A reliability anomaly detection method based on enhanced GRU-Autoencoder for Vehicular Fog Computing services

With the rapid development of the Internet of Vehicles (IoV), Vehicular Fog Computing (VFC) reduces communication latency by pushing computational resources from the cloud to the fog layer close to the vehicle. However, in dynamic IoV systems, VFC services face the challenge of abnormal reliability, which are the phenomenon of decreased reliability caused by various factors. Most of the existing IoV anomaly detection methods are designed to protect data security, thus ignoring service reliability anomalies in IoV. Therefore, in order to ensure the stability of the VFC service during runtime, in this paper, we propose an enhanced Gated Recurrent Unit (GRU)-Autoencoder method to detect VFC service reliability anomalies. Our method uses GRU-Autoencoder as the anomaly detection model and uses the Bayesian Optimization with the Tree Parzen Estimator (BO-TPE) algorithm to select the optimal threshold coefficient. The method identifies reliability anomalies in the VFC service by estimating the reconstruction loss of the input samples. To address the challenge of time series data anomaly detection methods that rely heavily on representative normal reliability data for training, we propose an algorithm called Multi-stage Scored Reservoir Sampling (MSRS), which enhances the model by automatically selecting representative normal reliability data as the training set. Moreover, we elaborately designed the checkpoint restart algorithm to ensure that the VFC service rolls back to the most recent checkpoint state when reliability anomalies occur. We extensively evaluated the proposed method on five simulated datasets and validated its effectiveness. Our method brings a novel, automated and efficient reliability anomaly detection solution to VFC services.