Research on Adaptive Model Pooling Method for Data Stream Anomaly Detection Based on Variational Auto-Encoder

Abnormal detection is vital for fault warning in industrial production and safety in daily life. However, the challenge of constantly evolving data streams, influenced by external factors and performance degradation, presents a significant hurdle in this regard. Unfortunately, existing anomaly detection methods are unable to cope with these issues effectively. The models are unable to adapt to data streams with constantly changing distributions, which results in poor anomaly detection performance. To address this problem, this paper proposes VAE-ARCUS, an adaptive model pooling online abnormal detection method based on variational autoencoders. This is achieved by modifying the original autoencoder model, which is implemented within the ARCUS framework, into a variational autoencoder. This enables the model to achieve greater accuracy and to grasp the characteristics of the data distribution more effectively, thus alleviating the issues associated with the overfitting of the base autoencoder model to noisy data. Finally, experimental results conducted on six high-dimensional and concept drift datasets show that the model outperforms other autoencoder-based models and common stream anomaly detection methods in detecting anomalies and generalizing the model's capabilities.