SeLog: A Log Anomaly Detection Method on Log Event and Variable Semantic

Detecting log anomalies based on semantics is an important approach. However, the existing methods often ignore the semantic information of variables. Variables in logs contain important indicative information, which is of positive significance for anomaly detection. Therefore, this paper proposes a log anomaly detection method based on events and fusion variable semantics, referred to as SeLog. SeLog is based on the method of keyword table and positive antonym table, combined with expert experience, to extract valuable variable information from the log. We use these valuable log variables and events as data sources, use the FastText algorithm to obtain their semantics, and construct an input vector suitable for the Bi-LSTM model to detect log anomalies by learning the semantic information of the log context. Experimental verification shows that the F1 value of the model reaches 99 %, which is better than baseline methods such as DeepLog.