Research on a Temporal Semantic-Enhanced Transformer-Based Anomaly Awareness Method for Complex Enterprise Data Pipelines

Abstract

Complex enterprise data pipelines are influenced by multiple factors during operation, including changes in data sources, task scheduling, field structures, resource status, and upstream/downstream dependencies. Anomalies often exhibit temporal continuity, semantic relevance, and link propagation. To address the challenge of traditional anomaly detection methods simultaneously characterizing changes in operational state and business semantic constraints, this paper proposes a temporal semantic enhancement Transformer anomaly perception method for complex enterprise data pipelines. This method first jointly encodes multi-source operational observations and node semantic attributes, mapping indicator fluctuations, task status, data quality information, and semantic metadata to a unified representation space to enhance the completeness of pipeline node state representation. Subsequently, temporal position encoding introduces the sequential information of the operational window, enabling the model to perceive the changes in pipeline state over continuous time. Furthermore, the Transformer structure is used to model long-distance temporal dependencies, and a dependency attention mechanism is combined to characterize the relationships between different pipeline nodes, thereby improving the model's ability to identify cross-node anomaly propagation and contextual anomaly patterns. Based on this, a semantic gating enhancement module adaptively adjusts temporal features according to node semantic attributes, ensuring that anomaly judgment considers both numerical changes and business implications. Comparative experimental results show that the proposed method achieves better performance in metrics such as Acc, Pre, Rec, and F1, indicating that the constructed temporal semantic joint modeling framework can effectively improve the accuracy and stability of anomaly detection in complex enterprise data pipelines.