Microservice architecture has advantages such as independent development, independent deployment, scalability, and reusability. However, due to its complexity and large scale, faults are inevitable during the operation of microservice system. Although the multi-source data generated by the system provides an overall description of the system, existing research methods are aimed at single data sources and ignore the correlation information between multi-source data. In this paper, a novel anomaly detection approach is introduced, incorporating a multi-level attention mechanism and a multi-scale convolutional neural network. This approach effectively captures the patterns and dependencies within both logs data and KPI data, enabling accurate detection of anomalies. For this purpose, A feature extraction module is designed for log features with multiple influencing factors, and another feature extraction module for KPI features with a multi-scale attention network. Then these features are input in parallel to the attention network for weighted fusion and anomaly detection. We deployed an open-source benchmark microservice system TrainTicket and injected various typical faults to validate our approach. Experimental results show that the approach can identify anomalies more accurately than the existing methods.