Transcript Slide 1

Data
Smoothing and
Noise Removal
D. Gordon E. Robertson, PhD, FCSB
School of Human Kinetics,
Faculty of Health Sciences
Issues
• All data collected from human subjects has some imbedded
error (noise).
• A signal is the desired information in a waveform.
• Noise is the unwanted part.
• What are the characteristics of your signal?
• What is the noise in your data?
• What factors need to be considered to faithfully record your
signal?
• What is the best technique for reducing the influence of the
noise?
Characteristics of a Signal Waveform
• What is the signal’s amplitude, measured by “dynamic range”?
– dynamic range = maximum – minimum
– obtained from theory, literature, or a pilot study
• Is signal AC or DC?
– average value of an AC signal is zero, e.g., ECG, EMG
– some transducers are AC-coupled and therefore cannot
record DC signals such as force, displacement, temperature,
e.g., piezoelectric transducers
• What is the frequency spectrum (range of frequencies)?
– muscle force is DC to 10 HZ, EMG is 20 to 500 Hz
– do all systems record across this range
– is sampling rate of collection system high enough (>2x) ?
Characteristics of a Measurement
System
• AC or DC coupled? Electronics can compensate, i.e., permit DC
signal with AC transducer or remove DC offsets from AC signals
• Is system capable of amplifying signal without causing
saturation. How do you test if it has correct amplification? Must
never saturate any instrument in the measurement chain. This
cannot be corrected post facto.
• Do all transducers in the measurement chain have correct
frequency spectrum to match signal’s spectrum? If violated
cannot be corrected.
• Where can interference or random noise be introduced and how
can they be reduced or eliminated. I.e., proper grounding and
shielding. Noise may be possible to attenuate.
• Does the system introduce a known or unknown delay?
Measurement System
• Each measurement system is a series of instruments in the
measurement chain.
• E.g., transducer, pre-amplifier, signal conditioner, output
amplifier, A/D converter, input amplifier, storage medium
Calibration
signal
Feedback
Measured
signal
Input
transducer
Input signal
Input power
source
Signal
conditioner
Transduced
signal
Output
transducer
Output signal
Auxillary power
supply
Measurement System
• Example, electromyography (EMG)
transducer (electrodes) signal conditioner
and pre-amp
(receiver), output amp
input amp, A/D
converter, storage
Measurement System
• Example, force transducer
transducer (strain
gauge)
signal conditioner
(bridge amp)
input amp, A/D
converter, storage
What is a Waveform?
• A waveform is any time-varying or spatial-varying series of
related data. Usually collected by A/D converters of computer.
– W(t)
• It can be a known mathematical function (e.g., sine wave)
– W(t) = a sin (2p f t + q)
generalized sine wave with frequency f (cycles per second),
amplitude a (arbitrary units) and phase lag q (radians).
or a series of data sampled at regular or irregular known
intervals.
– W(t) = a(t)
• Can contain both signal (information) and/or noise
What is a Signal?
• A signal is the information carried in a waveform or a physical
quantity that can carry information.
• It is the portion of the waveform that carries the desired
information of the researcher.
• In mathematical waveforms there is typically only signal in the
waveform.
• In data sampled from electronic or other devices there is always
some part of a waveform that is not signal, called noise.
What is Noise?
•
•
•
•
Noise is the part of a waveform that is not signal!
The unwanted errors in a waveform
Noise can be random (white) or have a statistical distribution
Noise can be due to interference from another signal (called
cross-talk) or induced by physical devices near the medium
carrying the waveform (e.g., electric motors, radiation, power
cords, radio waves).
• Noise can occur at random intervals (e.g., bumping of
electrodes, power surges, floor impacts nearby) or regular
(50/60 Hz line interference) or irregular intervals (e.g., ECG or
EEG interference of EMGs).
• Noise may have frequencies inside or outside the frequency
range of the signal (if outside, filtering can effectively reduce the
noise).
Examples of Noise in an EMG signal
heart rate
detected
60 Hz
noise
an impact
spike
What Techniques are Available?
• Moving averages (e.g., Chapman)
• Curve (spline) fitting and interpolation (e.g., Wood & Jennings
1979, Felkel 1951, Woltring 1986)
• Digital filtering (e.g., Winter et al. 1974)
• Fourier reconstruction (e.g., Hatze 1981)
Moving Averages
•
•
•
•
Very simple and easy to implement
Usually have time lags
Unweighted and weighted averaging are possible
Will always attenuate peaks and valleys even if they are
valid
• Need to select a window width (n), usually an odd number
1 n
W(ti )
– MAV(ti) =

n 11
Moving Averages
• rectified EMG (1010 Hz sampling rate)
• averaged EMG (moving average, 51 points)
Curve Fitting and Splines
• If signal has a known mathematical function (e.g., line, parabola,
exponential function) then a “best fit” criterion may be used to
extract the true signal from the waveform
• Piecewise polynomials (splines) may be used to fit curves of
long duration that cannot be fitted by a single function, such as a
polynomial.
• Spline functions do not require equally time intervals in the
waveform and therefore may be used to fit gaps in data files.
• The “Woltring” filter, aka generalized cross-validatory (GCV)
spline filter uses b-splines to filter noise. It behaves similar to a
bidirectional, zero-lag, Butterworth filter.
Curve Fitting and Splines
Vaughan’s
(1982) golf
ball data with
and without
noise (±0.1)
and fitted
polynomial of
order 2.
Digital Filtering
• Used on data that have been sampled with fixed time intervals.
• Types:
– low-pass, high-pass, band-pass, band stop and notch (single
or small band-stop filter, useful for AC interference)
• Designs:
– Butterworth (optimally flat in bandpass), critically-damped,
Chebyshev etc. (sharper cutoffs), Generalized Crossvalidatory (GCV also called Woltring filter, 1986)
• Problems:
– Noise spikes alter a localized period in the signal
– Phase distortion (usually phase-lags occur)
– Does not reduce size of data file
Fourier Reconstruction
• Once a Fourier series has been extracted from a waveform
many cycles may be created
• Requires signal to be cyclic or made cyclic with “windowing”
functions (Hamming, Blackman, Cosine bell, etc.)
• Can reduce a complex signal to a very few number of
coefficients
• Problems:
– Noise spikes can significantly alter overall cyclic pattern
– Can distort signal in unpredictable ways
How to Decide on the Best Technique?
• Visually compare original noisy signal with smoothed signal
(Pezzack et al. 1977). Smooth signal should pass through
middle of the noisy waveform without distorting peaks and
valleys.
• Evaluate smoothing technique against known mathematical
functions (e.g., Robertson & Dowling 2003)
• Evaluate smoothing technique against published data (e.g.,
Wood & Jennings 1979; Hatze 1981; Lanshammar 1982; vs.
Pezzack et al. 1977)
• Evaluate residuals mathematically (e.g., Jackson 1979; Winter
et al. 1984).
• Simulation (Walker 1998, Nagano et al. 2003)
Removing High Frequency Noise
• Essential for data that are to be doubly-differentiated (e.g.,
computing acceleration from displacement data)
• Low-pass filtering is the most common (Winter 1974, Pezzack et
al. 1978)
• Need to select an appropriate cutoff frequency and roll-off (filter
order)
• Critically-damped filters may be better for rapid transients
(Robertson & Dowling 2003)
• Butterworth filters have better roll-offs, i.e., better attenuation in
the “reject band”
• Zero-lag filters can be achieved by filtering forwards and
backwards digitally
Removing Low Frequency Noise
• Essential for doubly-integrating data (e.g., integrating force to
obtain displacement)
• Bias removal is critical
• High-pass filters (Murphy & Robertson 1992)
With bias (pad with means,
• End-point problems need to be considered
zeros, reflexively, e.g., Smith 1989, Walker 1998)
Subtract mean
High-pass filter (10 Hz)
Removing Noise Spikes
• Low-pass filtering may not be effective, perhaps use higher
order, Butterworth filter, or critically-damped filters
• Interpolate across spike or artifact
• Moving median (use smallest window possible)
– Sine wave with spikes
(all single samples
except one)
– All but one removed
(it had double the width)
How to Prevent Phase Distortion
• Centrally weighted moving averages
• Filter in both directions (Winter et al.1974)
• Zero-lag filters (b-splines, Woltring 1986)
Green is original
Cyan is zero-lag
Red has lag
How to Prevent End-point Transients
• Collect extra data before and after critical period
• Padding points
– zeros
– means
Triangle
– reflexive (Smith 1989)
– linear extrapolation (Vint & Hinrichs
1996)
No
padding
• Windowing functions are useful for Fourier analysis
– Cosine bell, Blackman, Hamming, Pad
Hanning,
with triangular
means etc.
Pad reflexively
References
•
•
•
•
•
•
Felkel, E. 0. (1951) Determination of acceleration from displacement-time data.
Prosthetic Devices Research Project, Institute of Engineering Research,
University of California, Berkeley, Series 11, 16.
Hatze, H. (1981) The use of optimally regularised Fourier series for estimating
higher-order derivatives of noisy biomechanical data. Journal of Biomechanics,
14:13-18.
Jackson, K.M. (1979) Fitting of mathematical functions to biomechanical data.
IEEE Transactions on Biomedical Engineering, BME-26(2):122-124.
Lanshammar, H. (1982) On practical evaluation of differentiation techniques for
human gait analysis. Journal of Biomechanics, 15:99-105.
Murphy, S.D. & Robertson, D.G.E. (1992) Construction of a high-pass digital
filter. Proceedings of NACOB II, Chicago, 95-96.
Pezzack, J.C.; Winter, D.A. & Norman, R.W. (1977) An assessment of derivative
determining techniques used for motion analysis. Journal of Biomechanics,
10:377-382.
References cont’d
•
•
•
•
•
•
•
•
Robertson, D.G.E. & Dowling, J.J. (2003) Design and responses of Butterworth
and critically damped digital filters. Journal of Electromyography and
Kinesiology, 13(6):569-573.
Smith, G. (1989) Padding point extrapolation techniques for the Butterworth
digital filter. Journal of Biomechanics, 22:967-971.
Vaughan, C.L. (1982) Smoothing and differentiation of displacement-time data:
An application of splines and digital filtering. Int. J. Bio-Med. Comp. 13:375-396
Vint, P.F. and Hinricks, R.N. (1996) Endpoint error in smoothing and
differentiating raw kinematic data: an evaluation of four popular methods.
Journal of Biomechanics, 26:1637-1642.
Winter, D.A.; Sidwall, H.G and Hobson, D.A. (1974) Measurement and reduction
of noise in kinematics of locomotion. Journal of Biomechanics, 7:157-159.
Woltring, H.J. (1986) A Fortran package for generalized cross-validatory spline
smoothing and differentiation. Advances in Engineering Software, 8:104-113.
Wood G.A. and Jennings, L.S. (1979) On the use of spline functions for data
smoothing. Journal of Biomechanics, 12:477-479.
Wood, G. (1982) Data smoothing and differentiation procedures in
biomechanics. Exercise and Sport Sciences Reviews, 10:308-362.