<p>In recent years, with the increasing scale and complexity of microservice systems, timely and accurate detection of system anomalies has become crucial for the stability and reliability of these systems. However, existing approaches that rely on single-modal data, such as logs or metrics, fail to capture complex and diverse anomaly patterns. While recent studies have explored multi-modal anomaly detection, their failure to sufficiently address spatial correlations in temporal metrics, local anomaly dilution in long log sequences, and effective complementation between heterogeneous modalities may result in a high false positive rate. To tackle these challenges, we propose a novel semi-supervised anomaly detection framework, named KANAD, which integrates fine-grained modeling of multi-modal structural and multi-modal fusion module. Specifically, for metrics, we design a K-Nearest Neighbors-based graph to model spatial dependencies via graph neural networks. Then, for logs, we propose a point-wise mutual information word-level graph construction approach to capture fine-grained structural semantic or co-occurrence relations among log words, effectively mitigating local anomaly loss. Finally, we introduce multi-modal fusion module aligns heterogeneous representations in a unified semantic space and feeds them into Kolmogorov-Arnold networks for nonlinear fitting to enhance the model’s generalization ability in identifying anomalous data. Experimental results on two public and one real-world datasets show that KANAD generally outperforms baseline methods, achieving higher F1 scores and demonstrating more effective performance in anomaly detection. The source code of KANAD is publicly available at <a href="https://github.com/wuyangwu/KANAD">https://github.com/wuyangwu/KANAD</a>.</p>