Skip to main content
The Anomaly detection learning block is a flexible anomaly detection option available when designing an impulse in Edge Impulse Studio. Unlike specific anomaly algorithms such as GMM or K-means, this block does not enforce a particular method — instead, it provides a generic container for any supported custom anomaly detection approach. This is the default destination for custom anomaly detection blocks (excluding visual anomaly detection, which uses the Visual anomaly detection block).

​
Overview

Anomaly detection is the process of learning a representation of normal (non-anomalous) behavior and identifying deviations from that baseline. Models trained for anomaly detection output an anomaly score, with higher scores indicating greater deviation from the learned normal pattern. Typical anomaly detection workflows:
  • Train on samples that do not include anomalies
  • Build a model of normal behavior
  • Evaluate new inputs to compute an anomaly score
The Anomaly Detection block acts as a general host for any custom algorithm that implements the expected learning block interface. This includes Edge Impulse-provided blocks and custom learning blocks.

​
Setting up the Anomaly Detection learning block

To add this block to your impulse:
  1. In Edge Impulse Studio, open Create impulse
  2. Click Add learning block
  3. Select Anomaly Detection from the list
  4. Configure any available parameters exposed by your custom learning block implementation
  5. Click Save impulse

​
Example configuration

Custom anomaly detection blocks may vary, but typical parameters include:
  • Thresholds or confidence settings — may be exposed to control when an anomaly is triggered.
  • Model-specific hyperparameters — depending on the algorithm.

​
Train

Once configured, click Start training to begin learning your anomaly detection model. Unlike classification or regression, training for anomaly detection normally expects no anomalies in the training set. After training completes, Edge Impulse Studio will show the training status and allowable controls such as Model testing.

​
Testing the Anomaly Detection learning block

Navigate to the Model testing page and click Classify all to score your test dataset. You can inspect individual samples and their anomaly scores. If your custom block supports it, you may also see options for:
  • Anomaly Explorer visualizations of multi-dimensional feature space
  • Threshold adjustment to tune sensitivity

​
How does it work?

Unlike fixed algorithm blocks (e.g., GMM or K-means), the Anomaly Detection block itself does not define a specific method. Each custom block packaged under this learning block must:
  1. Receive training features (from upstream processing blocks)
  2. Learn a model of normal behavior
  3. Output a per-sample anomaly score at inference
These steps are implemented by the custom block according to its algorithm.

​
Additional resources