Physical Invariant Subspace Based Unsupervised Anomaly Detection for Internet of Vehicles
Man Zhou, Lansheng Han, Jingpei Wang · IEEE Transactions on Intelligent Vehicles · 2024
Due to the dynamic and uncertain nature of the autonomous driving environment, traditional detection methods are ineffective in detecting complex and diverse anomalies. Additionally, existing detection methods relying on deep neural network may lack interpretability, making it challenging to infer the accuracy of the algorithms or the reasonableness of anomalies. To address these issues, this paper proposes an unsupervised anomaly detection method based on the physical invariant subspace. This method utilizes a Kalman Variational Autoencoder (KVAE) to learn the normal trajectory of vehicles and incorporates a subspace extraction (SSE) layer in the encoder to identify and capture discriminative subspaces relevant to the anomaly detection task in the latent space representation. The entire anomaly detection framework consists of a data preprocessing module, a KVAE with a SSE layer (SKVAE), an iteratively reweighted least squares state estimator, and an adaptive non-parametric cumulative sum anomaly detector. Firstly, the SKVAE separates spatial representation from action recognition and employs a Gaussian state space model to capture latent dynamic information in the sequential data, filtering out outliers and noise to enhance the accuracy and robustness of anomaly detection. Moreover, the adaptive non-parametric cumulative sum detector automatically adjusts the detection threshold to adapt to different distributions and variations in the anomaly data. This method demonstrates perfect detection accuracy and efficiency in numerical experiments, and the proposed detection model offers interpretability of the predicted results through the physical invariant characteristics and the extracted interpretable subspaces.