A multi-scale one-class time series anomaly detection method based on Mamba

Abstract

The disclosure provides a Mamba-based multi-scale single-class time series anomaly detection method. For the time series data input by the sensor input of the industrial water treatment system, firstly, the multi-scale dependent features of the time series are modeled to identify local fluctuations and global time series correlation features; then the global time series correlation features are input into a global context feature extraction module to extract global context features, the local fluctuations are input into a local dynamic feature extraction module to extract local dynamic features, and a multi-scale semantic representation is formed through a feature fusion module; finally, a denoising single-class projection module and a customized loss function are fused and applied, and the compact clustering and robust anomaly discrimination of the normal mode of the industrial water treatment system are realized through the value range judgment of the anomaly score. To solve the problems that the existing method is difficult to balance multi-scale features and long-term correlation, is easy to fall into the optimization error of low-level features, has weak anti-noise and data pollution ability, and has insufficient scene adaptability.

Images