Energy-Efficient Anomaly Detection and Chaoticity in Electric Vehicle Driving Behavior

Area of Science:

• Electrical Engineering
• Data Science
• Automotive Engineering

Background:

• Anomaly detection in mobile systems can predict risks and enhance energy efficiency.
• Existing research often focuses on cybersecurity or visual data for anomaly detection.

Purpose of the Study:

• To develop unsupervised hybrid anomaly detection approaches for battery electric vehicles (BEVs).
• To evaluate the energy recovery potential and chaotic characteristics of driving data.
• To establish an energy efficiency-based anomaly detection strategy independent of additional sensors.

Main Methods:

• Utilized Long Short-Term Memory (LSTM)-Autoencoder, Local Outlier Factor (LOF), and Mahalanobis distance for hybrid anomaly detection.
• Assessed model performance using silhouette score, Davies-Bouldin index, and Calinski-Harabasz index.
• Analyzed driving datasets for chaotic aspects using Lyapunov exponent, Kolmogorov-Sinai entropy, and fractal dimension.

Main Results:

• Hybrid models outperformed individual sub-methods in anomaly detection.
• Hybrid Model-2 achieved 2.92% higher success in anomaly detection than Hybrid Model-1.
• Hybrid models demonstrated significant potential energy savings (31.26% for Hybrid Model-1, 31.48% for Hybrid Model-2).
• A strong correlation was observed between system anomalies and chaoticity.

Conclusions:

• The developed hybrid models offer superior anomaly detection and energy efficiency for BEVs.
• The study highlights the link between anomaly detection and chaotic analysis in driving data.
• This approach provides an energy-saving anomaly detection strategy using existing sensor data, differentiating from cybersecurity-focused methods.