Blind Signal Separation

1
Blind Signal Separation

• Mike Wininger
• Introduction to Prosthetics
• Rutgers BME
• 8december2004

2
What is Blind Signal Separation (BSS)?

• In Abstract
• Recovers signals (sources detected by surface
  EMG) of which only observations are available.
• Blind because in vivo, the electrical signal is
  obscured by confounding factors.

• In Application
• Complex muscle movements involve combinations of
  contractions/generate combined signals.
• Applications to volition control in prosthetics
  and site-specific diagnostics.

3
EMG detection and theory

• EMG employs surface electrodes to detect
  electrical signal in local muscles.
• Problems of clean signal detection
• Interference from fat, skin, surface products
• Cross-talk (neighboring sources generate
  detectable signals)
• Signal degradation is non-linear

Figure In Farina, et.al, the Pronator Teres and
Flexor Carpi Radialis were isolated for signal
detection.
4
Basic Principles
The observations as a function of the sources, as
shown on left. Aim of BSS to find the mixing
matrix, A, so as to pull out the sources, as
defined as



xt Ast nt xt observations A mixing
matrix st sources nt noise matrix
st Âxt Cst  nt
 Estimate of A  Moore-Penrose Inverse of
 C BSS Indeterminacies matrix
5
Experiment Model

• Step 1 Simulation Model
• Volume conductor is a nonhomogenous anisotropic
  medium. (muscle anisotropic fat isotropic,
  3mm skin isotropic, 1mm).

• Two muscles, 3 electrodes. 10mm inter-electrode
  distance, electrodes 1x5mm. Electrodes at ends,
  1 between muscles.
• Sensors placed between the innervation zone and
  tendon region,
• Crosstalk (over muscles) scaled by inverse linear
  relationship

Figure Placement of sensor and sources
6
Model

• Simulation parameters
• Signal sources intracellular action potentials,
  travel with 4m/s mean velocity.
• Finite-length muscle fibers (65mm).
• Uniform fiber distribution (range 50-800)
  innervated by same motorneuron.
• Motor unit is recruited when force developed by
  muscle exceeds 15 Maximal output.
• Protocol
• Two cases simulated corresponding to two muscles
  active during short intervals of time.

7
Experiment Subject Setup

• Two muscles considered Flexor Carpi Radialis,
  Pronator Teres.
• Selective control for isolated movements
• FCR Wrist Flexion, PT Wrist rotation

• Cycle of contractions
• 3s flexion (50 MVC), 1s rest
• 3s rotation (50 MVC), 1s rest
• 100s survey time
• Forces recorded separately, displayed on
  oscilloscope for subject feedback (thus, it was
  known a priori in which intervals of time 1
  muscle was active, and the other inactive)

Figure Placement of sensors on subject
8
Results

• Simulation Results
• Choi-Williams chosen as preferred filter (over
  Wigner-Ville).
• Single-auto term criterion yielded highest
  correlative values
• Results improved when 3 muscles/5 sensors employed

Figure Results of simulation.
9
Results

• Experimental signals
• After BSS, the cross-correlation between
  reference source and reconstruction
• FCR .98 .02
• PT .94 .07
• Mean Frequency (N16) reference
• FCR 101.4 21.8 Hz
• PT 85.6 15.3 Hz
• Mean Frequency (N16) after sepn
• FCR 104.4 20.4 Hz
• PT 84.9 14.1 Hz

Figure Results of subject study. Notice the
recognition of Source 2 in Signal 1 and
vice/versa.
10
Headlines
Correlation coefficient and similar trends in
mean frequency indicate that the hypothesis of
linear instantaneous mixtures was met in the
experimental recordings. (Reference/Reconstructed
Correlation coefficients exceeded 0.9) A fair
indicator of performance is the ability to reduce
the influence of second source from recordings
over the two muscles
11
Discussion

• Apparent Benefits
• BSS well-suited for separation of surface EMG
  generated by different muscles
• Possible to separate closely situated muscles
• Promising applications for small parallel muscles

• Limitations
• Attenuation of second source
• Linear instantaneous mixtures is only an
  approximation (limited to linear mixtures)
• Filter-dependent (signal amplitude much smaller
  with Laplacian vs. Gaussian)

12
New York Marathon 2004