Log Anomaly Detection Method Based on Transformer and Temporal Convolutional Networks

As record files generated during system operation, logs play a vital role in ensuring system stability. However, the rapid development of distributed technologies has led to the increasing scale and complexity of log data, resulting in existing methods facing difficulties in fully mining log features and information loss during parsing. To overcome these difficulties, this study proposes TranConvAD, a global and local feature extraction method for log anomaly detection. The method extracts multidimensional features of logs from three perspectives: time, risk level, and message content, and designs FastText and BERT based multi-type word embedding methods for generating log representation vectors, aiming to minimize noise introduction and information loss. In addition, TranConvAD combines for the first time the global dependency capture of Transformer with the local temporal modeling of temporal convolutional networks. Furthermore, a separable self-attention mechanism and pointwise convolution are introduced to optimize the model. Experimental results demonstrate that TranConvAD achieves a maximum precision of 99.8% on the HDFS dataset, outperforming LogAnomaly by 3.8%, and 98.6% on the BGL dataset, surpassing LightLog by 3.4%, thereby providing a more efficient and scalable solution.