ETD

• AIOps
• anomaly detection
• predictive analytics
• server monitoring

Anomaly detection techniques have been widely used in time series contexts, and they have been applied to a large number of areas, including server monitoring tools. With the advancement and spreading of Artificial Intelligence (AI), these methods have evolved towards more automated approaches, giving rise to the concept of Artificial Intelligence for IT Operations (AIOps). In this work, a monitoring system for Sybase ASE servers was developed based on AIOps principles, using a predictive and explainable approach to detect anomalies in multivariate time series composed of several aspects of performance counters. To achieve this, a modular architecture was proposed, that covered data ingestion, preprocessing, and anomaly detection through unsupervised models, as well as the inclusion of an Explainable Artificial Intelligence (XAI) component. Among the tested models, it was found that those based on Variational Autoencoders (VAE), such as LSTM-VAE and OmniAnomaly, were more effective in capturing temporal dependencies and relationships among metrics, which can be complex. The anomaly explanations were generated both globally and locally, using Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) respectively, in order to identify the influence of server performance metrics on the occurrence of anomalies. Finally, the results were integrated into a dashboard that displays historical metrics, predictions, and their corresponding explanations, becoming a tool that supports transparent and informed decision-making.