MADMM: microservice system anomaly detection via multi-modal data and multi-feature extraction

Accurately detecting anomalies in microservice systems is crucial to avoid system failures and economic losses for users. Existing approaches detect anomalies by extracting sequential information from single-modal data, such as metrics or logs. However, they do not clearly analyze the temporal and spatial dependencies of each data type and ignore the correlation between different types of data, which can lead to a significant number of false positives. In this paper, we propose a novel Microservice system Anomaly Detection method via Multi-modal data and Multi-feature extraction (MADMM), which performs a joint analysis of metrics and logs from a spatial and temporal perspective. Specifically, we firstly construct separate feature graphs for metrics and logs in each time window. A graph convolution network is then utilized to capture the spatial correlation among different metrics, while a graph attention network is employed to analyze the contextual relationships among different log events. Then, we design a Cross-Modal Attention-based Gate Recurrent Unit (CMA-GRU) to capture the intricate temporal dependencies of each modal data and facilitate fulfilling cross-modal interactions. Finally, we introduce multi-grained contrastive learning methods to learn robust cross-modal features from both inter- and intra-modality aspects. Experimental results on real-world datasets demonstrate that MADMM outperforms existing baseline methods and exhibits better robustness.