An Online Anomaly Detection Algorithm with Adaptive Concept Drift
Mengyuan Li · 2024
In response to the existing offline anomaly detection algorithms performing well on static data, but struggle to meet the requirements of accuracy and real-time processing when the data distribution changes. This paper proposes an online anomaly detection algorithm with concept drift adaptation to adapt to continuously changing data distributions. Firstly, the data is subjected to offline training using a Long Short-Term Memory autoencoder to obtain anomaly scores. And unsupervised anomaly detection is performed based on dynamic thresholds. Then, an adaptive sliding window is employed for concept drift detection, with the model being updated and adjusted through concept drift adaptation. Finally, the improved Particle Swarm Optimization algorithm is utilized to optimize the hyperparameters that contribute more to the accuracy. Results indicate that the proposed algorithm effectively adapts to evolving data distributions, meeting the real-time processing and accuracy requirements essential for anomaly detection.