Anomaly-PTG: A Time Series Data-Anomaly-Detection Transformer Framework in Multiple Scenarios

• DOI: 10.3390/electronics11233955
• Publication Date: 2022-11-30T09:32:53Z
• Authors: Gang Li; Zeyu Yang; Honglin Wan; Min Li
• Keywords: Autoencoder; Anomaly (physics); Robustness

In actual scenarios, industrial and cloud computing platforms usually need to monitor equipment traffic anomalies through multivariable time series data. However, the existing anomaly detection methods can not capture long-distance temporal correlations of data potential relationships between features simultaneously, only have high accuracy for specific sequence scenarios without good generalization ability. This paper proposes a time-series anomaly-detection framework multiple Anomaly-PTG (anomaly parallel transformer GRU), given above limitations. The model uses GRU as information extraction module learn correlation timestamps global feature relationship multivariate series, which enhances ability extract hidden from After extracting information, learns sequential representation data, conducts modeling, transmits full connection layer prediction. At same time, it also autoencoder reconstruct two are optimally combined form an model. combines timestamp prediction with reconstruction, improving rate rare accuracy. By using three public datasets physical devices one dataset network intrusion detection, model’s effectiveness was verified, strong robustness were demonstrated. Compared most advanced method, average F1 value on four increased by 2.2%, each over 94%.