Wearable EMG System with Signal Compression and Decompression - PowerPoint Presentation

Slide1

Wearable EMG System with Signal Compression and Decompression

Related reading: Effective

Low-Power Wearable

Wireless Surface EMG

Sensor Design

Based on Analog-Compressed Sensing,

Balouchestani

& Krishnan

(

2014

).

Sensors

14: 24305-24328.

W. Rose 201704011

Background

“

sEMG

signals exhibit

good level of sparsity in the time and frequency

domains.”

“Conventional data

acquisition approaches

rely

on the Shannon sampling

theorem, which

says

a signal must be sampled at least twice its bandwidth in order to be represented without error

.”

Sparse signal = a signal with most values zero or lacking information

Drawbacks of conventional approach

Generates

huge intolerable

number of

samples for many applications with a large bandwidth.

Even

for low signal

bandwidths, including

some biomedical signals,

this produces

a large

number

of redundant digital samples.

Cure

Use compressive sampling to reduce

the number of acquired samples by utilizing sparsity

.

Specifically, use analog compressive sampling before the analog-to-digital conversion step

No compression

Digitize, then compress (more common than #3)

Compress, then digitize

Balouchestani

& Krishnan (2014)

Question for the Authors

Balouchestani

& Krishnan (2014) say they apply compressive sampling to the

analog

signal before A-to-D conversion. (See bottom part of figure in previous slide.) But they implement their compression algorithm in computer code (C, hSpice

, or Matlab

). And their sample signals are from online databases. The fact that they implement in code (not with a circuit) and that their test data is already digitized means the ADC has already happened.

How do they account for this discrepancy?

How did they get it published?

L-1 norm

The L1 norm is a way of measuring the “length of a vector” by the sum of the absolute values of its components.

L-2 norm

The L2 norm is the Euclidean norm, in which we measure the length by the square

root of the

sum of the squares of the components.

L

-

norm

The L-



norm means measuring the length of a vector by the length of it longest component.

L-n norm

The L-n norm is a way of measuring the “length of a vector” by the n

th

root of the sum of the n

th power of each element.

BSBL

Block-sparse Bayesian Learning

Data Compression Examples

No

compression: standard audio

CD, .wav, .

aiff

Lossy

compression:

MP3, AAC Decompressed signal is not perfect, but very close

Typical compression factor = 4 to 20

Lossless compressed audio

formats

Decompressed signal is perfect copy

Typical compression

factor = 2

Codec Software or hardware to compress and decompress

Transmitter and Receiver Design

Balouchestani

& Krishnan (2014)

Eq

. 7

in

Balouchestani

& Krishnan (2014

) should be*

Sensitivity = % of “positives” that are correctly identified.

Specificity

= % of

“negatives

” that are correctly identified

.

T

P

= true positive;

F

N

= false negative, etc.* B.&K. 2014, eq.7, says:

Sens. = T

P

/ (T

P

+TN) (wrong)

Step

7

of Table 2 in

Balouchestani

& Krishnan (2014) says sparsity level is given by

Sp

= (N/N-K)

w

hich is an unclear or wrong choice of parentheses, and K is undefined, and it is inconsistent with later figures, which show that Sparsity is a percentage that is 100% or less – which will not be true if computed with the equation above.

Fig. 10 does not make sense. It indicates that the accuracy is highest when sampling rate is lowest, just 10% of the Nyquist rate.