Experimental Investigation For Improvisation Of Soc & Soh Of Battery By Using Ai Embedded Bms In EV

The research work describes a unified battery management technique that involves a digital twin-based method that encompasses practically validated theory, data-oriented state evaluation, embedded management, and Internet of Things supervision. The task employs two connected phases of experimentation. The first phase consists in the analysis of a six-cell lithium-ion battery pack 3S2P 18650 in twelve operation modes, which consists of using both 0.5C, 1C, and 2C discharges under ambient temperatures of 25°C, 35°C, and 45°C with dynamic loading. The data obtained in the analysis of the voltage, current, temperature, capacity, and energy indicators will serve as calibration data for a second-order Thevenin equivalency circuit digital twin in MATLAB. Phase II benchmarks Random Forest, XGBoost, and LSTM models on 9,742 records from the NASA lithium-ion battery dataset and integrates the selected Random Forest model with an ESP32-based BMS prototype. The embedded platform combines dual INA219 current sensors, a 2-channel relay module, local LCD monitoring, MQTT communication, a Flask-SocketIO dashboard, and a staged voltage-protection strategy. The digital twin reproduces the measured voltage response with RMSE below 0.08 V and R² above 0.98 across the reported constant-current validation cases. In the AI benchmark, Random Forest reaches R² = 0.976, compared with 0.951 for XGBoost and 0.637 for LSTM. Feature-importance analysis identifies current and cycle count as the leading predictors of energy-loss behaviour. Live prototype measurements of 9.87 V, 39.4% SOC, and 100% SOH demonstrate end-to-end sensing, communication, decision, and protection operation. Together, the results establish a practical path from physically validated battery modelling to AI-assisted embedded BMS control.