Time-Series to Image-Transformed Adversarial Autoencoder for Anomaly Detection

Abstract

The automation of systems and the accelerated digital transformations across various industries have rendered the manual monitoring of systems difficult. Therefore, the automatic detection of system anomalies is essential in diverse industries. Various deep learning-based techniques have been developed for anomaly detection in multivariate time-series data with promising performance. However, there are several challenges: 1) difficulty in understanding the relationships among time-series data due to their complexity and high-dimensionality, 2) limitation in distinguishing anomalies from normal data that exhibit similar distributional patterns, and 3) lack of intuitive interpretation of anomaly detection results. To address these issues, we propose a novel approach referred to as the time-series to image-transformed adversarial autoencoder (T2IAE), which adopts image transformation techniques and convolutional neural network (CNN)-based adversarial learning. Image transformation techniques were used to effectively capture the local features of adjacent time points. Two CNN-based adversarial autoencoders competitively learned to distinguish between normal and abnormal data. We experimentally analyzed five real-world multivariate time-series datasets, wherein the proposed model achieved superior anomaly detection performance compared with state-of-the-art methods. Moreover, the proposed model enables humans to intuitively interpret the detection results, facilitating appropriate explanations of the results and enhancing the model's usability.