Incorporating prior knowledge for domain generalization traffic flow anomaly detection

Traffic flow anomaly detection is crucial for traffic management and reducing adverse impacts. However, due to the lack of labeled information for anomaly events and the highly complex nature of road networks, practitioners find it difficult to detect anomaly information. Currently, many anomaly detection methods perform well when there is no statistical difference between training and testing data (same road), ignoring the unique traffic patterns of different road and the dependency on prior knowledge, which are difficult to dynamically adapt to the detection model. In order to effectively integrate prior knowledge and generalize this knowledge to different road for anomaly detection, we use domain-specific prior knowledge as the initial weights of convolutional kernels, fine-tuning the convolutional kernel weights using classification loss, allowing prior knowledge to adapt to extracting anomaly features. To further increase the generalization capability of features, we perturb the extracted features with Gaussian noise specific to the direction of the abnormal class during training, enabling the anomaly detection classifier to have domain generalization performance across roads with different traffic patterns. We compare the proposed model with six baseline methods on real and synthetic datasets. The results demonstrate that the proposed model can effectively detect anomalies and outperforms the baseline methods.