Fault diagnosis and root cause localization in large-scale server systems are critical challenges in the field of IT operations and maintenance. Timely and accurate localization of root causes is essential for ensuring system stability. However, existing root cause localization approaches primarily focus on the service node level, lacking fine-grained analysis at the indicator level. This paper addresses the granularity limitations of traditional device-level root cause localization techniques and proposes a fine-grained solution for fault diagnosis in intelligent operation systems. Specifically, we construct an indicator-level topology graph to capture deep dependencies among system monitoring indicators, design a multi-head embedding module to enhance node feature representation, and adopt a Graph Convolutional Network (GCN) to model nonlinear dependencies among indicators, achieving end-to-end indicator-level root cause localization. By mapping the device-level topology to corresponding indicator nodes, we build a fine-grained indicator graph for fault localization. The introduction of multi-head embedding further enhances the expressiveness of indicator node features. Moreover, a GCN-based model is trained on historical fault data to capture spatial features and fault propagation patterns within the indicator graph, significantly improving root cause identification accuracy. Experimental results on public datasets show that the proposed method significantly outperforms existing device-level root cause localization approaches in terms of accuracy.