This paper studies and compares several machine learning algorithms to characterize and cluster different types of concrete based on their ultrasonic response. We focus on identifying properly fabricated ultra-high-performance concrete (made with extremely fine aggregate and containing thin steel fibers). Ultra-high-performance concrete achieves much higher strength characteristics than traditional mixes and therefore can be used with less material and lower maintenance. Yet, improperly fabricated ultra-high-performance concrete can become a safety concern if characteristics do not match the design specifications. Hence, the goal of our research is to create a nondestructive system to verify the proper construction of this concrete, and in properly fabricated UHPC, determine the quantity and orientation of fibers.  
We compare four methods in the paper: a least mean square adaptive filter, a normalized least mean square adaptive filter, a k-nearest neighbor algorithm, and a multi-layer perceptron based neural network algorithm. The first two algorithms are studied because adaptive filters are commonly used to learn from time-domain signals. When using an adaptive filter, we assume one measurement acts as a reference (similar to a baseline) and another measurement acts as the filter input. If the reference and the filter output belong to the same group, the mean square error will be small. The last two algorithms are used for many traditional machine learning classification tasks. For both methods, we use each ultrasonic measurement’s autocorrelation as an input feature. The autocorrelation expresses the relationship between a measured signal and shifted versions of that signals.  
Two types of high-performance concrete are classified by our algorithms: concrete without steel fiber and concrete with steel fibers. We collect 252 groups of data from 10 cm by 10 cm by 35 cm concrete cuboids, made at the University of Florida. The concrete cuboids are constructed with a metal casing mold. We use indirect transmission in which the transmitter and receiver are located on the same surface. The transmitter sends a 50 kHz center frequency Gaussian pulse with a 200 kHz bandwidth, which is then measured at the receiver after propagating through the concrete. Measurements are taken across the concrete with a fixed distance between the two transducers and with one transducer fixed to one end of the cuboid. Each algorithm was tested using 10-fold cross-validation. We found the normalized least mean square filter failed to distinguish the presence of fiber (achieving an accuracy of 62%) while the k-nearest neighbor algorithm achieves an average 89% accuracy. Overall, the algorithms demonstrate a strong potential to distinguish the concrete mixtures with steel fibers from the ones without.

Acknowledgement  
The authors will like to thank the Florida Department of Transportation (FDOT), for  
providing the funding for this project under the task work order # BDV31 TWO 977-105.


Distinguishing Steel Fiber Content in Ultra-High Performance Concrete (Uhpc) Using Machine Learning Algorithms

No proceedings papers in 2020. Video only.

QNDE2020 - Virtual Conference

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.