sEMG ADC

teensy3.2 ADC is 3.3v tolerant

MPU angle

• Gesture detection and data fusion based on MPU9250 sensor
  • Kalman filter
  • Pitch only
• Code available: Arduino sketches for MPU9250 9DoF with AHRS sensor fusion
  • Ask if appropriate to use

• Acce. test
  https://github.com/dymnz/sEMG/blob/master/Angle_Arduino/Angle_calc/Angle_calc.ino
  https://github.com/dymnz/sEMG/blob/master/Reports/wang/exp4_2017_8_24.MD
• Online verification

Protocol

Trial

1. 90d zero-load
2. Fixed load, varying angle

Observe

1. 90d zero-load normalization
2. Angle's effect on Weight-sEMG curve, angle-sEMG curve
   • Angle's effect on sEMG should be a constant multiplier. see Surface EMG force modeling with joint angle based calibration

• Dry electrode test
  • Compare dry electrode & adhesive electrode signal noise #14
• Hardware design
  • Prototype, for dry electrode test
  • Finalized design

Need 5V / +2.5V / -2.5V / GND
Maxim IC?

processing_angle_read

Arduino disconnecting after upload.

https://github.com/dymnz/sEMG/blob/master/Reports/wang/Paper/Joint%20Angle%20Estimation.md

Using '1.lvm', find the power spectrum in the 'still' part at the beginning

Conclusion

It can be seen that 60 Hz PLI is the dominant noise

Folder | Template

• Important: ⭐
• Reading 🔥

Papers

• A Wearable System for Recognizing American Sign Language in Real-Time Using IMU and Surface EMG Sensors
• 🔥Surface EMG force modeling with joint angle based calibration

Meeting

• Meeting 3

1. • Title: Continuous Motion Decoding from EMG Using Independent Component Analysis and Adaptive Model Training
   • 6-ch sEMG, ICA, PCA
   • Example: 3-DOF shoulder angle & 1-DOF elbow angle
2. • Title: EMG-based simultaneous and proportional estimation of wrist/hand kinematics in uni-lateral trans-radial amputees
   • 7-ch sEMG, motion capture, TDAR feature set, Accuracy: 72+-8%
   • Example: 3-DOF wrist movement

Applying in 4-ch exp.

Summary

• PCB layout
  • Verify inverting summing amp.
  • Change the resistor on the design manual to the value on current plug board
• Test

samples/sec test SamplingRateTest1() vs. SamplingRateTest2()

Arduino M0 PRO

• With ADC_CTRLB_PRESCALER_DIV512
  • Function call: 10031
  • While loop: 10166

=> Difference = (10166 - 10031) / 10166 % = 1.33%

• WIth ADC_CTRLB_PRESCALER_DIV256
  • Function call: ~19500 (unstable)
  • While loop: ~19500 (unstable)

=> Difference = ?

• WIth ADC_CTRLB_PRESCALER_DIV128
  • Function call: ~32800 (unstable)
  • While loop: ~33500 (unstable)

=> Difference = (33500 - 32800) / 33500 % = 2.08%

• WIth ADC_CTRLB_PRESCALER_DIV64
  • Function call: 49450
  • While loop: 52480

=> Difference = (52480 - 49450) / 52480 % = 5.77%

• Simultaneous Force - sEMG recording
• 1-ch sEMG + 1-ch Force w/ graph=
  • ~9.4k SPS @ 512 clock div.
  • 13k ~ 16k SPS @ 128 clock div.
• 2-ch sEMG + 1-ch Force w/ graph
  • ~5.1k SPS @ 512 clock div.
• 2-ch sEMG + 1-ch Force + 1-ch MPU (naive angle calc.) w/ graph
  • 1024 SPS, exactly. Probably due to MPU I2C polling.
• 2-ch sEMG + 1-ch Force + 1-ch MPU (complex angle calc.) w/ graph
  • ~460 SPS @ 512 clock div.
  • ~580 SPS @ 128 clock div.
• 2-ch sEMG + 1-ch MPU (complex angle calc.) w/ graph
  • ~525 SPS @ 128 clock div.
• 4-ch sEMG + 1-ch MPU (complex angle calc.) + 1-ch pot. w/ graph
  • ~460 SPS @ 128 clock div.

Sampling rate @10kHz. FFT are correct up to 5kHz, ADC sample interval seems to be fixed.

Waveform & FFT

0.5kHz

1kHz

3kHz

4kHz

5kHz

How to find out electrode-muscle corresponding pair at startup.

ICA?

This is the main research point

On going.

Given a sEMG signal, find a way to detect if force is used.

• Signal files given

Parameter

• 1~3 electrodes
• 1~10kg dumbbell

ToDo

• Buy Dumbbell or Water bottle + scale
• ADC PCB: #21
• Check sEMG-Force curve

Result

• 1st experiment
• Inter-session experiment

• Similar parameter
• Clear methodology
• Classify paper for grouping
  • Action length
  • Action type
• Electrode
  • Placement
  • # of electrode

See #40 for papers

Preprocessing used in Wrist Angle Estimation Papers

From #40

• Surface EMG pattern recognition for real-time control of a wrist exoskeleton (2010)

• Simultaneous and Proportional Force Estimation for Multifunction Myoelectric Prostheses Using Mirrored Bilateral Training (2010)

  • Feature set performance: MSV(?) < TD = TD+6AR = TD+5 wavelet marginal(TDWV)(?)
  • 7-ch semg -> TD feature -> PCA -> MLP

• Comparison of regression models for estimation of isometric wrist joint torques using surface electromyography (2011)

• SVM for Estimation of Wrist Angle from Sonomyography And SEMG Signals (2011)

  • RMS for amplitude
  • then wavelet to remove fluctuation

• EMG-based simultaneous and proportional estimation of wrist/hand kinematics in uni-lateral trans-radial amputees (2012)

  • TD feature set + 6 AR, not suitable for RNN. See A new strategy for multifunction myoelectric control(1993)
    • MAV / MAV-slope / Zero-crossing / Slope sign change / Wave length
  • Window = 100 ms long, w/ 60 ms overlap

• Multichannel surface EMG based estimation of bilateral hand kinematics during movements at multiple degrees of freedom (2010)

  • Full-wave rectified, LPF-16Hz

• Simultaneous and proportional control of 2D wrist movements with myoelectric signals (2012)

  • 200mS of POWER/RMS/LOG-VAR. See Spatial Filtering for Robust Myoelectric Control
  • Characteristic
    • POWER: Linear transform of EMG
    • RMS: sqrt(POWER), non-linear
    • LOG-VAR: log(POWER), non-linear
  • Use non-linear transform to linearize relationship between Feature and Joint angle
    • Linear regression performance better w/ RMS, LOG-VAR
    • MLP is better w/ POWER
    • Linear regreesion w/ LOG-VAR has best performance

• Bayesian Filtering of Surface EMG for Accurate Simultaneous and Proportional Prosthetic Control (2014)

  • 140 ms long, w/ 100 ms overlap: MAV v.s. Bayesian filter for sEMG amplitude

• Quantifying Forearm Muscle Activity during Wrist and Finger Movements by Means of Multi-Channel Electromyography (2014)

  • Rectified + LPF-1Hz
  • 112-ch sEMG array -> Non Negative Matrix Factorization (NNMF) for dimension reduction

• EMG-based learning approach for estimating wrist motion (2015)

  • Rectified + LPF-20Hz then normalized

• Continuous Estimation of Wrist Angles for Proportional Control Based on Surface Electromyography (2016)

  • 250 ms RMS

• Continuous estimation of hand's joint angles from sEMG using wavelet-based features and SVR (2016)

  • W: 256ms O: 100ms. A set of feature per window
  • Time-Frequency feature of DWT-sEMG. 5th level 8th order Daubechies. Feature calculted w/ Sub-band: D1-5 & A5
    • MAV
    • SD
    • Autoregressive coefficients
    • Entropy
    • Energy

• EMG-Based Continuous and Simultaneous Estimation of Arm Kinematics in Able-Bodied Individuals and Stroke Survivors (2017)

  • Rectified, LPF-4Hz
  • then downsampled to 20Hz

• Continuous Motion Decoding from EMG Using Independent Component Analysis and Adaptive Model Training(2014)

  • LPF-300Hz, PCA, fastICA
  • MAV W: 15ms O: 5ms

May Be Relevant:

• Neural network committees for finger joint angle estimation from surface EMG signals
• Real-time simultaneous and proportional myoelectric control using intramuscular EMG (2014)
• Estimation of finger joint angles from sEMG using a recurrent neural network with time-delayed input vectors (2009)

Misc

• Electromyogram Whitening for Improved Classification Accuracy in Upper Limb Prosthesis Control(2013)

Reports

• Check 1S2WA_23_CV`
  • W/ nICA_2 as demixing baseline, the accuracy isn't better than RMSDown
• Noted in Meeting 2018/08/30 - S2WA_23

Format source of #4

### Sampling rate @10kHz. FFT are correct up to 5kHz, ADC sample interval seems to be fixed.
---
#### 0.5kHz
![0.5k plot](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/0.5k_plot.png)
![0.5k fft](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/0.5k_fft.png)

#### 1kHz
![1k plot](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/1k_plot.png)
![1k fft](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/1k_fft.png)

#### 3kHz
![3k plot](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/3k_plot.png)
![3k fft](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/3k_fft.png)

#### 4kHz
![4k plot](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/4k_plot.png)
![4k fft](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/4k_fft.png)

#### 5kHz
![5k plot 1](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/5k_plot_1.png)
![5k plot 2](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/5k_plot_2.png)
![5k fft](https://raw.githubusercontent.com/dymnz/sEMG/master/ADC_Arduino/Testing/ADC_FFT_Result/5k_fft.png)

TDSEP — an efficient algorithm for blind separation using time structure

• Subtle Hand Gesture Identification for HCI Using Temporal Decorrelation Source Separation BSS of Surface EMG
  • Performance in gesture rec. TDSEP > FastICA > Infomax > JADE > RawEMG
• Performance comparison of ICA algorithms for Isometric Hand gesture identification using Surface EMG
  • Same as above
• Automatic Classification of Artifactual ICA-Components for Artifact Removal in EEG Signals
• [Comparison of Temporal and Standard Independent Component Analysis (ICA) Algorithms for EEG Analysis]
  • FastICA v.s. Temporal FastICA v.s. TDSEP (best)
  • TDSEP source: http://www.user.tu-berlin.de/aziehe/code/
• Artifact Reduction in Multichannel ECG Recordings Acquired with Textile Electrodes

Spin-off from #42

Components

1. AED
2. LPF+HPF
3. N/I Amp
4. Level shifter

Notes

• All stages should be isolated by jumper
• Resistor value on AED = 6.8k ohm

Relevant

• #6 Hardware component list
• AED circuit
• Additional circuit

Test the applicability with more electrodes

• 6+
  • Check paper for suitable number
    • 6 should be enough, have enough component for 6
    • 6 AED made and tested, making pre-amp
  • Check electrode width for suitable number

• Current: Pot. to avoid Acce. gravity issue
• Question raised by teacher, as handmade == bad

• Pot:
  • Accuracy
  • Consistency
  • Can I prove it is good?
    • Compare it with MPU on pitch
      • Can't test with SUP/PRO
      • Pot. angle magnitude significantly smaller than IMU due to rotating axis issue
        • See 'Meeting 2018/05/16 - GroundTruth/SigPro'
• Is Magnetometer better?

After #44, see table 'paper_sel.xlsx'

• 4-ch sEMG w/ random placement. Try to achieve comparable performance.
• Compare separate FLX/EXT w/ most paper
• Compare separate PRO/SUP w/ paper 7 & 10 on table
  • 7: Bayesian Filtering of Surface EMG for Accurate Simultaneous and Proportional Prosthetic Control (2014)
  • 10: Continuous Estimation of Wrist Angles for Proportional Control Based on Surface Electromyography (2016)

Note

• No need for Magn/Pot. angle groudtruth in this case, since FLX/EXT won't happen w/ PRO/SUP
  • #45
• Separating FLX/EXT and PRO/SUP means that 4-ch sEMG is redundant since 2-ch sEMG is enough to capture the muscle activity of PRO/SUP and FLX/EXT separately
  • What is the rationale behind using 4-ch sEMG v.s. 2-ch sEMG when measuring separate FLX/EXT and PRO/SUP?
    • Accuray
      • Why stop at 4-ch then?

TODO

• Decide procedure

Folder | Template

• Important: ⭐
• Reading 🔥

Papers

• Surface electromyography applications in the sport
  • Effects of rate of force development on EMG amplitude and frequency
• 🔥 The Application of sEMG in Aging: A Mini Review
  • Spectral properties of the surface EMG can characterize/do not provide
    information about motor unit recruitment strategies and muscle fiber type
  • ⭐The extraction of neural strategies from the surface EMG: an update
• Surface Electromyography Signal Processing and Application: A Review

Meeting

• Meeting 1

Folder | Template

• Important: ⭐
• Reading 🔥

Papers

Find out what people do when they got sEMG signal:

• Transradial Amputee Gesture Classification Using an Optimal Number of sEMG Sensors: An Approach Using ICA Clustering
• Principal Component Analysis Applied to Surface Electromyography: A Comprehensive Review

Meeting

• Meeting 2
• Meeting 2-1

Interval between two consecutive ADC reading swing by 6 +- 1 microseconds, with 6 ~ 80% of the time.

• Caused by micros() call?
  • Check the interval between consecutive micros() call, should swing or have random peak based on previous test
• Effect of interval swing?
  • Ask teacher

Example

Relevant: LINPOYU/ultrasound#3

DC offset after inverting amp. should be 0, yet there's DC offset present, which limits the signal amplitude(gain can't be too large or the signal will be cutoff).

• Change the pot. for DC at Inv. Sum. to a larger value
• Change the res. for DC at Inv. Sum. to a larger value

200Hz 10mV in

• Signal measured in CH2, not quite accurate in a range of N mV
• Best output DC: 2.3V / Vpp: 4V
• After HPF: DC: 2mV / Vpp: 30mV
• After Inv amp: DC: -14mV / Vpp: 76mV

0 input

• Signal measured in CH2, not quite accurate in a range of N mV
• Best output DC: 2.34V
• After HPF: DC: 300uV
• After Inv amp: DC: -14mV

Conclusion

• INA128 x1 - Active electrode
• OPA2604 x2 - LPF+HPF, N/I amp
• LMC6482 x1 - LevelShifter

1. Look for snap button (less than NT$50)
2. 洞洞板 x1
3. Resistors xN
4. Capacitor xN
5. Dupont wire head x4 (female)
6. IC坐 x3

https://github.com/dymnz/sEMG/blob/master/Reports/wang/exp3_2017_8_9.MD

To prove the effectiveness of signal processing under different placement condition

• Multiple ring placement test. Equal distance placement with different starting point.
  • 90 degree done
• Follow the same evaluation process as before
  • Run 10 test for 60 segments for each of the 4 gesture
  • Remove outlier for each of the test
  • Plot the box plot of the median error for each gesture
  • Compare the performance w/ and w/o the signal processing

The current PCB is too fragile in

• Wire connection
  • Should have a dedicated connector
• Board itself
  • Thicker board
• Electrode connection
  • Through hole should be a lot bigger

• PCB tested
• Borrow 3D printer
• Design a casing

• -90 <- left | 0 | right -> 90
• +15
  *1 ~ 5 kg

• RNN test
  • Note
  • Vanilla
    • Convolution test
  • LSTM
    • Sensor fusion test
    • Cross-validation w/ patience
• 4-channel power supply PCB
  • Design
  • Missed P+9V power line connection, to be fixed
    • Temp fixed, but still significant voltage drop w/ load. Use PSU instead.
  • Test
    • Voltage drop >10% w/ load, need more test. Use PSU now
• Additional circuit PCB
  • Design
  • Test
• Angle measurement:
  • Acce.
  • Pot. (Unusable, see #43)
  • Magn. (#45)
• Grip force measurement:
  • Note
  • Solder + test
  • Woodwork
  • Calibration
    • Calibration factor : -200000
    • Accuracy: to be calculated