Learning Unified Representations Across System Telemetry Modalities for Automated Incident Detection

Modern distributed computing systems generate vast amounts of heterogeneous telemetry data across multiple modalities, including performance metrics, log files, network traces, and system calls. The effective detection of operational incidents requires integrating these diverse data streams into a unified representation that captures cross-modal dependencies and temporal patterns. Traditional unimodal approaches suffer from limited contextual awareness and fail to leverage complementary information across modalities. This paper proposes a novel framework for learning unified representations across system telemetry modalities using deep multimodal fusion architectures. We employ attention-based fusion mechanisms combined with Long ShortTerm Memory (LSTM) networks to capture both temporal dependencies within individual modalities and semantic correlations across different telemetry streams. Our approach incorporates a hierarchical encoder architecture that progressively refines cross-modal representations through multiple fusion layers. Experimental evaluation on real-world cloud infrastructure datasets demonstrates that our unifiedrepresentation learning framework achieves significant improvements in incident detection accuracy compared to traditional unimodal and simple concatenation-based fusion approaches. The proposed method exhibits robust performance across various incident types and systemconfigurations, with particular effectiveness in detecting subtle anomalies that manifest across multiple telemetry modalities. This work contributes to the growing body of research on intelligent system monitoring by demonstrating the substantial benefits of learning unified multimodal representations for automated incident detection in complex distributed systems.