CANet: Convolutional Attention-based Network for Multivariate Time Series Anomaly Detection

Anomaly detection of multivariate time series has been an active research topic for decades owing to the broad application of time series in various fields including e-commerce, smart city, robotic systems, etc. Accurate detection is therefore of great importance to the operational stability of a system or corporation. In this paper, we propose a new model composed of convolutional neural networks and a modified attention mechanism. Snapshot pictures of multivariate time series are adopted as the input for convolutional process. Convolutional blocks are responsible for spatial information extraction. And the modified attention mechanism can well assign weights to input segments since this mechanism emphasizes important sequences while depressing irrelevant segments. The comparison work with multiple baselines is carried out on datasets collected from the public archive and real world. Experimental results of multiple evaluation metrics prove that our model not only obtains higher accuracy but also gains a better recall rate.