Anomaly detection of extremely rare events based on VAE-LSTM-SelfAttention

In certain industrial application scenarios, the number of normal samples significantly outweighs that of abnormal samples. When abnormal samples account for less than 1% of the total dataset, this scenario is termed extreme rare events. In the context of anomaly detection for such events, dataset imbalance often leads to critical challenges, including insufficient feature learning for abnormal samples and poor model generalization. To address these issues, this paper proposes a novel combined model that integrates a Variational Autoencoder (VAE), Long Short-Term Memory network (LSTM), and Attention mechanism for detecting extreme rare events. Experimental results on public benchmark datasets demonstrate that the proposed method outperforms state-of-the-art algorithms in terms of key evaluation metrics, including AUC-ROC, F-score, and the precision-recall (PR) curve, validating its effectiveness in handling extreme class imbalance and improving rare anomaly detection performance.