DDConv: Disentangled Dual Branch Convolutional Network for Anomaly Detection in Server Machines

Detecting anomalies in server machines is critical for providing reliable Web services. Usually, working status of server machines are characterized by multivariate time series (MTS). However, complex patterns in MTS, including both temporal and intervariable ones, make it hard to precisely differentiate anomalies from normal sequences. In this article, we propose a model called disentangled dual branch convolutionary networks (DDConv), a reconstruction-based fully convolutional network to detect anomalies in server machine MTS. The core of DDConv is to disentangle pattern learning in time and variable dimensions and separate the intradimension and interdimension feature learning processes for thoroughly exploring all these kinds of patterns. Patterns in time and variable dimensions are captured independently in parallel via a dual branch convolutional network working under the macro-group convolution mechanism. A novel regularization is further proposed to enable precise reconstructions by shrinking the error space formed by reconstruction errors. Extensive experiments on public benchmarks demonstrate DDConv achieves superior anomaly detection performance over state-of-the-art baselines while maintaining model efficiency.