Embedded Machine Learning (TinyML) Crash Course

TinyML from zero to hero 

(The first technical TinyML course in Egypt)

Projects per course:

• Gesture Recognition Using Camera

• Voice Command Recognition

• Smart Home Temperature Control

• Anomaly Detection in Sensor Data

• Predictive Maintenance for Machines

• Real-Time Object Classification Using Camera Module

• Energy Consumption Prediction

• Wearable Fitness Tracker

• Smart Irrigation System

Course content

Lecture 1: Introduction to TinyML

Project: Overview of TinyML and Gesture Recognition Using Camera
Concepts:

1. What is TinyML? Applications and relevance in IoT.

2. Overview of ML on constrained devices (MCUs, ESP32, Raspberry Pi).

3. Dataset collection using camera modules for gesture recognition.

4. Introduction to TensorFlow Lite for Microcontrollers.

5. Python basics for preprocessing image data and building ML models.

Practical:

• Use Python and OpenCV to capture images for gestures.

• Train a CNN model to classify gestures (e.g., wave, thumbs up).

• Deploy the trained model on an embedded device.

Lecture 2: Voice Command Recognition

Project: Building a TinyML system for voice recognition.
Concepts:

1. Fundamentals of sound waves and audio features (MFCC, spectrograms).

2. Dataset preparation for voice commands (e.g., "On", "Off", "Stop").

3. Neural network architecture for speech recognition.

4. Quantizing models for deployment on low-power devices.

Practical:

• Record audio commands using a microphone in Python.

• Train a model to classify commands using TensorFlow.

• Deploy the model on ESP32 with a microphone module.

Lecture 3: Smart Home Temperature Control

Project: Automated temperature control based on DHT22 sensor data.
Concepts:

1. How sensors interact with TinyML models.

2. Regression models for continuous data prediction.

3. Preprocessing time-series data for training.

4. Creating control logic for smart devices.

Practical:

• Collect temperature and humidity data using Python.

• Train a regression model to predict comfort levels.

• Deploy the model and control actuators (e.g., fans or heaters).

Lecture 4: Anomaly Detection in Sensor Data

Project: Detects anomalies in environmental sensor readings.
Concepts:

1. Understanding anomaly detection and unsupervised learning.

2. Feature extraction from sensor data.

3. Using Autoencoders for anomaly detection in TinyML.

4. Implementing real-time monitoring and notifications.

Practical:

• Simulate sensor data (e.g., temperature spikes) in Python.

• Train an Autoencoder to detect anomalies.

• Deploy the model and trigger alerts for anomalies.

Lecture 5: Predictive Maintenance for Machines

Project: Monitoring vibrations to predict machine failures.
Concepts:

1. Introduction to predictive maintenance and its use cases.

2. Collecting and preprocessing accelerometer data.

3. Building and training a classification model to detect failure patterns.

4. Data transfer between sensors and microcontrollers.

Practical:

• Use Python to simulate vibration data or use an accelerometer.

• Train a model to classify normal vs. abnormal states.

• Deploy the model on an embedded device to monitor real-time data.

Lecture 6: Real-Time Object Classification Using Camera Module

Project: Classifying objects using a camera module (e.g., ESP32-CAM).
Concepts:

1. Image classification basics and MobileNet architecture.

2. Dataset preparation and augmentation for TinyML.

3. Optimizing ML models for camera input.

4. Real-time image processing on constrained devices.

Practical:

• Capture object images using Python and OpenCV.

• Train a MobileNet model on TensorFlow.

• Deploy the model to classify objects on a camera module.

Lecture 7: Energy Consumption Prediction

Project: Predicting energy usage based on historical data.
Concepts:

1. Working with time-series data for TinyML.

2. Using LSTM models for sequential data prediction.

3. Data preprocessing techniques: sliding windows, normalization.

4. Deployment challenges for LSTM models in TinyML.

Practical:

• Simulate energy consumption data using Python.

• Train an LSTM model to predict future consumption.

• Deploy the model to analyze real-time data from sensors.

Lecture 8: Wearable Fitness Tracker & Smart Irrigation System

Projects: Activity tracking and soil moisture-based irrigation.
Concepts:

1. Data collection from wearable sensors (e.g., accelerometer, gyroscope).

2. Classification models for fitness tracking (e.g., steps, jumps).

3. Regression models for moisture prediction and irrigation control.

4. Practical challenges in battery optimization for TinyML devices.

Practical:

• Build a fitness tracker using Python to process accelerometer data.

• Train a model to classify activities like walking and running.

• Use a soil moisture sensor to control a water pump via TinyML.

Tools & Frameworks:

1. Python Libraries: TensorFlow, Numpy, Pandas, Scikit-learn, Matplotlib.

2. Hardware: ESP32, Raspberry Pi, DHT22, accelerometers, microphones, and cameras.

3. Deployment: TensorFlow Lite for Microcontrollers, Arduino IDE, PlatformIO.