Microservices-based systems are complex and dynamic. To manage such systems, different tools are installed to monitor and test its performance, which produces a significant volume of monitoring data in real time. How to model these multi-source data and mine the relationships among these data is a challenge for building microservice anomaly detection systems. This paper presents MADM, a novel framework for anomaly detection that leverages Graph Neural Networks. MADM models time series data from various sources, incorporating predictive techniques to enhance detection accuracy. By efficiently pre-processing and utilizing data from multi-sources, including logs, metrics, and traces, MADM establishes a unified graph representation for all data. This representation can describe complex service interactions and topological information of microservice systems. Experiments conducted on two widely used microservice benchmark datasets demonstrate that MADM surpasses existing methods in the anomaly detection task.