Dual‐Encoder GAN With Dynamic Thresholding for Time Series Anomaly Detection

ABSTRACT Time series anomaly detection (TSAD) is utilized for identifying abnormal patterns from normal patterns in various fields. Defining semantic anomaly patterns in a specific sequence is a problem for existing methods, resulting in loss of diversity and incorrect detection problems. In order to address these problems, this study recommends using a dual‐encoder generative adversarial network. The proposed model employs a generator and two encoders to comprehend the intricate dependencies and subtle alterations of data, guaranteeing accurate generation. An adaptive threshold is used with anomaly scores that combine reconstruction and discrimination errors to clearly identify anomalies when detecting anomalies. Experimental results in NASA, NAB and UCR benchmark datasets show that the proposed model outperforms ARIMA, LSTM‐AE and TadGAN in terms of F1 score by 0.135, 0.222 and 0.05, respectively.