International Concrete Abstracts Portal

TUBEDECK, a one-way spanning voided composite slab, has been utilized in the construction field over the years to enhance the efficiency, constructability, and environmental performance of structures. TUBEDECK incorporates both cast-in-situ reinforced concrete slabs and profiled steel decks. However, there is a need to clarify the shear resistance capacity in this slab because the shear strength of the member reduces as concrete volume is eliminated to optimize flexural strength. Therefore, this study applied the artificial neural network (ANN) technique to determine the shear strength of TUBEDECK. By varying factors in the ANN features, several ANN models were developed. Out of many models developed, an optimal model was selected, having a maximum/mean relative errors of 5.1% in a dataset.

With IoT sensors gaining widespread adoption in recent years for monitoring in-situ concrete properties, the volume of data generated using these sensors is growing at a significant rate. These sensors are typically used for several purposes among which temperature, humidity and strength monitoring (using the maturity method) are currently the most common. This data is typically collected at centralized cloud-based databases where they can be accessed by sensor end-users as well as algorithm developers. This work presents on how data from these IoT sensors has been used by the authors to train machine learning algorithms to perform several tasks including but not limited to: detecting anomalies, detecting events in the service life of the sensor (e.g. concrete pouring,) suggesting mixture alterations to optimize performance and predicting future performance. These capabilities are currently being used by concrete practitioners on daily basis. This is done using data collected from tens of thousands of sensors, used in over 7500 projects representing geographical regions of over 45 countries and representing several thousand unique concrete mixtures. This, to the authors’ knowledge, is the largest dataset available for training such algorithms. Two use cases are presented for how this data is utilized to train machine learning algorithms to assist practitioners in day-to-day activities such as mixture optimization.

This paper presents and explains an implementation of artificial intelligence for the real-time crack detection of ultra-high-performance concrete (UHPC). A deep learning algorithm is employed to process image data and to identify physical cracks. The state-of-the-art object detection method generates accurate results with small datasets. To provide training and validation images, UHPC specimens are cast with various fibers and loaded per an ASTM standard, including steel and synthetic (collated and monofilament polypropylene) fibers. After testing, sample images are labeled with an annotation tool and the algorithm is trained and validated with an image recognition approach, leading to a mean average precision (mAP) of 99%. The occurrence of cracking and propagation are linked with the applied load level to appraise the influence of the mixed fibers in the crack development of UHPC. It needs to be noted that the adopted deep learning architecture is incapable of quantifying crack width and area directly; therefore, a Java-based image processing program is used to measure these properties of the specimens. The characteristics of the load-induced cracks are dominated by the fiber types. Plain UHPC fails rapidly and the flexural capacity of UHPC increases with the presence of the fibers; especially, the UHPC with steel fibers demonstrates higher flexural capacities than other cases.

In this paper, the artificial neural network (ANN) method is utilized to predict ground-penetrating radar reflection amplitudes from four different inputs, namely, temperature, ambient relative humidity, chloride level, and corrosion condition on the surface of the reinforcing bar. A total of 288 ground penetrating radar (GPR) data points were collected from a series of chloride-contaminated concrete slabs under various environmental profiles that were used to train, validate, and test the proposed ANN model. The ANN model performed well in predicting the GPR reflection signals, with the overall coefficient of determination (R2) being 0.9958. The overall mean squared error (MSE) and root mean squared error (RSME) values are 0.015 and 0.122, respectively. These values are very low, which means that the ANN model has an excellent prediction capability. The research results show that the GPR reflection amplitudes are more sensitive to the temperature changes and chloride level parameters than the ambient relative humidity and rust condition on the reinforcing bar surface. Using the ANN method to predict the GPR reflection amplitudes is relatively new for structural concrete applications. This study paves the way for further developments of neural networks in civil and structural engineering.

The structural deterioration of aging infrastructure systems is becoming an increasingly important issue worldwide. To compound the problem, economic strains limit the resources available for the repair or replacement of such systems. Over the past several decades, structural health monitoring (SHM) has proven to be a cost-effective method for the detection and evaluation of damage in structures. Visual inspection and condition rating is one of the most commonly applied SHM techniques, but the effectiveness of SHM varies depending on the availability and experience of qualified personnel and largely qualitative damage evaluations. Simply supported three-dimensional reinforced concrete T-beams with varying geometric, material, and cracking properties were modeled using Abaqus finite element (FE) analysis software. Up to five cracks were considered in each beam, and the ratios of stiffness between cracked and healthy beams with the same geometric and material parameters were measured at nine equidistant nodes along the beam. A feedforward ANN utilizing backpropagation learning algorithms was then trained on the FE model database with beam properties and nodal stiffness ratios serving as inputs for the neural network model. The outputs consisted of the predicted parameters of location, depth, and width of up to five cracks. This inverse problem is very difficult or impossible to solve with the training done by the Artificial Neural Network. One ANN was trained to predict the parameters of the cracks using the full database of FE simulations. The damage prediction ANN achieved fair prediction accuracies, with coefficients of determination (R2) equal to 0.42. This result was the outcome of the no uniqueness in the prediction of this inverse analysis. Nevertheless, this ANN model provides a rough estimate of the cracking type and damage content in bridge girders once the nodal stiffness ratios are measured by applying a field vehicle loading and measuring the deflection using a theodolite. A touch-enabled user interface was developed to allow the ANN model to predict the crack configurations. The application was given the acronym DRY BEAM, for Damage Recognition Yielding Bridge Evaluation After Monitoring.