The seismic response of a wide variety of structures, from small but irreplaceable museum exhibits to large bridge systems, is characterized by rocking. In addition, rocking motion is increasingly being used as a seismic protective strategy to limit the amount of seismic actions (moments) developed at the base of structures. However, rocking is a highly nonlinear phenomenon governed by non-smooth dynamic phases that make its prediction difficult. This study presents an alternative approach to rocking estimation based on a physics-informed convolutional neural network (PICNN). By training a group of PICNNs using limited datasets obtained from numerical simulations and encoding the known physics into the PICNNs, important predictive benefits are obtained relieving difficulties associated with over-fitting and minimizing the requirement for a large training database. Two models are created depending on the validation of the deep PICNN: the first model assumes that state variables including rotations and angular velocities are available, while the second model is useful when only acceleration measurements are known. The analysis is initiated by implementing K-means clustering. This is followed by a detailed statistical assessment and a comparative analysis of the response-histories of a rocking block. It is observed that the deep PICNN is capable of effectively estimating the seismic rocking response history when the rigid block does not overturn.
]]>Deep reinforcement learning (DRL) is promising for solving control problems in fluid mechanics, but it is a new field with many open questions. Possibilities are numerous and guidelines are rare concerning the choice of algorithms or best formulations for a given problem. Besides, DRL algorithms learn a control policy by collecting samples from an environment, which may be very costly when used with Computational Fluid Dynamics (CFD) solvers. Algorithms must therefore minimize the number of samples required for learning (sample efficiency) and generate a usable policy from each training (reliability). This paper aims to (a) evaluate three existing algorithms (DDPG, TD3, and SAC) on a fluid mechanics problem with respect to reliability and sample efficiency across a range of training configurations, (b) establish a fluid mechanics benchmark of increasing data collection cost, and (c) provide practical guidelines and insights for the fluid dynamics practitioner. The benchmark consists in controlling an airfoil to reach a target. The problem is solved with either a low-cost low-order model or with a high-fidelity CFD approach. The study found that DDPG and TD3 have learning stability issues highly dependent on DRL hyperparameters and reward formulation, requiring therefore significant tuning. In contrast, SAC is shown to be both reliable and sample efficient across a wide range of parameter setups, making it well suited to solve fluid mechanics problems and set up new cases without tremendous effort. In particular, SAC is resistant to small replay buffers, which could be critical if full-flow fields were to be stored.
]]>A significant challenge of structural health monitoring (SHM) is the lack of labeled data collected from damage states. Consequently, the collected data can be incomplete, making it difficult to undertake machine learning tasks, to detect or predict the full range of damage states a structure may experience. Transfer learning is a helpful solution, where data from (source) structures containing damage labels can be used to transfer knowledge to (target) structures, for which damage labels do not exist. Machine learning models are then developed that generalize to the target structure. In practical applications, it is unlikely that the source and the target structures contain the same damage states or experience the same environmental and operational conditions, which can significantly impact the collected data. This is the first study to explore the possibility of transfer learning for damage localisation in SHM when the damage states and the environmental variations in the source and target datasets are disparate. Specifically, using several domain adaptation methods, this article localizes severe damage states at a target structure, using labeled information from minor damage states at a source structure. By minimizing the distance between the marginal and conditional distributions between the source and the target structures, this article successfully localizes damage states of disparate severities, under varying environmental and operational conditions. The effect of partial and universal domain adaptation—where the number of damage states in the source and target datasets differ—is also explored in order to mimic realistic industrial applications of these methods.
]]>Population-based structural health monitoring (PBSHM) systems use data from multiple structures to make inferences of health states. An area of PBSHM that has recently been recognized for potential development is the use of multitask learning (MTL) algorithms that differ from traditional single-task learning. This study presents an application of the MTL approach, Joint Feature Selection with LASSO, to provide automatic feature selection. The algorithm is applied to two structural datasets. The first dataset covers a binary classification between the port and starboard side of an aircraft tailplane, for samples from two aircraft of the same model. The second dataset covers normal and damaged conditions for pre- and postrepair of the same aircraft wing. Both case studies demonstrate that the MTL results are interpretable, highlighting features that relate to structural differences by considering the patterns shared between tasks. This is opposed to single-task learning, which improved accuracy at the cost of interpretability and selected features, which failed to generalize in previously unobserved experiments.
]]>Offshore wind turbines intend to take a rapidly growing share in the electric mix. The design, installation, and exploitation of these industrial assets are regulated by international standards, providing generic guidelines. Constantly, new projects reach unexploited wind resources, pushing back installation limits. Therefore, turbines are increasingly subject to uncertain environmental conditions, making long-term investment decisions riskier (at the design or end-of-life stage). Fortunately, numerical models of wind turbines enable to perform accurate multi-physics simulations of such systems when interacting with their environment. The challenge is then to propagate the input environmental uncertainties through these models and to analyze the distribution of output variables of interest. Since each call of such a numerical model can be costly, the estimation of statistical output quantities of interest (e.g., the mean value, the variance) has to be done with a restricted number of simulations. To do so, the present paper uses the kernel herding method as a sampling technique to perform Bayesian quadrature and estimate the fatigue damage. It is known from the literature that this method guarantees fast and accurate convergence together with providing relevant properties regarding subsampling and parallelization. Here, one numerically strengthens this fact by applying it to a real use case of an offshore wind turbine operating in Teesside, UK. Numerical comparison with crude and quasi-Monte Carlo sampling demonstrates the benefits one can expect from such a method. Finally, a new Python package has been developed and documented to provide quick open access to this uncertainty propagation method.
]]>Physics-informed neural networks (PINNs), which are a recent development and incorporate physics-based knowledge into neural networks (NNs) in the form of constraints (e.g., displacement and force boundary conditions, and governing equations) or loss function, offer promise for generating digital twins of physical systems and processes. Although recent advances in PINNs have begun to address the challenges of structural health monitoring, significant issues remain unresolved, particularly in modeling the governing physics through partial differential equations (PDEs) under temporally variable loading. This paper investigates potential solutions to these challenges. Specifically, the paper will examine the performance of PINNs enforcing boundary conditions and utilizing sensor data from a limited number of locations within it, demonstrated through three case studies. Case Study 1 assumes a constant uniformly distributed load (UDL) and analyzes several setups of PINNs for four distinct simulated measurement cases obtained from a finite element model. In Case Study 2, the UDL is included as an input variable for the NNs. Results from these two case studies show that the modeling of the structure’s boundary conditions enables the PINNs to approximate the behavior of the structure without requiring satisfaction of the PDEs across the whole domain of the plate. In Case Study (3), we explore the efficacy of PINNs in a setting resembling real-world conditions, wherein the simulated measurment data incorporate deviations from idealized boundary conditions and contain measurement noise. Results illustrate that PINNs can effectively capture the overall physics of the system while managing deviations from idealized assumptions and data noise.
]]>