Journal of Computing in Civil Engineering

Although deep convolutional neural networks (DCNNs) have been widely adopted for crack segmentation, they often demonstrate performance degradation on data with real-world complexities. To achieve consistent and accurate prediction performance with complex and feature-rich real-world data, DCNN hyperparameters must be properly selected or optimized. The goal of this study is to provide a novel hyperparameter optimization framework for future crack segmentation DCNN designs to follow, and gain insights into hyperparameter importance on segmentation performance. In this study, a Bayesian optimization framework and an accompanying global sensitivity analysis have been proposed to guide the search for optimal crack segmentation DCNNs using real-world 3D roadway range images. The proposed Bayesian optimization framework can determine the optimal configurations for both training- and architecture-related hyperparameters. In addition, the probabilistic models developed during Bayesian optimization are leveraged by the accompanying global sensitivity analysis to interpret and rank the hyperparameter importance on DCNNs’ segmentation accuracy.

Abstract

The prevalence of run-off-road crashes, particularly in rural areas, underscores the significance of roadside characteristics in safety analysis. This paper proposes a novel approach for automated roadside safety assessment using deep convolutional neural networks (CNNs) and Generative Adversarial Networks (GANs) for data augmentation. The CNN models evaluate roadside features through two-dimensional (2D) image analysis, whereas GANs expand the data set by generating additional diverse samples. The proposed framework aligns with the standard rating system of the Federal Highway Administration (FHWA) and encompasses four distinct models for guardrail detection, clear zone width assessment, rigid obstacle detection, and sideslope estimation. The performance of each model is compared against non-GAN augmented models to assess the efficacy of using GANs for data augmentation. The results show that the proposed approach outperforms existing methods in terms of accuracy, which is measured with 96% in detecting guardrails, 88% in detecting clear zones, 80% in detecting rigid obstacles, and 84% in detecting roadside slopes. Compared with manual approaches, the proposed method offers advantages such as cost-effectiveness, ease of implementation, and the ability to rapidly rank state roads. The developed framework can assist departments of transportation (DOTs) in efficiently identifying problematic road segments and prioritizing safety improvement projects based on FHWA standard rating system.

Practical Applications

This research focuses on the development of computer vision models for roadside safety assessment. The models successfully detect and classify important features such as guardrails, clear zones, rigid obstacles, and roadside slopes in images. The practical applications of this research are significant for transportation authorities, engineers, and practitioners involved in roadway safety. By utilizing these computer vision models, they can efficiently analyze large amounts of visual data and identify potential safety hazards along roadways. These models can aid in identifying areas with inadequate clear zones, which are crucial for preventing roadside crashes. They can also detect the presence of guardrails, which play a vital role in redirecting vehicles and minimizing the severity of accidents. Moreover, the models provide insights into the presence and distance of rigid obstacles, such as trees, rocks, and mountains, which are essential for assessing the overall safety of roadside environments. Last, the models assess roadside slopes, allowing practitioners to identify areas with steep inclines that may pose a higher risk to motorists. Overall, the practical applications of these computer vision models enable stakeholders to prioritize safety interventions, implement targeted improvements, and enhance roadway safety for both drivers and pedestrians.

The energy efficiency optimization of Florida’s residential buildings is essential for the reduction of fossil fuel consumption and greenhouse gas emissions. This optimization requires designers to accurately calculate the building energy loads at the design stage to efficiently design the cooling and heating systems. However, due to time and cost constraints, designers usually explore only a few design alternatives and simulate their energy use. This study developed an artificial-neural-network (ANN)-based generative design (GD) framework to automate the design process of detached residences while optimizing their energy performance. The ANN model was developed using the machine learning platform TensorFlow and some Python-based Keras libraries based on a big data set of about 17,000 newly constructed detached residences in Florida between the years 2009 and 2021. The GD framework was established using Autodesk Dynamo and the Autodesk Revit GD add-on that uses the nondominated sorting genetic algorithm (NSGA-II), in which multiple design parameters mainly relating to the house geometry and its energy performance were incorporated. Considering 10 independent variables, including the total wall area, roof area, floor area, and the windows’ U-value, the ANN model predicted the required capacities of the cooling and heating systems in detached houses, with $R^2$ values of 0.955 and 0.904, respectively. The ANN then was integrated within the GD framework for performance evaluation purposes. The findings of this study resulted in a fully automated 3-min design process of an energy-efficient detached house envelope, maximizing the productivity of designers and developers and greatly reducing the financial strain and time consumed using traditional techniques.

Maintaining the quality of indoor environments in educational facilities is crucial for student comfort, health, well-being, and students’ learning performance. Current indoor environment maintenance practices and building systems for educational facility spaces often fail to include feedback from students and exhibit limited adaptability to their needs. To address this problem, this paper introduces a novel artificial intelligence of things (AIoT)-based framework to predict multidimensional indoor environment quality (IEQ) conditions. The proposed framework integrates internet of things (IoT) systems with deep learning algorithms to systematically incorporate multidimensional IEQ data and multimodal feedback from occupants. By collecting, fusing, and analyzing real-time IEQ and occupant feedback data, the proposed framework predicts the future IEQ condition based on current conditions. This framework yields insights into the IEQ conditions and their potential impacts on student well-being, thereby facilitating the future development of climate-adaptive, data-driven, and human-centric educational facilities. This framework was deployed, validated, and tested in selected educational facilities at the Virginia Tech Blacksburg campus, with encouraging results.

Abstract

Structural health monitoring (SHM) is critical in identifying the degradation of infrastructure systems to ensure structural integrity and safety. Vibration-based SHM approaches, including numerical-physics-based modeling and data-driven strategies, are commonly used to detect damage. This study proposes a method for predicting damage conditions using a hybrid vibration-based approach with convolutional neural networks (CNNs) trained with physics-based data sets. The method is evaluated using a five-story reinforced concrete building that undergoes multiple base excitations, resulting in cumulative damage that affects the building’s stiffness and dynamic responses. A set of damage states is defined based on the structure’s response, and simplified models of the building are used to create a training database for the CNNs. The CNNs are trained on noise-free dynamic responses (i.e., accelerations or displacements) from numerically simulated white noise (WN) sequences and then tested with the appropriate floor response data from different types of base shaking. The accuracy of the models is consistently high, with noise-free acceleration and displacement responses yielding results of 99.9% and 93.9% for numerically simulated WN base excitations, respectively. The accuracy remained high when tested with 30 dB signal-to-noise ratio (SNR) noisy acceleration and displacement responses, with accuracies of 99.9% and 93.8%, respectively, and 100% when using acceleration responses from experimentally measured WN base excitations with a similar SNR. Ambient microtremor acceleration data collected within California’s Central Valley were used to validate the approach for low-amplitude ambient ground vibrations, achieving an accuracy of 86.69% when tested with noisy acceleration responses with the measured microtremors as base shaking. The proposed method has limitations in identifying bordering damage states and reduced accuracy when tested on field data, but overall shows promise for damage state identification and story stiffness reduction analysis.

Practical Applications

This study presents a new method for detecting damage scenarios in buildings using vibration data and machine learning (ML). We used CNNs, a learning algorithm, to analyze vibrations from a five-story building with varying damage levels. The CNN models were trained with computer simulations and real-world measurements to recognize various damage states. The proposed method proved to be highly accurate and efficient in detecting and associating damage with reductions in floor stiffness, even with noisy data, achieving up to 99.9% accuracy. The study also found that detailed computer models are not necessary for generating training data. A simplified model resembling a real structure is effective, making the method more practical and computationally efficient. This allows the technique to be applied in real-world situations, where limited measurements of the structure’s response can be input into the CNN model to assess if the building has experienced damage over time. In summary, this research shows that CNN models trained with data from simple numerical models can effectively identify damage scenarios in a building using real-world measured vibrations from the building, even during minor shaking.

This study introduces a framework for transplanting a building information modeling (BIM) library. Design detailing constitutes 50%–60% of the total design time, even within the BIM context. Previous studies have highlighted the potential of integrating BIM and artificial intelligence (AI) for enhanced productivity. However, challenges arise due to architects’ preferences for unique project-specific details when applying generalized AI approaches based on big data. To address this, we propose a BIM library transplant framework. This framework automatically identifies objects at a high level of development (LOD) from a selected existing BIM model (i.e., a donor model) and matches them with low-LOD objects in a new model (i.e., a recipient model). Subsequently, it replaces the low-LOD objects with corresponding high-LOD objects. The framework involves three steps: (1) extracting the library from the donor model, (2) matching the library, and (3) transplanting the library from the donor to recipient model. To validate its efficacy, we implemented the BIM library transplant framework as a Revit add-on, employing the random forest classifier as the object-matching AI model. Our results indicate that the implemented framework has the potential to reduce detailing time by approximately 60%–70%, while achieving an accuracy of 65%–80%.

Even with the utilization of software tools like Navisworks to automate clash detection, clash resolution in construction projects remains a slow and manual process. The reason is the meticulous nature of the process where coordinators need to ensure that resolving one clash does not lead to new clashes. The use of machine learning to automate clash resolution as a potential option to improve the clash resolution process has been suggested with research showing positive results to support the implementation. While the research shows high accuracy in predicting clash resolution options to support automation, the scope limits the discussion on the complex and often lengthy process of developing a machine-learning model. Based on this research gap, the authors in this paper discuss the development of a prediction model to identify clash resolution options for given clashes. The discussion is focused on individual steps involved in creating machine-learning models like data collection, data preprocessing, and machine-learning algorithm development and selection. The authors also address common challenges in the development of machine-learning models including class imbalance and availability of limited data. The authors utilize a multilabel synthetic oversampling method to generate different percentages of synthetic data to account for class imbalance and limited data sets. Using this data set, the authors trained five machine-learning algorithms and reported on their accuracy. The authors concluded that increasing the data set with 20% synthetic data, and using an artificial neural network to develop the machine-learning model to automate the resolution of clashes have generated better results with an average accuracy of around 80%.