Professional Documents
Culture Documents
Richard Burdett
Signal Recovery, Wokingham, UK
signal and noise spectra and improving a signal-to-noise the instrumentation techniques used to reduce the remaining
ratio must be done at the expense of the response time noise content. Finally, special considerations involved in
or measurement time (T ); with random white noise inter- recovering pulse signals from photon (light), ion, or elec-
ference, the output signal-to-noise ratio is proportional to tron beams are covered in this section Part 12, Ele-
√
T . The bandwidth reduction technique is best looked ments: F – Signals and Noise, Volume 1.
at from a frequency-domain point of view; signal aver-
aging and correlation techniques lend themselves to time-
domain analysis. 3 NATURE OF NOISE
In this set of articles, – Part 12, Elements: F – Signals
and Noise, Volume 1 – mathematics, and theoretical con- Noise is an undesired signal. All systems have it present
siderations are kept to a minimum and the models presented to some degree. It usually becomes of interest when it
are basic. Practical best practice requires considerably more obscures a desired signal. Figure 1 shows the power spec-
processing to precondition signals than is shown here. For tral density (power/unit bandwidth) of the most commonly
further simplicity, it is assumed that all noise processes are encountered types of noise.
stationary and that both signal and noise are ergodic, analog Deterministic noise can range from simple discrete-
variables; we will not concern ourselves here with digi- frequency components such as power-line hum at har-
tal signals or discrete-time (sampled) signals except where monics of 50 or 60 Hz, to radio frequency interference,
such signals are involved in the enhancement techniques. RFI, caused by narrow, high-energy pulses from power-
They are essential in modern application methods but it is line switching spikes, pulsed lasers, radar transmitters, and
the basic ideas that drive the digital methods. the like.
In addition, only signal recovery techniques will be con- Stochastic or random noise is found in most systems
sidered. Further processing, such as least-squares polyno- both as white noise, where the power spectral density is
mial smoothing of a waveform or Fourier transformation to independent of frequency, and also as 1/f or flicker noise,
obtain a frequency spectrum, are not considered here. where the power spectral density decreases as frequency
Discussion is started by reviewing some basic concepts, increases. Power spectral density is usually measured in
then moving on to ways of avoiding adding noise (e.g. mean-squared-volts/Hz or mean-squared-amperes/Hz; for
hum pickup and preamplifier noise) and finally covering noise, such specifications are usually referred to as spot
Year−1
Power line
Change of classes,
work shifts, etc 50/60 Hz
Lifts, 150/180 Hz
106 Day−1
elevators Switched mode
Power/unit bandwidth (Arbitrary units)
PSUs PC monitors
100/120 Hz
Hour−1
AM
Temperature radio Analog
104 Min−1 TV
Typical RFI
102 frequency
envelope
1/f Noise
1
White Noise
Figure 1. Environmental noise. (Reproduced by permission of SIGNAL RECOVERY, part of AMETEK Advanced Measurement
Technology.)
Signals in the Presence of Noise 3
noise data and usually are a function of frequency. Notice What this model represents is that it is increasingly harder
that for an rms voltage of v (volts) and a frequency range to remove noise as the frequency is lowered below 1 Hz or
of f (Hz), the power spectral density, S, is given by so. This makes practical sense when it is considered that
2 there is a link between noise generation and temperature
v2 v cycling of mechanical systems. It becomes harder to inte-
S= = √ (1)
f (f ) grate temperature cycling noise as the cycle time lengthens.
√
The quantity v/ (f ) is usually referred to as voltage
√
spectral density and is measured in rms volts/ Hz (volts 6 IMPORTANCE OF BANDWIDTH
per root hertz). Similarly, we can refer to current spectral
√
density specifications in units of rms amperes/ Hz. What do we mean by bandwidth? In the simple low-pass
filter circuit shown in Figure 2, for example, we usu-
ally and somewhat arbitrarily define the signal bandwidth
4 WHITE NOISE (Figure 3) to be the cutoff frequency, fc , where vo /vi =
70.7% (−3 dB) or vo2 /vi2 = 50% (the half-power point).
White noise is usually found in one of two forms: Johnson
noise and shot noise. Johnson, or thermal, noise is caused by
random motion of thermally agitated electrons in resistive
7 EQUIVALENT NOISE BANDWIDTH
materials, and the mean-square noise voltage is given by
Notice that frequencies above fc will obviously pass
vn2 = 4kTRf (2)
(although attenuated) through the filter, and therefore are
not really cut off. For noise, it is convenient to think in
where k is Boltzmann’s constant (1.381 × 10−23 JK−1 ),
terms of an equivalent noise bandwidth, Bn , defined by the
T is the absolute temperature (kelvin) and relationship
R is the resistance (ohm). ∞
1
Alternatively, from Ohm’s law, the mean-square noise Bn = 2 |H (jω)|2 df (5)
G 0
current is given by
v 2 4kT f R
in2 = n
= (3)
R R
ni C no
Shot noise is caused by the random arrival of elec-
trons – see Article 185, Photon Counting, Volume 1 –
at, for example, the electrodes of electron tubes or tran-
sistor junctions. A DC current, I , will have a noise-current Figure 2. Low-pass filter circuit. (Reproduced by permission of
component, in , given by SIGNAL RECOVERY, part of AMETEK Advanced Measure-
ment Technology.)
in2 = 2AeI f (4a)
Noise bandwidth, B n
where e is the charge of one electron (≈1.6 × 10−19 C), A
Signal bandwidth, fc
is the mean gain experienced by each electron and I is in
0 Slope = −6 dB/octave
amperes. In many cases, A = 1, so that −3 (−20 dB/decade)
5 FLICKER NOISE
log(f )
fc Bn
Flicker noise has many different origins and is not clearly
understood but exhibits a 1/f n power spectrum with n Figure 3. Low-pass filter transfer characteristic. (Reproduced
usually in the range of 0.9 to 1.35. Note that DC drift is a by permission of SIGNAL RECOVERY, part of AMETEK
very low frequency form of flicker noise. Advanced Measurement Technology.)
4 Elements: B – Signal Conditioning
where H (jω) is the frequency response function of the sys- RELATED ARTICLES
tem and G is a gain parameter suitably chosen to be a
measure of the response of the system to some parameter Article 125, Outline of Purpose of Analog Data
of the signal: for low-pass systems (e.g. Figures 2 and 3), Filters, Volume 1; Article 177, Signals and Signal-to-
G is usually taken to be the zero-frequency or DC gain. noise Ratio, Volume 1; Article 179, Noise Matching and
For band-pass responses, G is usually made equal to the Preamplifier Selection, Volume 1.
maximum gain.
Using the above definition, and taking G to be the zero-
frequency gain (i.e. unity), we can readily calculate that for REFERENCES
the simple RC filter shown in Figure 2 that
Fellgett, P.B. and Usher, M.J. (1980) Fluctuation Phenomena
1 in Instrument Science, Journal of Physics E: Scientific and
Bn = RC Hz (6) Instrumentation, 13, 104–106.
4
Kester, W. (2002) Mixed-signal and DSP Design Techniques,
Noise, of the stochastic form, has been reviewed in Engineering Staff of Analog Devices Inc., Newnes, London.
relation to instrument systems in a classic paper, Fellgett Vainshtein, L.A. (1985) Extraction of Signals from Noise, reprinted
and Usher (1980). from Dover Publications, Wokingham, UK, 1970.