Time Series Anomaly Detection using Prediction-Reconstruction Mixture Errors

Abstract

Anomaly detection on time series data is increasingly common across various industrial domains that require monitoring metrics to prevent potential accidents and economic losses. The complications of anomaly detection revolve around a scarcity of labeled data and the need to learn temporal correlations between multiple variables. Most successful unsupervised methods either use single-timestamp prediction or reconstruct entire time series. However, these methods are not mutually exclusive and can each offer complementary perspectives. This work first explores the successes and limitations of prediction-based and reconstruction-based methods. Next, it compares the effect of attention-based architectures with LSTM-based architectures on existing models. Finally, this research proposes a novel autoencoder architecture capable of producing bi-directional predictions while simultaneously reconstructing the original time series by optimizing a joint objective function. An ablation study using a mixture of prediction and reconstruction errors demonstrates that this simple architecture outperforms other state-of-the-art models for anomaly detection on both univariate and multivariate time series.