LUAD: a lightweight unsupervised anomaly detection scheme for multivariate time series data

Anomaly detection of multivariate time series data has drawn extensive research attention recently, as it can be widely applied into various different domains, such as Prognostics Health Management, community behaviour monitoring, financial Anti-fraud and so on. Anomalies typically refer to unexpected observations or sequences within the captured data. The prevailing solutions of current anomaly detection methods are not only highly related to the individual use, but also rely on the domain-specific prior knowledge. Existing methods of anomaly detection by detecting aberrations encounter fundamental engineering challenges in terms of steam data online nature and the lack of expert knowledge for the training data set. Also, to meet the practical requirements, the anomaly detection model is often required to be used in edge architectures where the computing resources are limited, which leads to the demand for developing light-weight anomaly detection methods. To address these challenges, we propose a lightweight, unsupervised anomaly detection scheme, called LUAD. LUAD is consists of a detection model and a diagnosis model. The detection model learns the normal patterns of input data via an encoder–decoder scheme that combines Temporal Convolutional Network (TCN) and Variational Auto-Encoder (VAE) to deconstruct and reconstruct multivariate time series data. The diagnosis model improves LUAD's overall detection accuracy and provides a reasonable explanation for an anomaly. Experiments on three very different public datasets indicate that LUAD is both highly generalizable and more accurate than the two current state-of-the-arts. Overall, the LUAD model outperforms the baselines both in effectiveness (0.71%∼1.45% higher) and efficiency (31X smaller in model size, 1.9X faster in training time).