Time-Series Anomaly Detection in Automated Vehicles Using D-CNN-LSTM Autoencoder

It is reasonable to expect Connected and Automated Vehicles (CAVs) to revolutionize the intelligent world owing to the swapping of seamless and real-time data. Although CAVs provide many benefits to the environment and society, new challenges in term of security, privacy, and safety are emerged by anomalies, errors, cyber-security attacks, or malicious activities that led to accidents with fatal victims. This paper tackles the anomaly detection problem by introducing a novel framework in multi-sensor CAVs, which applies an efficient data preprocessing and deep learning method. In this paper, two preprocessing methods have been used and compared in the deep learning based models. The main contributions of this paper compared to previous works include two fundamental tasks. First, the quality of time series data is improved in the preprocessing phase by Differencing (DIFF) or Moving Standard Deviation (MSD). Second, the applied deep learning method is based on an autoencoder where a combination of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) is utilized as the autoencoder architecture, which is referred to as D-CNN-LSTM Autoencoder in this paper. The D-CNN-LSTM Autoencoder method optimizes the anomaly detection rate for all of the anomalies, specifically in the case of low magnitude anomalies, enhancing F1-score up to 18.12% in single types of anomalies and 32.83% in mixed types of anomalies. The experimental results indicate the superiority of the proposed method for time series anomaly detection over the state-of-the-art and benchmark methods.