Being able to detect anomalies is becoming an extremely useful technique in the world of embedded ML, and one that can be done on the most constrained always-on devices. Anomaly detection can be used for a multitude of use cases, from cold chain monitoring to fault detection on industrial machinery or satellites.

In this talk we will present how to use data-driven engineering to create your data set and use Edge Impulse to create a model able to classify anomalous sensor readings. We will do this by leveraging some new powerful features in Edge Impulse.

You will learn to:
· Implement and use custom DSP blocks to analyze your IoT data and extract the most important features
· harness the value of feature importance to zoom in on interesting frequency bands
· iterate over thresholds in anomaly detection blocks, in order to find the optimal configuration