ALGAN: Time Series Anomaly Detection with Adjusted-LSTM GAN

Abstract Anomaly detection in time series data-identifying points that deviate from expected patterns-is a common challenge across various domains, including manufacturing, medical imaging, and cybersecurity. Recently, Generative Adversarial Networks (GANs) have shown great promise in improving anomaly detection performance. The dual architecture of GANs, comprising a Generator and a Discriminator, enables more accurate modelling of normal behaviour, thereby enhancing the detection of anomalies. In this paper, we present a novel GAN-based model, adjusted-LSTM GAN (ALGAN), designed to improve anomaly detection in both univariate and multivariate time series under an unsupervised setting. ALGAN integrates an adjusted-LSTM network to enhance the quality of generated outputs, leading to more effective anomaly identification. We evaluated ALGAN on 75 real-world datasets from diverse domains and benchmarked its performance against traditional approaches, deep neural networks, and state-of-the-art GAN-based models. ALGAN outperformed these models and achieved an average F $$_1$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mmultiscripts> <mml:mrow/> <mml:mn>1</mml:mn> <mml:mrow/> </mml:mmultiscripts> </mml:math> score of 0.771 on 46 univariate time series from the Numenta Anomaly Benchmark (NAB), 0.851 on 28 multivariate datasets from the Server Machine Dataset (SMD), and 0.677 on a large-scale multivariate dataset from the Secure Water Treatment (SWaT) system. These results demonstrate ALGAN’s effectiveness and underscore the value of the adjusted-LSTM architecture in enhancing the quality of generated outputs and improving unsupervised time series anomaly detection.