRACT FROM AUTHOR]