Real Time Control of Prosthesis Using EMG
Transcription
Real Time Control of Prosthesis Using EMG
Real Time Control of Prosthesis Using EMG Aviv Peleg Or Dicker Supervised by: Tal Shnitzer Dr. Oscar Lichtenstein Introduction -In the USA alone there are 700,000 below the elbow amputees -Current hand prostheses capabilities range from single action to expensive (>20,000$) multi-action prostheses Single function prostheses 2 multifunction prostheses Introduction •Prosthesis for children is especially complex problem, since the growth of the child means many the need for many new prosthesis as the child grows. 3 Current Solution 4 Current Solution •3D blueprints for affordable single action prosthesis are currently available as “open source”. •With today’s technology, affordable multi-action prostheses can be developed for all in need. 5 Our Solution Sensor: Prosthesis MYO armband 6 3D printed hand with the Intel Edison Input-Output 7 The EMG The muscle Anatomical structure of muscle 8 The motor unit Generation of the EMG The EMG signal * = * = * = Amplitude ranging from -5mV to 5mV The frequency spectrum of the sEMG signal collected with commonly used sensors ranges from 0 to 400 Hz 9 N−1 X t = h i ∗ 𝛿(𝑡 − 𝑡𝑘 ) + n(t) i=0 Project Objectives Development of a control mechanism for a 3d printed Prosthesis: 10 Requirement Goal Real Time Less then 250ms from command to execution Reliability classification success rate of more than 90% Portability Fully portable Variety of gestures Range of 6 different gestures Short setup time Less then 1 minute for calibration All the above and also Affordable The Prosthesis 11 The Prosthesis 12 Algorithm Workflow Calibration Recording and labeling Feature extraction pattern recognition Hand Control Recording Decision Convert classification to servo movements 13 Real time classification Pre action Threshold Movement Calibration Calibration Recording and labeling Feature extraction pattern recognition Hand Control Recording Decision Convert classification to servo movements 14 Real time classification Threshold Movement Recording and Labeling Sensor: MYO armband EMG data from sensor: 15 Raw data Feature Extraction • Mean Absolute Value • Root Mean Square • Variance • Willison Amplitude • Modified Mean Frequency • Wavelet Coefficients • Double Threshold (for rest) 16 Feature extraction Mean Absolute Value 17 Data after Feature extraction Pattern Recognition • K Nearest Neighbors Algorithm Different K • Support Vector Machine Kernel 18 Example of K Nearest Neighbors boundary of decision Pattern Recognition • K Nearest Neighbors algorithm K=3 19 Example of K Nearest Neighbors boundary of decision Hand Control Calibration Recording and labeling Feature extraction pattern recognition Hand Control Recording Decision Convert classification to servo movements 20 Real time classification Threshold Movement Real Time Recording and Classification The hardware : The software : Microcontroller Intel Edison + self built shield 21 classification code in python Conversion of Classification to Servo Movements Servomotor: TowerPro MG996R Servo Command of Servo Angle engine Servo 1 Servo 2 Mechanical Output position Servo 3 gesture classification Thumb Index finger Digits 2 90° -90° 90° Index finger Closed Open Closed 3 90° 90° 90° fist Closed Closed Closed 4 -90° 90° 90° Thumb up Open Closed Closed Conversion table of classification to servo command to mechanical output 22 Movement of Prosthesis The six different hand gestures: fist, pinch, point, rest, Frisbee catch, like. 24 Overview BT 25 PWM Results Set goal 26 Real Time <250 msec Reliable >90% Portability Fully portable Variety 6 gestures set up time <1 minute Demo 27 Conclusion -All the objectives were met, proving the feasibility of our solution. -Further work is needed to reach a complete product. Mostly in regard to subjects outside the scope of this project. 29 Future Developments Remove need for daily calibration using Domain adaption algorithm. Implementation of control and feedback over amount of force generated. Combining the sensory control systems. Reduction of time from command to execution to Less then 100ms 30 ACKNOWLEDGEMENT 31 ▪ Dr. Oscar Lichtenshtein ▪ Mr. Sharon Ishar ▪ Prof. Yoav Medan ▪ Mr. Gal Pressman ▪ Mr. Hovav Gazit ▪ Mr. Johanan Erez ▪ Mr. Yair Herbst ▪ Mr. Koby Kohai ▪ Mr. Shlomi Dach ▪ Mr. Vasily Vitchevsky ▪ Mr. Oren Forkosh ▪ Mr. Yossi Bar Erez Questions? 32