Search

Efficiently monitoring the condition of civil infrastructure requires automating the structural condition assessment in visual inspection. This paper proposes an Attention-Enhanced Co-Interactive Fusion Network (AECIF-Net) for automatic structural condition assessment in visual bridge inspection. AECIF-Net can simultaneously parse structural elements and segment surface defects on the elements in inspection images. It integrates two task-specific relearning subnets to extract task-specific features from an overall feature embedding. A co-interactive feature fusion module further captures the spatial correlation and facilitates information sharing between tasks. Experimental results demonstrate that the proposed AECIF-Net outperforms the current state-of-the-art approaches, achieving promising performance with 92.11% mIoU for element segmentation and 87.16% mIoU for corrosion segmentation on the test set of the new benchmark dataset Steel Bridge Condition Inspection Visual (SBCIV). An ablation study verifies the merits of the designs for AECIF-Net, and a case study demonstrates its capability to automate structural condition assessment. • A deep learning model AECIF-Net is developed for automating the visual assessment of structural conditions. • SBCIV, an image dataset with annotations of structural elements and their surface defects, is created. • AECIF-Net outperforms current methods in segmenting structural elements and surface defects. • Merits of the designs for the AECIF-Net are verified by comprehensive experiments. [ABSTRACT FROM AUTHOR]

• A framework for predicting seismic responses based on limited building information is proposed. • A dataset of seismic responses corresponding to 162 typical RC frames of low-medium rise and 200 diverse ground motions is built. • A CNN-based prediction model (call StruNet here) is established for regular RC frame structures based on the dataset. • StruNet maps the limited building information and ground motion time histories into seismic responses (PFA and IDR) of each floor. • StruNet shows sufficient accuracy and high computational efficiency in predicting seismic responses for four actual cases. Building portfolio is the important urban engineering system, and the seismic resilience assessment of a city needs the quick and accurate prediction of the seismic responses of existed buildings. However, many existed buildings generally possess the problem that the design information materials are incomplete or completely lost. The major challenge in the seismic resilience assessment of building portfolio is how to predict the seismic responses of buildings quickly and accurately just using limited building information. This manuscript aims to develop a method for the seismic response prediction of the existed reinforced concrete (RC) frame buildings just using limited building information. A total of 162 typical RC frame buildings of low to medium rise are designed, and the inter-story drift (IDR) as well as peak floor acceleration (PFA) of each floor in each building are computed for 200 ground motions with nonlinear time history analysis (NLTHA) method. A convolutional neural network (CNN) is developed with ground motion records and five easy-getting building parameters as inputs. The outputs are IDR and PFA of each floor for the given building. Considering the physical means of an input parameter—number of stories, the modified loss function and modified evaluation function are proposed. The developed network is trained with the computed dataset and the modified loss function, and the trained model (referred to StruNet) can take the characteristics of ground motions and structures into consideration together comparing to previous studies. The proposed model is verified through four cases (i.e., 4 actual buildings with different construction time, occupancy types, and plane layouts), which are independent of the deep learning dataset. The results confirm that the proposed method offers prediction results with sufficient accuracy and shows high computational efficiency. [ABSTRACT FROM AUTHOR]

• A classification neural networks and a regression neural network were trained on simulated CBED patterns to predict TEM sample thickness from 4D STEM convergent beam diffraction patterns. • Sub-unit cell lateral resolution, single unit cell thickness resolution, and 1 nm deviation from quantitative HAADF STEM thickness estimates were achieved. • Convolution neural networks were trained on simulation data with automatic labelling, and do not require quantitative matching between experiment and simulation. • Similar methodology might be extended to learn other physical properties of a sample from large 4D STEM datasets. Two types of convolutional neural network (CNN) models, a discrete classification network and a continuous regression network, were trained to determine local sample thickness from convergent beam diffraction (CBED) patterns of SrTiO 3 collected in a scanning transmission electron microscope (STEM) at atomic column resolution. Acquisition of atomic resolution CBED patterns for this purpose requires careful balancing of CBED feature size in pixels, acquisition speed, and detector dynamic range. The training datasets were derived from multislice simulations, which must be convolved with incoherent source broadening. Sample thicknesses were also determined using quantitative high-angle annular dark-field (HAADF) STEM images acquired simultaneously. The regression CNN performed well on sample thinner than 35 nm, with 70% of the CNN results within 1 nm of HAADF thickness, and 1.0 nm overall root mean square error between the two measurements. The classification CNN was trained for a thicknesses up to 100 nm and yielded 66% of CNN results within one classification increment of 2 nm of HAADF thickness. Our approach depends on methods from computer vision including transfer learning and image augmentation. [ABSTRACT FROM AUTHOR]

A deep learning-based rapid peak seismic response prediction model for the most common two-story and three-span subway stations is proposed in this study. The established model predicts the peak seismic responses of subway stations with a data-driven fashion and using limited information. The prediction model extracts the features of ground motions using one-dimensional convolutional neural network (1D-CNN) and then integrates the information of subway stations (i.e., the seismic fortification intensity, buried depth, and shear wave velocity) through a fully connected neural network for regression, resulting in peak seismic responses, namely the peak floor acceleration (PFA) and maximum inter-story drift ratio (MIDR). The model is trained using 19,200 samples obtained from the nonlinear time-history analyses (NLTHAs) of the designed 48 typical subway station structures. Furthermore, the external model verification was performed on 960 additional samples. For the predictions of PFA and MIDR , the coefficient of determination (R 2) values are 0.967 and 0.986, respectively, and the damage states of subway stations are further evaluated, achieving an accuracy of 95.0%. These indicates that the model has good predictive performance and generalization ability. Moreover, the prediction model demonstrates a significantly higher computational efficiency compared to numerical simulation methods. ● Deep learning-based peak seismic response prediction model for two-story and three-span subway stations is established. ● Ground motion and structural characteristics can be considered simultaneously. ● The developed prediction model just requires the limited structural information. ● Peak seismic responses of each layer in subway stations can be obtained by the model. [ABSTRACT FROM AUTHOR]