Page 19 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

Van Vy and Hyungchul Yoon 3

inhomogeneous materials. This can result in low accuracy due to noise signals, non-
straight signal propagation, and variations in signal velocity.
In recent, deep learning techniques have begun to supplant conventional
approaches. A prominent example of such a technique is the use of convolutional
neural networks (CNNs) [8] that are adept at resolving the issue of crack detection.
CNNs are often utilized to scrutinize visual imagery or spatial data, allowing them to
learn the intrinsic characteristics of images, such as pixel positions, and identify
crucial features autonomously without human intervention. Due to their use of
specialized convolutional and pooling operations, and parameter sharing, CNNs are
also computationally efficient.
Nowadays, several researchers have demonstrated the applicability of CNNs to lo-
cate cracks through the analysis of AE signals. For instance, Barbosh et al. [9]
employed a 2D CNN and time-frequency decomposition methods to extract crack
information from AE data. Similarly, Sikdar et al. [10] suggested the use of CNN-
based techniques for automatically classifying crack locations using time-frequency
images. Nonetheless, these techniques relied on a classification network, which could
only approximate the location among candidate areas, without pinpointing the precise
coordinates of the location.
Perfetto et al. [11] proposed a regression network to address the previously
mentioned constraints. Nonetheless, this approach relied solely on data obtained from
a single sensor, leading to a single-input single-output (SISO) model. As a result, the
SISO model could not fully harness the relationship between signals obtained from
various sensor locations. Furthermore, this technique was tested solely through
simulations conducted on an aluminum plate, a homogeneous material, using the
Abaqus software. In practical scenarios, identifying cracks in heterogeneous materials
such as concrete is a more intricate task, owing to non-linear signal propagation and
variations in signal velocities within the material.
The structure of the paper is explained in the following sections. In section 2, read-
ers will find a comprehensive introduction to the proposed approach as well as an in-
depth explanation of the process. In section 3, the test and the results are shown. Fi-
nally, section 4 describes the conclusion and future works.

2 The Proposed Method

In this paper, a novel approach for crack detection using deep learning and AE
sensors is proposed. The method aims to automatically estimate the 3D coordinates of
the crack source by detecting the signal of sound waves. The process involves
converting signals collected from AE sensors into the time-frequency domain through
the continuous wavelet transform (CWT) [12]. Then, a convolutional neural network
(CNN) named AECWT-3DR-Net is designed to detect the crack location using CWT
images as input. Finally, the trained AECWT-3DR-Net is utilized to estimate the
crack source coordinates. The overview of the proposed method is shown in Figure 1.
The proposed method's effectiveness is demonstrated through experiments conducted
on a concrete block, which is discussed in detail in Section 3.

ISBN: 978-604-80-8083-9 CITA 2023