Unsupervised Log Sequence Anomaly Detection via Deep Embedded Fuzzy Clustering with FGM-VAE and LSTM

This paper presents FVLL, a novel unsupervised log sequence anomaly detection method that combines FGM-VAE and LSTM within a deeply embedded fuzzy clustering framework. Unlike existing methods, FVLL eliminates reliance on labeled data during both training and detection. It learns an end-to-end anomaly detection model using unlabeled log data, consisting of a deep feature extraction model and a fuzzy clustering model, optimized based on reconstruction and clustering losses. FVLL classifies log sequences into normal, abnormal, and fuzzy categories, further distinguishing the fuzzy class using cosine similarity. Experimental results on public datasets (HDFS and BGL) demonstrate that FVLL outperforms existing methods in detecting log sequence anomalies.