Instrumentation & Measurement Magazine 24-5 - 22

or nondestructive testing. The p-emf sensor used in this work
was fabricated in our labs from a commercially available
semiconductor GaAs crystal. The sensor was found to fit completely
the technical requirements of the application proposed
in this work, with pros and cons summarized in Table 1.
Liquid Level Measurement Using
Optical Interferometry with a P-emf
Sensor and Neural Networks: A Caseof-Study
The
optical arrangement proposed as a proof of concept to
measure the tube vibrations is depicted in Fig. 2. The beam of
a low power He-Ne laser is divided in two beams by a 50:50
beam splitter. One of the beams, the signal beam, was pointed
to a small mirror glued to the tube and the back reflection was
made to impinge onto the p-emf sensor, where it interferes
with the reference beam. The path length of both beams was
equalized by reflecting the reference beam on a prism. The
p-emf sensor was fabricated from a commercially available
semiconductor GaAs crystal with a dimension of 3×3×0.5 mm.
Two parallel stripe electrodes were deposited on the front surface
in such a way that the inter-electrode area was 3×1.6 mm.
Fig. 3 shows a picture and schematic representation of the fabricated
p-emf sensor.
In the described experiment, the container vibrations induce
phase modulations on the reflected signal beam, and
the superposition of the signal and reference beams form an
interference pattern that vibrates in accordance to the tube vibrations.
In response to this vibrating interference pattern,
the p-emf sensor generates a temporal electrical current that
resembles the tube vibrating pattern. The arrangement constitutes
a simple coherent optical range finder which provides
an output photocurrent linearly proportional to the displacements
of the target.
Neural Networks-based Non-linear
Function Approximator
Applied Artificial Intelligence (AI) in instrumentation and
measurement is emerging as a promising technique in a variety
of fields, with a growing interest in Deep Learning (DL)
approaches. DL techniques include recurrent neural networks,
convolutional neural networks, deep belief networks, generative
adversarial networks, and autoencoders, allowing to do
tasks such as classification, clustering, prediction, decision
making, estimation, and others, without the need to first build
an analytical model of the problem [28]. Most of these methods
rely on a training process designed to deliver a model that best
matches a set of input data to the desired output. That model
is further used to obtain output values for new incoming data.
The use of neural networks in content level measurement
techniques has been explored in some recent approaches on
soft sensors with good results, mainly due to their nonlinear
function approximation properties, and the mentioned ability
to learn by examples. The work in [29] presents a system in
which the frequency response of slosh waves inside an automotive
tank are fed into a neural network aiming to determine
the content level using a capacitive sensor on a running vehicle
at various tank levels with good results. The work in [30] presents
the design of an adaptive calibration technique using an
optimized artificial neural network for liquid-level measurement
in support of linearity improvement, with the NN used
in approximation function mode. Similarly, a nonlinear system
identification approach for liquid level modeling in an industrial
coke furnace using fuzzy neural networks is reported in
[31]. With these antecedents, we propose in this work the use of
a backpropagation trained, multilayer perceptron neural network,
which carries on a nonlinear mapping of the numerical
information obtained from the natural frequencies (NF) as a
function of the content level through data fitting.
Fig. 4 shows an example of the behavior of the normalized
Fig. 2. Experimental optic setup to detect the vibrations induced in a glass
tube container using a p-emf sensor.
NF corresponding to the first four vibration modes as a function
of the liquid level. The curves are normalized with respect
to foi,
which represents the frequency of the ith
NF obtained
Fig. 3. P-emf sensor and its schematic representation. S and R: signal and
reference beams. Lx
×Ly
22
is the interelectrode area.
when the container is empty. It can be noticed that the first NF
shows a tendency to decrease linearly in levels above 50%, so
theoretically it could be used directly to determine the filling
level in this range; however, the situation is different in the remaining
range, so the use of information obtained from the rest
of vibrational modes becomes pertinent.
Accordingly, in this paper we propose the use of a neural
network architecture used in the form of a non-linear function
approximator. The NN consists of a multilayer perceptron
with a single hidden layer, trained using a back propagation
Levenberg Marquardt algorithm within a data fitting scheme.
The activation functions are sigmoids in the hidden layer and
IEEE Instrumentation & Measurement Magazine
August 2021
Instrumentation & Measurement Magazine 24-5

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-5
