A Log Anomaly Detection Method for Power System Operation and Maintenance Scenarios

With the increasing scale and complexity of modern power systems, massive operation and maintenance logs are continuously generated, making manual inspection inefficient. Existing log anomaly detection methods mainly rely on log key sequences or template occurrence patterns, while the semantic information of templates and the contribution of instance-level parameters are often insufficiently utilized. To address this issue, this paper proposes a template-parameter fusion framework for log anomaly detection. Log templates are encoded by a pre-trained language model, and log parameters are represented through a lightweight parameter encoder. A template-semantics-guided dynamic weighting mechanism is further designed to adaptively assign importance to different parameters. The fused event representations are then fed into a BiLSTM-based sequence detector for anomaly prediction. Experiments on HDFS and BGL datasets demonstrate the effectiveness of the proposed method.