WKN-OC: A New Deep Learning Method for Anomaly Detection in Intelligent Vehicles

Connected and automated vehicles (CAVs) play a vital role in transforming human mobility, tackling road congestion and road safety. However, CAVs rely heavily on the security, accuracy, and stability of sensor readings and network data. Abnormal sensor readings caused by malicious cyberattacks or faulty car sensors can have devastating consequences, possibly even lead to a fatal crash. In order to avoid the CAVs data anomalies caused by network attacks or data failures, we propose a Wavelet Kernel Network with Omni-Scale Convolutional (WKN-OC) for anomaly detection in intelligent transportation systems (ITS), which can select the optimal scale adaptively and processes high-frequency signals better. The proposed method pays more attention to the high-frequency components of input data, and fully extracts valuable features through the feature extraction framework, so that the model has strong anomaly detection performance. We verify the reliability of the WKN-OC method on the Safe Pilot Model Deployment (SPMD) data set. It is shown that the proposed WKN-OC model has good detection performance for various anomaly data, especially in the mixed anomaly experiment. We have achieved 96.78% average accuracy in mixed anomaly experiments and 96.13% accuracy in multi anomaly experiments. The results show that the model has strong generalization performance for the anomaly detection problem faced by the Internet of Vehicles (IoVs) and can identify the unknown anomalies in reality well.

Automated summary

Connected and automated vehicles (CAVs) are crucial in transforming human mobility, addressing road congestion and enhancing road safety. However, their reliability heavily relies on the security, accuracy, and stability of sensor readings and network data. To tackle the issue of anomaly detection in intelligent transportation systems (ITS), we propose a Wavelet Kernel Network with Omni-Scale Convolutional (WKN-OC) model that adapts scales optimally and processes high-frequency signals more effectively. The model demonstrates strong generalization performance and achieved 96.78% average accuracy in mixed anomaly experiments and 96.13% accuracy in multi anomaly experiments on the Safe Pilot Model Deployment (SPMD) dataset.