In the domain of Industrial Internet of Things (IIoT), anomaly detection in time-series is very important. The systems of IIoT are composed of a large number of connected sensor devices, and produce data as multivariate time-series data containing operation status. Such data is of great significance to identify anomalies and ensure the safe and stable operation of the system. When dealing with multivariate time-series, traditional models for anomaly detection have not fully considered the multi-scale time characteristics, medium and long-term dependence patterns, and the complex correlation among characteristics. All those lead to inevitable false positives and bring challenges for anomaly detection in the IIoT domain. To solve these problems, we propose a multivariate time-series anomaly detection framework, TopoGAT, to capture the potential dependencies with integral topology analysis. Extensive experiments are carried out on four public datasets and show that our work has better performance than baselines and provides strong technical support for anomaly detection in the IIoT domain.