A hybrid deep autoencoder-random forest model for internal short-circuit fault detection in lithium-ion batteries of electric two-wheelers

With the increasing adoption of electric two-wheelers, there is a critical need for reliable and real-time fault detection systems to ensure the safety and longevity of lithium-ion batteries. This paper proposes a hybrid anomaly detection framework that integrates a Deep Reconstruction Autoencoder with a Random Forest (RF) classifier to identify internal short-circuit faults in lithium-ion battery packs. The system is modeled on a modular battery configuration inspired by a commercial electric scooter and is simulated under realistic driving and thermal conditions using MATLAB/Simulink. The autoencoder learns normal battery behavior and generates reconstruction errors, which, combined with raw sensor data, are used by the Random Forest classifier to detect anomalies. The proposed hybrid model is contrasted with the conventional methods of anomaly detection, such as Isolation Forest, One-Class Support Vector Machine (SVM), Local Outlier Factor (LOF) (density-based) and Logistic Regression (LR). The framework covers a range of fault severities. Reliable early fault detection ( R fault = 2.0 Ω , temperature deviation < 5 K above baseline ) and moderate faults ( R fault = 0.5 Ω , temperature deviation 5 -- 15 K ) constitute the primary diagnostic challenges, as electro-thermal indicators are weak and can be easily masked by normal operating variations. A severe short-circuit condition ( R fault = 0.005 Ω ) is also included as a boundary validation case to ensure that the framework remains robust across the full spectrum of fault severity levels. Findings indicate a detection rate of 99%, a precision rate of 98%, and a high level of robustness in both low and high levels of faults. These results emphasize the fact that the model can be implemented in real time in the battery management systems of electric two-wheelers.