Skip to main content

Create Impulse

An Impulse is the machine learning pipeline in Edge Impulse. It defines how raw sensor data is processed, what features are extracted from it, and what learning algorithm is used to classify or predict from those features.

Think of it as the blueprint for your model.

How It Works​

An Impulse has three stages:

Raw Data  →  Signal Processing (DSP)  →  Learning Block  →  Output
  1. Input block — defines the type of data (time series, image, audio) and the window size.
  2. Processing block (DSP) — extracts meaningful features from the raw data (e.g. spectral features from accelerometer data, MFCCs from audio).
  3. Learning block — the ML algorithm that trains on those features (e.g. a neural network classifier, K-means anomaly detection).

Creating Your Impulse​

  1. In Edge Impulse Studio, go to Impulse design → Create impulse.
  2. Set the input block — choose your sensor type and window size. For gesture recognition with the IMU, a window of 1000ms with a stride of 500ms is a good starting point.
  3. Add a processing block — click Add a processing block. For IMU data, select Spectral Analysis. For audio, select MFCCs.
  4. Add a learning block — click Add a learning block. For classification, select Classification (Keras). Edge Impulse will recommend one based on your data type.
  5. Set your output labels — these are the class names from your dataset (e.g. wave, punch, idle).
  6. Click Save Impulse.

Configure the Processing Block​

After saving, click on the processing block (e.g. Spectral features) in the left sidebar to configure it.

  • You'll see a live preview of how the raw data is transformed into features.
  • Use the default settings to start — they work well for most IMU and audio projects.
  • Click Save parameters, then Generate features.

Generating features processes your entire dataset and produces a Feature Explorer — a 3D visualisation showing how well-separated your classes are. Well-separated clusters mean the model will likely perform well.

Configure the Learning Block​

Click on the learning block (e.g. Classifier) to set up training:

  • Number of training cycles — how many epochs to train for (start with 30–50).
  • Learning rate — how fast the model learns (default 0.0005 is a safe starting point).
  • Neural network architecture — Edge Impulse provides a default architecture. You can customise it as you get more comfortable.

You don't need to change these settings for your first model. The defaults are well-tuned for small edge datasets.

Your Turn​

  • Go to Impulse design and create a new Impulse for your project.
  • Add the correct processing block for your data type.
  • Add a Classification learning block and set your output labels.
  • Save the Impulse and generate features.
  • Open the Feature Explorer and check that your classes form distinct clusters.