Evaluation of thermal damages of concrete subjected to high temperatures using recurrent neural networks for ultrasonic pulse waves

This study aims to propose the use of deep learning methods to classify the thermal damage of concrete cylinder specimens based on ultrasonic pulse wave data. Data are collected from laboratory experiments using concrete specimens with three different water-to-binder ratios (0.54, 0.46, and 0.35). The specimens are subjected to different target temperatures (100 °C, 200 °C, 300 °C, 400 °C and 600 °C) and another set of cylinders are subjected to just the room temperature to represent the normal temperature condition (20 °C). Thermal damages are classified using deep learning classification models based on three recurrent neural networks (long short-term memory, gated recurrent unit, and bidirectional long short-term memory) of the ultrasonic pulse wave data. It is demonstrated that deep learning yields a high accuracy in classifying thermal damage of concrete with a value of 88.24% based on the combination of instantaneous frequency and spectral entropy of ultrasonic times series with a sampling frequency of 1 MHz, as an input of the gated recurrent unit. The performance of the deep learning classification model is superior to models based on two conventional ultrasonic parameters (ultrasonic pulse velocity and signal coherence).