IOT108 · Edge AI and Smart Automation

Run intelligence where the data is born: prepare sensor data, train and compress small machine-learning models for microcontrollers and combine them with rules into safe, automated responses.

Advance · 8 hours · 4 lessons, a lab, an assessed assignment and a portfolio piece

Who it is for

Learners who can build connected sensor systems and want devices that recognize patterns and act locally without depending on the cloud.

Prerequisites

IOT106 (IoT platforms) and basic Python; IOT107 is recommended for secure deployment.

What you will be able to do

• Explain when edge inference beats cloud inference using latency, bandwidth, privacy and cost.
• Prepare windowed sensor data, extract features and train a small classifier or anomaly detector.
• Quantize and size a model for a microcontroller and evaluate its accuracy and false-alarm rate.
• Design a smart automation that combines model output, rules, fail-safes and human confirmation.

Lessons

• Lesson 1 · Why put AI at the edge
• Lesson 2 · From sensor signals to a deployable model
• Lesson 3 · Smart automation: models, rules and safety
• Lesson 4 · Lab and assignment

Lab: Vibration anomaly detection for a small workshop motor

You generate and label accelerometer windows for a motor running normally, with imbalance and with a loose mount, extract features and train a compact model in Colab. You quantize it, compare size and accuracy and then implement the resulting decision logic on a simulated ESP32 that raises a local alert and publishes only events, not raw data.

Free tools: Google Colab (free) with scikit-learn and TensorFlow Lite, Wokwi ESP32 simulator with MPU6050 accelerometer; Edge Impulse free developer plan optional

Assessed assignment

Submit your Colab notebook, the Wokwi project and a two-page design note for a workshop owner. Report model accuracy, the confusion matrix, the false-alarm rate and the model size before and after quantization. Describe the automation logic: what the device does on each state, which actions require human confirmation and what happens if the sensor fails. Compare the bandwidth of event-only reporting with raw streaming and discuss how you would detect model drift over time.

Portfolio piece

A documented edge-AI pipeline, from sensor windows to on-device classification with event-only reporting, that demonstrates applied TinyML skills.

Self-paced, 12-month access. Questions are answered in writing through the student panel. A completion record is issued when you finish the course.