Controlling Wheelchair Motion with Electroencephalography

Fourth Year Design Project Seminar Slides

Transcript

1. Team #43 Controlling Wheelchair Motion with Electroencephalography Calvin Ng, Edmund Lo, Jeff Tran, Jessica Woo & Safwan Choudhury Project Supervisor: Dana Kulić
2. Problem Background Some disabled individuals still immobile Thought-controlled actuation could one day restore mobility Challenges of thought-controlled actuation via electroencephalography Brain-computer interfaces (BCI) are an active area of research 2
3. General Framework Brain wave activity is recorded via EEG Offline training of machine learning algorithm Online classification of brain wave signals passed to microcontrollers Incoming signal processed and corresponding control signals generated Motor drives execute motion through drivetrain Trajectory corresponding to user trained brain wave pattern is executed Tracking achieved via closed loop control with sensory feedback 3
4. Proposed Solution Robust platform for future research in the field of BCIs Extensible BCI interface allowing for various novel EEG classification Overall low cost of equipment setup and assembly Efficacy of solution demonstrated via wheelchair application 4
5. EEG Interface Requirements Ability to extract raw EEG signals over USB for analysis Extensive SDK available for interfacing with microcontrollers Wireless operation desired Self-contained hardware and software package requiring minimal configuration 5
6. EEG Interface Selection Surveyed available solutions Selected Emotiv EPOC headset Hardware:14 electrodes, wireless Software: Machine learning algorithm, training interface, extensive API 6
7. EEG Interface Actions Determines 5 cognitive actions: 1: Push (Forward) 2: Pull (Back) 3: Left (Turn Left) 4: Right (Turn Right) 5: Neutral (Halt) 7
8. EEG Interface Training Split into 3 stages Step 1: Undistracted environment with software assist Step 2: Undistracted environment without software assist Step 3: Distracted environment 8 Day 1 Day 1 Day 2 Day 2 Day 3 Day 3 Day 4 Day 4 Day 5 Day 5 Day 6 Day 6 Day 7 Day 7 Action Hours Action Hours Action Hours Action Hours Action Hours Action Hours Action Hours Push Neutral 2 Pull 3 Push Pull 4 Left 3 Right 3 Left Right 4 All 5
9. EEG Interface Training 8-second training cycle for each action Repeatedly train according to schedule Completed offline and user profile stored for future use in online classification 9 Increasing the number of concurrent actions increases the difficulty in maintaining conscious control over the Cognitiv detection results. Almost all new users readily gain control over a single action quite quickly. Learning to control multiple actions typically requires practice and becomes progressively harder as additional actions are added. Although Emotiv Control Panel allows a user to select up to 4 actions at a time, it is important that each user masters the use of the Cognitiv detection one action at a time, only increasing the number of concurrent actions after he has first gained confidence and accuracy with a lower number of actions. Figure 11 Cognitiv Suite Panel 3.5.2 Understanding the Cognitiv Panel Display The Cognitiv Suite panel uses a virtual 3D cube to display an animated representation of the Cognitiv detection output. This 3D cube is also used to assist the user in visualizing the intended action during the training process. The Power gauge to the left of the 3D display is an indicator of the “action power”, or relative certainty that the user is consciously visualizing the current action. The default tab on the Cognitiv panel is the Action tab. This tab displays information about the current state of the Cognitiv detection and allows the user to define the current set of actions. In order to enable the Cognitiv detection, each chosen action, plus the Neutral action, must first be trained. For more information about the training process please refer to
10. Dynamic simulation used to provide design requirements for wheel torque Approximate wheelchair model reconstructed in SolidWorks to obtain estimated inertial parameters. Frictional force on wheels obtained through experimental trials Simulation Result: Required Wheel Torque = 14 Nm 10 Electromechanical Design Dynamic Simulation
11. Low cost actuator design One actuator per wheel Brushed vs. Brushless DC Motors Control and drive considerations Selected Actuator: 12V CIM Brushed DC Motors 11 Electromechanical Design Actuator Selection
12. Low cost gear train design Need 14:1 reduction ratio while operating at peak motor efficiency. Planetary/spur gear heads too expensive Decision: Design simple & custom gear train to achieve desired reduction ratio at minimal cost. 12 Electromechanical Design Gear Train Selection
13. 13 Electromechanical Design Gear Train Design Space Cost Direct Drive (Friction) 1 Stage Gearing 2 Stage Gearing Low Low High Moderate High High Final Design 1 stage gearing with 14:1 reduction between 112 teeth sprocket and 8 teeth sprocket coupled together with ANSI standard chain #25
14. Low cost requirement for batteries. Need 12V DC for selected motors Rechargeable via lead acid battery recharging station (wall plug-in) Battery Selection: Motomaster 12V 35AH Sealed Lead Acid Battery (for electric wheelchairs) 14 Electromechanical Design Power System
15. 15 Electromechanical Design Power System + Encoder Encoder 5V GND Input Input GND PWR Enable M2+ M2- GND PWR Enable M1+ M1- M2 _ + M1 _ + 12V Driver 12V Driver 12V Battery Toggle Switch Arduino 5V GND 5V GND USB Out 60 Amp Fuse
16. Sprocket and chain mechanism enclosed by fenders 60A 12V Toggle Switch as hardware failsafe mechanism 60A Fuse on Battery Circuit for added protection 4 AWG wires to safely handle high current draw when starting up Battery encased separately 16 Electromechanical Design Safety Features
17. Control Strategy Overview Arduino Duemilanove Closed loop PD controller Pololu High Power Motor Driver Microcontroller puts out PWM value to be amplified by electric motor drives Cytron Simple Rotary Encoders Optical encoders mounted to output shaft of DC motors for speed control 17
18. Control Strategy Microcontrollers 2x Arduino Duemilanove’s (16MHz ATmega328 Microprocessor) Master/slave(s) configuration between main controller and sensory inputs Communication between microcontrollers via I2C bus Expandable architecture to include more sensory feedback sources for better control 18
19. Control Strategy Master Controller Executes main control loop at 200Hz Interfaces with PC running machine learning algorithm over serial port Periodically requests the slave(s) for updated sensory input values via I2C Logs diagnostic information in its onboard EEPROM 19
20. Control Strategy Slave Controller Calculates the wheel RPM at 400Hz via optical encoder pulses Calculates the battery terminal voltage via separate circuitry to update the masters motor speed constant value Periodically services the master for updated sensory input values via I2C 20
21. 21 Control Strategy Control Loop Kp d/dt Kd DC Motor Encoder ωreference - + Σ Σ + + ωactual
22. 22 Control Strategy Step Response
23. 23 Control Strategy PWM Signals
24. Microsoft Robotics Studio 2008 Microsoft Simulation Environment Preprogrammed 3D environment to provide a virtual training ground for thought-controlled actuation. 3 modes of operation: empty field, obstacle course and maze. Simulation environment updates in parallel with the physical wheelchair moving in real time. 24 Simulation Overview
25. 25 Simulation Obstacle Mode
26. 26 Simulation Maze Mode
27. 27 Summary Project Costs Item Usage Quantity Cost/Unit EEG Headnet Microcontrollers DC Motors Motor Drivers 12V Battery 112 Teeth Sprocket 8 Teeth Sprocket ANSI Chain #25 Wheelchair Misc Control 1 $750 Control 2 $75 Mechanical 2 $32 Control 2 $50 Power 1 $130 Mechanical 2 $65 Mechanical 2 $15 Mechanical 1 $30 Mechanical 1 - Various ~ $70 $1,454
28. Summary Future Work Improve classification accuracy by developing a better algorithm Revise electromechanical design to enable variable speed control Embed new sensory information (i.e. proximity sensing) to improve navigation Develop a better method of enclosing electromechanical components (gear train, wires from controllers, etc) 28
29. Thank you for your time. Questions?