Need improvements in AIOPs MAD functionality

Criteria

·       Continuously ingest live metric streams from PRTG and SolarWinds, ensuring consistent formatting and data quality validation.

·       Identify and alert on metric anomalies, such as real-time bandwidth spikes or drops on MITA’s devices and sensors (monitored via PRTG and SolarWinds), by recognizing unusual patterns, deviations from historical baselines, or outliers, using adaptive thresholds that continuously adjust to changing conditions rather than relying solely on fixed rules.

·       Anticipate anomalies through predictive analytics based on historical trends and behavioral patterns.

Current Status and why we need improvements 

Basic functionality has been implemented but we feel that further fine tuning is still required here. For instance anomaly detection sensitivity, as slight deviations sometimes trigger alerts while obvious spikes or drops do not. This tuning must also account for seasonality, network activity varies significantly by time of day, weekday vs. weekend, and differs across individual devices.

Furthermore, recently we had pointed out if there is a way where we can access data about old anomalies, since as it stands one does not have a way to check which anomalies were reported if they have already been resolved. Currently, one needs to use the Metric search functionality to directly access the metrics for a specific device, and see if/when it had experienced anomalies. As a workaround, I had suggested to maybe have some sort of email notification when anomalies are raised, so that one would be able to access them accordingly in a back-dated fashion. This is just a basic idea, of course we are open to any idea from IBM's end.

Also, our understanding is that anomalies should be active during the entire period where the value is below the baseline. However, in the below example this is not the case. And this leads to multiple (potentially unnecessary) alerts.