Abstract
This paper studies how an edge recurrent neural network can improve a
Battery Monitoring System (BMS). The proposed monitoring measures
current, voltage, and temperature and infers the State of Charge (SoC)
and State of Health (SoH) values through machine learning in an embedded
system. The study relies on two test cases: a theoretical one using
NASA’s battery dataset, where the high volume of data is best suited for
a study, and a second one with a system built in the University of São
Paulo where this paper can analyze more profound the practical results
of its use. This system of the second test case consists of peripheral
sensors integrated into an Internet of Things (IoT) platform, sending
the data collected from a VRLA battery to a Single Board Computer (SBC)
via Bluetooth Low Energy (BLE). The edge SBC concentrates this received
data - from one or more IoT nodes - generating new data for supervision
and enabling control. The SBC communicates with a Web server in a
one-way route to send the battery data without any data request. The
algorithm developed for the SoC uses a dense recursive network with Mean
Absolute Error (MAE) up to 0.2, and for SoC above 10%, the Mean
Absolute Percentage Error (MAPE) can reach 0.16%. As an SoH calculation
method, the battery replacement performs the methodology when the
capacity reaches 80% of its nominal capacity. It is essential to
highlight that these results are from a devices with limited hardware
availability without cloud communication.