![](bg2.png)
Figure 1: Types of foundations (Association et al., 2012).
2 RELATED WORK
The first SHM study dates back 50 years ago. In
70s and 80s, the oil industry faced the problem of
identifying damage in offshore platforms (Martinez-
Luengo et al., 2016). They struggled to develop meth-
ods based on identifying vibrations to locate the dam-
age. At the same time, the aerospace community
began investigating the use of vibration-based strate-
gies. This approach has continued with the current
research of the National Aeronautics and Space Ad-
ministration (Seshadri et al., 2016). Currently, SHM
has been developed in the fields of civil aviation in-
dustry (Khan et al., 2014) and civil structures (Song
et al., 2017). SHM is highly multidisciplinary, and
advances in other areas of study can probably be re-
cruited for SHM’s progress.
Figure 2 shows a general classification for differ-
ent types of strategies for SHM.
Figure 2: Algorithms classification (Martinez-Luengo et al.,
2016).
The methodology implemented in this work is based
on supervised learning algorithms. Specifically on
neural networks (NN), which are recently used in
structural health monitoring to identify, locate, and
quantify damage in different types of structures (Liu
et al., 2017). One of the best-known deep NNs is
the convolutional neural network (CNN). A CNN is
commonly used to recognize objects in images given
their ability to exploit spatial or temporal correlation
in the data (Albawi et al., 2017). A CNN has multi-
ple layers; including fully connected layers, grouping
layers, convolutional and nonlinear layers. Fully con-
nected layers and convolutional layers have parame-
ters, however non-linearity and grouping layers have
no parameters.
Some research has been conducted related to CNN
in the field of SHM. For example, (Tabian et al., 2019)
proposes to collect impact waves using piezoelectric
sensors (PZT) to detect and locate impacts (this ap-
proach was tested on a rigid panel). Another method
with piezoelectric sensors is used in (De Oliveira
et al., 2018), where the signals from the sensors are
transformed to RGB images. A different study fo-
cused on transfer learning (TL) techniques to train
with discrete histogram data (compressed data) is pre-
sented in (Azimi and Pekcan, 2019). Their results in-
dicate that deep TL can be effectively implemented
for SHM of similar structural systems with different
types of sensors. However, these previous works used
known-input vibration signals. In this work, it is pro-
posed to use CNN for damage diagnosis in wind tur-
bines foundations by using only vibration-response
data. The strategy consists on transforming the vi-
bration signals into images (with as many channels as
sensors), and then classify the images with its corre-
sponding structural state label.
3 EXPERIMENTAL SET-UP
The general overview of the experimental testbed is
given in Figure 3 and explained as follows.
The experiment starts with a white noise signal
given by the function generator. This signal is am-
plified and passed to the inertial shaker. This is re-
sponsible for generating vibrations (similar to those
produced by gusts of wind on the blades) to the lab-
oratory tower structure. The shaker is placed at the
upper part of the structure, thus simulating the nacelle
mass. The simulation of different wind speed is also
simulated with this shaker, by changing the amplitude
of the input electrical signal. In particular, multiply-
ing it by the factors 0.5, 1, 2, and 3. Finally, the struc-
ture is monitored by 8 triaxial accelerometers which
are connected to the data acquisition system. Thus,
data from 24 sensors is collected. The nomenclature
used for each sensor is given in Table1.
The real structure used in this work is a tower model.
From Figure 3 (offshore platform) it can be seen the
components of the structure: jacket, tower and na-
celle. As a whole, this structure is 2.7 m high. The
tower is composed of three sections joined with bolts.
ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics
182