predictive log analytics

Project Goal: To develop an automated Machine Learning (ML) classification system to analyze raw system log data. The primary objective is to accurately predict system failures and detect anomalies in system behavior, thereby enhancing operational efficiency and improving proactive maintenance strategies.Dataset: The project utilizes the BGL (BlueGene/L) open-source log dataset from the Lawrence Livermore National Labs, which is rich with labeled alert and non-alert messages, making it ideal for prediction tasks.Core Methodology & Techniques:Data Preprocessing and Transformation:Performed data cleaning, type conversion, and handling of missing values.Employed CountVectorizer and SentenceTransformer to compute similarity between text columns, leading to the strategic removal of redundant columns (e.g., NodeRepeat, EventTemplate) to improve data quality.Exploratory Data Analysis (EDA):Identified that the KERNEL component is most frequently associated with FATAL errors.Analyzed common log entries, revealing technical patterns related to error handling and interrupt management (e.g., words like interrupt, error, except).Advanced Log Classification (NLP):Model Training: Trained and evaluated three state-of-the-art Transformer models: BERT, DistilBERT, and RoBERTa.Performance: The RoBERTa model delivered the highest performance, confirming its effectiveness for this text classification task:Accuracy: $\approx 94.00\%$F1 Score: $\approx 94.12\%$Conclusion: The project successfully implemented and validated a highly accurate RoBERTa-based classification model, providing a robust solution for critical operational issue identification and predictive maintenance.