Dynamic Global-Local Adversarial Training-Based Temporal Anomaly Detection Model

Timely and accurate anomaly detection (AD) in time series data is essential for ensuring the reliability and efficiency of industrial systems. However, the high-dimensional and dynamic characteristics of industrial data pose significant challenges for existing methods, particularly in capturing global-local features and complex temporal dependencies. To address these challenges, this paper introduces the Dynamic Adversarial Global-Local Variational Autoencoder-Long Short-Term Memory Framework (DAGLVAE). The proposed framework integrates Long Short-Term Memory (LSTM) and Variational Autoencoder (VAE) in a multi-layer architecture, embedding an attention mechanism in the encoder to dynamically extract critical temporal features. By incorporating adversarial training and dynamic weight adjustment, DAGLVAE achieves a synergistic optimization of global and local feature modeling, enhancing AD accuracy and robustness. Experimental results on multiple industrial datasets demonstrate that the proposed method outperforms existing baseline methods, achieving superior precision and reliability.