Volume 2 - Issue 1, January - February 2026

Logs from distributed systems carry vital information about system behavior and failures. Detecting anomalies in logs (due to failures or attacks) is crucial for system reliability and security. We propose a self-learning AI framework for online log anomaly detection that continuously adapts to evolving log patterns. The system ingests raw logs (e.g. via Kafka), applies a streaming log parser (e.g. Spell) to extract log templates, and transforms them into feature vectors. A self-learning anomaly model (e.g. an LSTM autoencoder or an Incremental Isolation Forest) is trained on normal log streams. It outputs anomaly scores for each log sequence, and retrains itself automatically when drift is detected (with minimal human intervention). We evaluate on public benchmark logs (HDFS, BGL) under a simulated streaming setup, reporting Precision, Recall, F1, and detection latency. The proposed method achieves high detection accuracy (∼95–96% F1 on HDFS/BGL) with low false positives and fast response (∼200ms latency) comparable to state-of-art methods. We also incorporate explainability: e.g. attention mechanisms or SHAP/LIME scores highlight which log events or features caused an alarm. This helps analysts interpret anomalies. We discuss strengths (real-time, minimal manual tuning), limitations (heavy models require tuning, explanation overhead) and future work.

Nagarjun K, Pravin K S, Dr Kavitha V, Uthra V,"Log Anomaly Detection Using Self-Learning AI" International Journal of Advanced Multidisciplinary Research and Educational Development, V2(1): Page(583-589) Jan-Feb 2026. ISSN: 3107-6513. www.ijamred.com. Published by Scientific and Academic Research Publishing.