Abstract

Model reduction for fluid flow simulation continues to be of great interest across a number of scientific and engineering fields. In a previous work [1], we explored the use of Neural Ordinary Differential Equations (NODE) as a non-intrusive method for propagating the latent-space dynamics in reduced order models. Here, we investigate employing deep autoencoders for discovering the reduced basis representation, the dynamics of which are then approximated by NODE. The ability of deep autoencoders to represent the latent-space is compared to the traditional proper orthogonal decomposition (POD) approach, again in conjunction with NODE for capturing the dynamics. Additionally, we compare their behavior with two classical non-intrusive methods based on POD and radial basis function interpolation as well as dynamic mode decomposition. The test problems we consider include incompressible flow around a cylinder as well as a real-world application of shallow water hydrodynamics in an estuarine system. Our findings indicate that deep autoencoders can leverage nonlinear manifold learning to achieve a highly efficient compression of spatial information and define a latentspace that appears to be more suitable for capturing the temporal dynamics through the NODE framework.

Abstract

During each aircraft program a vast amount of aerodynamics data has to be generated to judge performance, structural loads as well as handling qualities. Within the past years the usage of computational fluid dynamics has significantly increased providing accurate insights into aircraft behaviour at early design stages and therefore at least partially enabled the mitigation of costly design changes. However, fully relying on high fidelity aerodynamic data is still computational prohibitive. Hence, data-driven models have gained an increasing attention in recent years. These methods not only provide continuous models but also enable the inclusion of highly accurate aerodynamic results in time-critical environments. This paper aims at applying deep learning techniques to derive such models and compare them to state of the art reduced order modeling techniques. In particular, three deep learning methods, a Multilayer perceptron for distribution predictions, a Multi-layer perceptron for pointwise predictions and an Autoencoder coupled with an interpolation technique are compared to Proper Orthogonal Decomposition and Isomap with latent space interpolation. For all methods an efficient methodology to determine hyperparameters is outlined and applied. Results are presented for an Airbus provided XRF1 dataset which includes surface pressure distributions at various Mach numbers and angles of attack.

Abstract

During their life-cycle, civil infrastructures are continuously prone to significant functionality losses, primarily due to material's degradation and exposure to several natural hazards. Following these concerns, many researchers have attempted to develop reliable monitoring strategies, as integration to visual inspections, to efficiently ensure bridge maintenance and early-stage damage detection. In this framework, recent improvements in sensor technologies and data science have stimulated the use of Machine Learning (ML) algorithms for Structural Health Monitoring (SHM). Among unsupervised learning techniques, the potential of autoencoder networks has been attracting notable interest in the context of anomaly detection. In this light, the present paper proposes two different autoencoder-based damage detection techniques, focused on the Multi-Layer Perceptron (MLP) and the Convolutional Autoencoder (CAE) networks, respectively. During the training, the selected ML models learn how reconstructing raw acceleration sequences acquired from sound conditions. Unknown data, including both healthy and damaged bridge responses, are afterwards used to test the implemented networks and to detect damage occurrence. To this aim, a specific index of reconstruction loss is selected as a damage sensitive feature with the aim to quantify the errors between the original and reconstructed sequences. The performance exhibited by the two approaches is compared and evaluated by application to the Z24 benchmark bridge. Results demonstrate the effectiveness of the proposed methodology to perform feature classification and real time damage detection at the level of macro-sequences as new sensor data is collected, resulting suitable for continuous assessment of full-scale monitored bridges.