EV Battery Health and Fire Predictor IoT System with GSM

Abstract

—Fires in electric vehicle battery packs, whether they begin with thermal runaway, overcharging, or an internal short circuit, continue to erode public confidence in electric mobility. At the same time, most conventional battery management systems respond only after the damage has started: they isolate the pack once a fault is already under way and almost never warn the owner when the vehicle is out of sight. This paper presents the design, construction, and bench evaluation of an EV battery health and fire predictor built around the ESP32. The unit brings together five physical measurements, namely pack voltage, pack current, cell surface temperature, combustible gas concentration, and open flame radiation, and combines them into a compact logistic fire risk score that is computed entirely on the edge node. The prototype operates with a 3S lithium ion pack, an ACS712 Hall effect current sensor, a precision 30 kiloohm and 10 kiloohm resistive divider, a DS18B20 1-Wire probe, an MQ-2 gas sensor, and an infrared flame sensor, and it drives a relay cut-off, an L298N motor stage, a 2000 Hz buzzer, and an SSD1306 OLED display. The owner can supervise the vehicle over three independent channels: Bluetooth Classic, Wi-Fi linked to the Blynk IoT cloud, and a SIM800L GSM modem that delivers SMS alerts. A five-condition sensor fusion rule, whose adaptive term is a sigmoid risk score built from temperature, its rate of rise, and current magnitude, latches the hazard condition and executes a timed safety sequence: an immediate audible alarm, battery disconnection three seconds after the latch, drive train inhibition, and an SMS to the registered owner. Because the inference reduces to a fixed set of floating point operations, the edge decision latency stays well below ten milliseconds per cycle, while the one second sampling period and the three second cut-off delay form a deterministic and auditable safety timeline. Bench tests on a 3S pack reproduced the intended trip behaviour, delivered early warning three to five seconds before any visible flame, and held the false alarm rate below four percent across the load profiles tested. The design is deliberately inexpensive, can be assembled from commodity modules, and transfers directly to e-rickshaws, e-scooters, solar battery banks, and uninterruptible power supplies. The complete firmware, the locked pin map, and every diagram are released alongside this paper so that the experiment can be repeated without ambiguity.