Page 180 - Kaleidoscope Academic Conference Proceedings 2021

2021 ITU Kaleidoscope Academic Conference

Culex mosquitoes, usually known as common house Mel–Filterbank Energy (MFE), Neural Network (NN)
mosquitoes, are the primary vectors of West Nile fever, St. classifier, model t raining, t esting, and deployment were all
Louis encephalitis, and Japanese encephalitis, as well as viral performed using Edge Impulse.
illnesses of birds and animals. Culex mosquitoes can also
spread lymphatic filariasis, a parasitic illness, and tularemia, 2.1 Database acquisition
a bacterial disease. With the exception of extreme northern
latitudes, Culex mosquitos are found across the world in In order to train the Convolutional Neural Networks (CNN)
tropical and temperate climates. They are found both indoors model, the database acquired by https://humbug.ox.ac.uk/
and outside and feast on humans and animals at night[8]. was re–sampled to 16 kHz. This database was created to help
people in Sub–Saharan Africa to detect and identify mosquito
In most developing countries, the conventional technique species. The training and testing data comprises three labels:
of identifying mosquitoes or mosquito breeding places Anopheles, Aedes, and Culex, with a total training duration
includes single or several laboratory analyses of larva samples of 9 minutes and 17 seconds and a total testing duration of
collected from various locations by the municipality. As 2 minutes and 14 seconds. The following figures (Figure
a result, the result is delayed sometimes, and manpower is 2a & Figure 2b) illustrates the training and testing database
required. Delay in obtaining results might be harmful to respectively.
the community. The local community’s low involvement
in battling mosquito breeding places may be due to a
lack of situational knowledge about mosquito population
in a location. It should be automated to be successful,
primarily to minimize false reporting and to reduce the
work required to gather mosquito incidence data. However,
(a) Training database
the difficulty in obtaining these mosquitoes may be to
blame for the lack of mapping data. The identification
procedure in developing countries requires manual capture
and identification. Hence, it necessitates people and time
resources. To successfully combat mosquito–borne disease
vectors, thorough and up–to–date data on their mapping
is required. Thus, low–cost surveillance systems that can (b) Testing database
identify mosquito populations are required[9, 10].
Figure 2 – Train and test database
This paper explores the possibility of identifying Anopheles,
Aedes, and Culex mosquitoes using tiny machine learning
2.2 Impulse design and model training
(TinyML) algorithms and audio data collected from
smartphones. The technique for identifying the mosquito
An impulse takes raw data, uses signal processing to extract
from its sound is to compare the smartphone–recorded
features, and then uses a learning block to classify new data.
wing beat sound to a database of mosquito wing beat
There are four blocks in impulse including, Input, MFE,
recordings. The dataset gathered by https://humbug.ox.ac.uk/
Neural Network, and Output. The input block contains the
was re–sampled to 16 kHz to train the Convolutional Neural
time series data in raw form with audio as axes, as shown
Networks (CNN) model. Using a spectrogram, the frequency
in the figure 3. The Window size parameter specifies how
of mosquito wing beats was visualized as a feature. After
long each data window should be in milliseconds. Each raw
evaluation, the model had a precision of 88.3 percent and a
sample is split into several windows. The Window increase
loss of 0.26. This model is then uploaded to the Arduino Nano
parameter sets the offset of each successive window from the
33 BLE Sense, which is connected to an OLED screen. The
first[11].
Arduino Nano 33 BLE Sense combines a small form factor
with a variety of environmental sensors such as an IMU,
microphone, gesture, light, proximity, barometric pressure,
temperature, and humidity. I2C pins are used to connect the
OLED to the Arduino. TinyML and TensorFlow Lite may be
used to run AI on it. When a mosquito is identified from its
wing beat, the screen displays a photograph of it. This whole
system may be integrated into a wearable or portable gadget
that can help detecting mosquitoes in our environment.
2. EDGE IMPULSE
With integrated machine learning, Edge Impulse allows
developers to design the next generation of intelligent
device solutions[11]. Database acquisition, impulse design, Figure 3 – Time series data

– 118 –