Physical Anomaly Detection in EV Charging Stations: Physics-based vs ResNet AE

Abstract

The number of electric car users has grown in recent years, increasing the demand for reliable electric vehicle charging stations (EVCS). The safety of EVCSs is very important as compromised charging stations can disrupt the grid, injure the end-users, and damage the vehicle. In this paper, we will focus on the physical security of EVCS, because physical attacks tend to be more harmful to the end user. Two anomaly detection approaches were presented for detecting physical anomalies: physics-based anomaly detection and deep learning-based anomaly detection (ResNet Autoencoder). The presented approaches were trained and tested using data collected from the EV Charging Station System testbed of the Idaho National Laboratory. Anomaly detection performance was evaluated on three different attack scenarios, targeting various parts of the system including power transfer subsystems and the cooling subsystem of the charger. The presented approaches were compared against two widely used unsupervised anomaly detection algorithms: OCSVM and LOF. Moreover, we evaluated the advantages and limitations of the physics-based vs ResNet Autoencoder approaches for each of the three attack scenarios. The ResNet Autoencoder approach showed the highest performance in terms of accuracy, F1, recall, and precision. Furthermore, this approach demonstrated a number of advantages including automated non-linear feature extraction and unsupervised learning.