The increasing scale of large-scale scientific facilities and scientific data center network systems makes the system vulnerable to internal failures or external attacks during operation, leading to system paralysis and significantly impacting the normal work and life of users. Timely and accurate fault location is crucial for ensuring stable system operation. Log data, which records the system's running status, plays a vital role in identifying and rectifying anomalies that occur during system operation. Therefore, timely troubleshooting of log anomalies is essential for maintaining the stability of complex systems and ensuring the safe operation of large scientific facilities and scientific data centers.
This paper proposes a log anomaly detection method based on local information extraction in Transformer model, which learns the deep language features and context information of log recording through Sentence-BERT model. The local feature extraction of convolution operation and the Transformer model are used to capture context information in the sequence to improve the model's ability to recognize complex patterns and speed up model training and reasoning. We have carried out experiments on log data sets, and the experimental results show that this method can provide a reliable and efficient solution for log anomaly detection of large scientific research facilities and scientific data center network systems.