Page 20 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)
P. 20
4
Fig. 1. The overview of the proposed method
The initial step in the proposed approach is to convert the time-history data
obtained from the AE sensor network into the time-frequency domain using the CWT.
The CWT is a mathematical technique used for decomposing the signal into a series
of wavelets that are localized in both frequency and time. It can be used to describe
the behavior of the signal at different scales. This is a widely used technique in
acoustics processing and pattern recognition, as it decomposes the information from
the mother wavelet into basic forms. The CWT is particularly useful for analyzing
signals that contain non-stationary or time-varying features, such as speech signals,
music signals, and biomedical signals. The CWT of a function x(t) at a scale (a>0)
+*
a R and shifting value b R is expressed by the following integral [13, 14]
, (1)
where is a continuous function in both the time domain and the frequency
domain called the mother wavelet and the represents the operation of the
complex conjugate.
Using the CWT in the pre-processing stage before feeding it into a neural network
offers several advantages such as noise reduction or feature extraction. The process
was illustrated as shown in Figure 2.
Fig. 2. Converting signal to time-frequency domain using CWT
The next step involves designing a deep learning model called AECWT-3DR-Net,
which is a multi-input multi-output (MIMO) architecture used to determine the crack's
coordinates. The AECWT-3DR-Net is a Convolutional Neural Network that
comprises eight input branches, each corresponding to the signal obtained from
different sensors, and three outputs that correspond to the x, y, and z coordinates of
the crack source. The AECWT-3DR-Net architecture, illustrated in Figure 3,
comprises twelve convolutional layers, nine pooling layers, and other layers.
Convolutional layers are designed to extract features efficiently and automatically
from raw data inputs, such as images or sound waves. Convolutional layers can be
implemented efficiently using matrix operations and parallel processing, making them
CITA 2023 ISBN: 978-604-80-8083-9