Log anomaly detection based on parallel fusion of CNN and GRU
Li Cheng, Jiaqing Mo, Gang Zhou, Guangming Li · 2024
With the development of deep learning and big data technology, system log anomaly detection technology has received widespread attention. Although many methods have achieved good results in anomaly detection in logs, it remains an extremely difficult task mining the rich dependencies hidden within log data from both global and local perspectives. In this paper, we propose a multi-scale parallel fusion network based on GRU and CNN, which more comprehensively captures the overall dependencies of log data. Specifically, we employ BERT to semantize the preprocessed log sequences, subsequently feeding the semantic vectors into the neural networks of a dual-branch CNN and a variant GRU. This approach enables the capture of both local and global features of the log sequences. Finally, we parallelly fuse the outputs of the dual-branch CNN and variant GRU. We conduct extensive experiments and evaluate three well-known benchmark datasets, BGL, Thunderbird, and Spirit. The results show that our method is better than other methods.