Initialize Model¶
Initializing an Insights anomaly detection model primarily involves ensuring the right data is available and organized, and then interacting with the system, often via APIs, to define the learning scope and to activate real-time monitoring and alerting.
Model Initialization and Learning (via API)¶
While Insights is largely self-supervised, you interact with it programmatically to define its learning scope and manage its lifecycle.
Identify Signal Sampling Rate¶
Insights learns what "normal" behavior looks like by observing data within specified learning periods.
- Identify time ranges in the recent historical data that represent the expected, healthy operation of your systems
- Choose the specific numeric signals for which you want Insights to detect anomalies
- Use the ANOMALYLEARN flow to initiate the model learning process. Every successful completion of ANOMALYLEARN will generate a unique anomaly detection model with a unique index
M[#]and ID. This API call will require:- The timeRanges that contains largely "normal" operations
- The signal ID for which the model should be learned
- Optionally, config settings can be applied for specialized circumstances
- Aperture: Typically, it is observed that using the default setting for this parameter builds a good-enough anomaly detection model. This size controls the output latency. Consult Falkonry Support for best practice guidance on what aperture to set for your use case.
- Sampling period
Together, aperture and sampling_period define how much data and how frequently that data is sampled to build the baseline behavior for Insights anomaly detection. They are critical to tuning the model's sensitivity and accuracy for different types of operational signals.
Incremental Learning¶
You can build upon existing Insights models by providing additional "normal" learning ranges to the ANOMALYLEARN API, which helps the model adapt to new behaviors or reinforce learned patterns.
How to initialize an anomaly model using API has more details.