DBB220 / 2019 Semester B Quartile 4 / TU Eindhoven Industrial Design
Responsible Lecturer: Dr. Rong-Hao Liang
- Assistant Professor, Future Everyday Group, Industrial Design, TU Eindhoven
- Assistant Professor, Signal Processing Systems Group, Electrical Engineering, TU Eindhoven
- Homepage: http://www.ronghaoliang.me/
Sensors not only enable the interactivity of products but also generate data. Machine intelligence leverages computational power and data to empower further the product to deal with design problems involved prediction, decision, and adaptation. This course aims to help the students understand the main paradigms in sensing, data collection, signal processing, and machine learning to apply them in meaningful design solutions of intelligent interactive products.
The course combine theories and practices. Through a series of lectures and workshops, you will learn the principle and functionality of the sensors and machine intelligence. You will develop the ability to use these signal processing methods and machine learning algorithms to deal with everyday life problems where real-world complexity, uncertainty, and changing conditions make the use of these technical solutions a necessity. Understanding the type of problems that really will benefit from the utilization of signal processing methods and machine learning algorithms and creating realistic scenarios of use is equally important.
Arduino, Processing, and Weka are used as the development platform. Ready available example codes are available in Github repository, so you don’t have to implement the algorithms from scratch.
- Microcontrollers: Please get at least one of the following microcontroller boards from an online store.
- All the example codes used in this course were made for Arduino Leonardo. The small Arduino board that I used in the lecture is DFRobot Beetle, which is available in Kiwi electronics. Teensy and Circuit Playgroundare also Arduino-compatible. For this course, I highly recommend Circuit Playground or Circuit Playground Express platforms, which have various on-board sensors so that you don't need to solder or interface additional sensors by soldering.
- Sensors: Please get at least one of the following sensor modules from an online store.
- The 3-Axis Accelerometer module, which will be used throughout all the exercises. It's important to note that most of the accelerometers only support 3.3V power supply. For Arduino users, you should either plug the sensor to the 3.3V or simply find a 5V-compatible Adafruit ADXL335. I recommend this LilyPad ADXL335 because it supports analog output, instead of an I2C one that needs more complex wiring and an extra library.
- for Teensy users, you have more choices because Teensy uses 3.3V by default. I recommend this LilyPad ADXL335 because it supports analog output, instead of an I2C one that needs more complex wiring and an extra library. If you would like to try more sensors, consider Teensy Prop Shield.
- for Circuit Playground / Circuit Playground Express users, you have an on-board accelerometer that works well, so you don't need to worry about this.
- If you cannot get an accelerometer, I would recommend you to find other analog sensors (e.g., LDR, FSR, Capacitive Sensors) instead. The more reactive the sensor, the better, because it's easier for debugging.
- Equipment (Optional)
- To connect the sensors and the microcontroller, you might need solder equipment, wires, and breadboard. For debugging, you might need a multimeter. Please try to get your "home lab" ready so you will be free to explore.
We use Processing as the main software programming environment because it's powerful, simple, and versatile. It also support popular machine learning libraries such as WEKA. Please get the Processing IDE installed before lectures.
-
Installation
- Get the latest version (3+) from the Processing download page: https://processing.org/download/?processing (Links to an external site.)
- Unzip the downloaded file and run the “processing” program. You should be able to run Processing without installation.
- For Windows, it is recommended to move the program file to
D:\processing\ - For Mac OS X, move it to
/Applications/ - For GNU/Linux, move it to
/usr/local/processing/
- For Windows, it is recommended to move the program file to
-
Overview Every Processing project is a sketch of codes. There are multiple built-in examples that could help us understand sketch files.
As a Hello-World example for checking whether the installation is successful, please
- Click File > Examples…
- Open Basic > Input > Mouse2D example.
- Hit the Run button to play with it. More tutorials on Processing can be found on the official “Hello Processing" website. http://hello.processing.org/
We use Arduino IDE and the Wiring language for hardware configuration, regardless of which microcontrollers (e.g., Arduino, Teensy, Circuit Playground) you are using. Arduino IDE is an open-source micro-controller development platform that allows us to connect electronic devices, such as LED, servo motor, button, or sensor and build the features we want. Please get the Arduino IDE installed before lectures.
- Install the Arduino Software Get the latest version from the Arduino download page: https://www.arduino.cc/en/Main/Software. The page provides install files for Windows, Mac OS X, and Linux as shown below. Download the install file of your operating system. Open file after downloading and proceed with the installation. (You could also take a look at the official documents for Windows, Mac, and Linux.) More tutorials of Arduino can be found Here.
- Overview
- Prerequisites
- Schedule
- Module 1: Machine Learning
- Module 2: Signal Processing
- Synergies Between The Two Modules
- Why Machine Learning?
- Formulate Supervised Learning problems
- Types of Learning
- CSV vs. ARFF
- Save data as ARFF files
- Load data as ARFF files
- Timer and Uniform Sampling
- Send Sensor Data Stream to Processing
- Save Serial Data as ARFF files
- Bidirectional Serial Data Streaming between Arduino and Processing
- Noise in Data
- Linear Support Vector Classification
- Linear Support Vector Classifier
- Train a classifier from data
- Confusion matrix and In-sample accuracy
- Overfitting and out-of-sample accuracy
- Regularization
- k-fold Cross-Validation
- Optimize a prediction model
- Load a test set and get the out-of-sample accuracy
- Real-Time Posture Classification
- Linear Regression
- Linear Regressor
- Train a regressor from data
- Loss function
- Linear Regression in higher dimension
- Real-Time Posture Regression
- Visualizing Sensor Data Stream
- Segmentation and Activation Threshold
- Windowing and Statistic Feature
- Real-time motion classification
- Real-time motion regression
- Using More Features
- More Classifiers
- K-Nearest Neighbors Classifier
- Kernel Support Vector Classifier
- More Regressors
- Linear Support Vector Regressor
- Kernel Support Vector Regressor
- Hyper-parameter Tuning
- More Evaluation Metrics
- True Positive Rate
- False Positive Rate and Specificity
- Precision
- F1 score
- Area Under ROC
- More than Two Classes
- Reporting
- Validation Methods
- Validate the product with users
- Comparing different algorithms
- Reporting
- Time Series in Frequency Domain
- Sampling Theory
- Fast Fourier Transform
- Window Overlapping vs. Information Loss
- Segmentation, Windowing, and Filtering
- Real-time audio classification using a microphone
- Real-time vibration classification using an accelerometer
- Human Activity Recognition Systems
- Synchronizing Sensor Data Streams
- Context Recognition using Sensor Fusion
- Feature Selection
- Dimensionality Reduction
- Image Recognition and Computer Vision
- Bitmap Data Structure
- Statistic Features: Histograms
- Structural Features: Edges
- Simplification and Downsampling
- Segmentation methods
- Haar Cascade Classifier
- Application: Real-time Face Detection
- Marker-based Object Tracking
- Infra-red Cameras
- Artificial Neural Network
- Neuron and Activation Functions
- Forward and Backward Propagation
- Gradient Descent and Learning Rate
- Training an ANN
- Convolutional Neural Network