One-step Predictive Encoder - Gaussian Segment Model for Time Series Anomaly Detection

Unsupervised anomaly detection for time series is of great importance for various applications, such as Web monitoring, medical monitoring, and device fault diagnosis. Time series anomaly detection (TSAD) aims to find the observations that most different from others in a sequence of observations. With the development of deep learning, deep-autoencoder-based methods achieve state-of-the-art performance. These methods are usually able to find single anomaly points but fail to detect the anomaly segment and the change point. To tackle this problem, this paper proposes a novel TSAD method, which consists of a bidirectional LSTM (BiLSTM) autoencoder and a subsequent Gaussian segmentation model. BiLSTM encodes a time series in a predictive format from both positive and negative time directions, then outputs the latent feature vectors and restructured errors. After that, the latent features are used to find anomaly segments by the Gaussian segment model; the restructured errors are used to find change points and extreme single anomaly by a scoring function. In this way, our method can find all three kinds of anomaly points. Experiments on two real-world datasets demonstrate the effectiveness of the proposed method.