BGformer: An Abnormal Event Detection Model For Distributed Software Operation And Maintenance Logs

To address log time series uncertainty and high redundancy challenges associated with distributed software intelligent operation and maintenance, a model based on Transformer and Bi-GRU is proposed to improve the efficiency and accuracy of anomaly detection in distributed highconcurrency execution environments. Transformer multi-head attention captures the correlation between different execution contexts in a high-concurrency environment, thereby extracting global time dependencies;For distributed log data exhibiting long-term and short-term dependencies, the Bi-GRU bidirectional mechanism is used to capture the forward and reverse dependencies of the log sequence, and deeply understand the local time characteristics of the log data;Finally, multiple attention mechanisms are used to process repeated log sequences. It not only improves the detection accuracy of the model when processing large-scale log data, but also reduces its computational complexity. According to the experimental results, the model BGformer proposed in this paper shows good Precision, Recall, and F1-Score on the Spark dataset published by LogHub. Precision reaches 0.910, which is significantly better than LogGPT(0.487) and about the same as TLAN(0.912). While improving time efficiency and addressing the complexity of distributed log data, BGformer is more practical and robust. MultiLog and LogPal have slightly higher F1 scores, but BGformer shows higher scalability and practical deployment value in high concurrency environments. The work can be applied to fault detection and automated operation and maintenance of distributed software systems.