280 likes | 490 Vues
Forearm Surface Electromyography. Activity Detection Noise Detection, Identification and Quantification Signal Enhancement. Aim of research. Make myoelectric forearm prostheses more useable So far Onset detection Noise reduction. Today.
E N D
Forearm Surface Electromyography Activity Detection Noise Detection, Identification and Quantification Signal Enhancement
Aim of research • Make myoelectric forearm prostheses more useable • So far • Onset detection • Noise reduction
Today • Introduction to myoelectric signals, prostheses and control • Onset and activity detection • Carleton University’s CleanEMG - Noise detection, identification, quantification • Signal enhancement
Forearm Prosthesis Control • None (passive) • Realistic looking • Has a few basic uses • Body powered • User shrugs to open and close claw • Proprioception • Limited orientation • Myoelectric • Pick up muscle signals and interpret them into open and close commands • Mostly claw/pincer-type • First commercial limb in 1964
What myoelectric prostheses are not • No sensory feedback • No proprioception • One gesture at a time • Not as dextrous as natural hands • - No direct control of fingers • Not part of your body • Doff every night to charge • Takes a while to don the socket every morning
The iLimbState-of-the-Art Forearm Prostheses • Made by Touch Bionics in Livingston • Individually articulated fingers • Motors stall when ‘enough’ grip has been applied • Monitored by microprocessor • Clever re-use of open/close to allow more gestures • Can ‘pulse’ the motors to increase grip
Limitations of myoelectric prostheses • iLimb shares limitations with all modern commercial myoelectric prostheses: • Amplitude-based commands do not directly relate to desired gesture • Not all users can do all ‘double impulse’-type commands • Cannot address individual fingers • Manual thumb rotation for pinch and grip • Limited battery life – a day of normal use
Multi-channel raw sEMG signal (live or recorded) Generic Pattern Recognition System Sample Filter Onset/activity detection Windowing Feature extraction Classifier Majority vote Dimensionality reduction Class label stream
One-Dimensional Local Binary Patterns for Surface EMG Activity Detection
2-D Local Binary Patterns • For image analysis • Spatiotemporal LBP for video analysis http://www.scholarpedia.org/article/File:LBP.jpg
One-Dimensional (1-D) Local Binary Patterns • Take windows of signal • Calculate LBP codes within window • Form normalised histogram x[n] Sample number n 0 0 1 1 0 0 20 21 22 23 24 25 = 12 in decimal
x[n] 1-D LBP Activity Detection LBP code calculation 1-D LBP histogram calculation ‘Activity’ bins ‘Inactivity’ bins NO No activity Activity bins> Inactivity bins YES Activity
1-D LBP Bin Behaviour • Test on a synthetic signal (bandlimited Gaussian noise with AWGN 6dB)
1-D LBP bin behaviour • Test on single gesture of real EMG recording
1-D LBP Activity Detection • Once activity is detected, pattern recognition can be started • Can sum the LBP codes from multiple channels within a window to get a single decision
Carleton University’s CleanEMG • Access to an expert to manually identify and/or mitigate noise is not always possible • EMG can be contaminated with several types of noise • For each type, do some or all of these: • Detect • Identify • Quantify • Mitigate
Types of EMG noise • Power line (50Hz or 60Hz) • ECG • Clipping • Quantisation • Amplifier saturation Also • Baseline wander • RF
Features • Signal to Quantisation Noise Ratio • Signal to ECG Ratio • Effective Number of Bits • Signal to Motion Artefact Ratio • Power line Power (Least Squares Identification) SQNR SNR (ECG) ENOB SMR
Why a classifier? • Contaminants can be mistaken for each other if a single feature type is used • Motion artefact and ECG • Clipping and quantisation • Training a classifier should help to address this
Work done at Carleton • Improved Prof Chan’s and Graham Fraser’s CleanEMG Matlab code • Trained classifiers to identify contaminants using artificially-contaminated real and synthetic EMG • Indicated that detection and identification are harder for signals with higher SNR
Classification accuracy • The techniques lead to improvements in classification accuracy for noisy data • Data Set 1 (Recorded at Strathclyde) – a little, especially Channel 2 • Data Set 2 (Prof Chan’s) – improved • Data Set 3 (Italian) – improvement in some subjects • Classification accuracy is improved for noisy data
Raw sEMG signal (measured or recorded) PR system with a new stage Sample Filter Noise Detection, Identification, Quantification, Mitigation Data Windowing Onset Detection Feature Extraction Classifier Median Filter (Majority Vote) Dimensionality Reduction Class label