Tài liệu Multisensor thiết bị đo đạc thiết kế 6o (P3) doc

3 ACTIVE FILTER DESIGN WITH NOMINAL ERROR Although electric wave filters have been used for over a century since Marconi’sradio experiments, the identification of stable and ideally ter

3

ACTIVE FILTER DESIGN

WITH NOMINAL ERROR

Although electric wave filters have been used for over a century since Marconi’sradio experiments, the identification of stable and ideally terminated filter net-works has occurred only during the past 35 years Filtering at the lower instru-mentation frequencies has always been a problem with passive filters because the

required L and C values are larger and inductor losses appreciable The

band-lim-iting of measurement signals in instrumentation applications imposes the

addition-al concern of filter error additive to these measurement signaddition-als when accurate nal conditioning is required Consequently, this chapter provides a development oflowpass and bandpass filter characterizations appropriate for measurement signals,and develops filter error analyses for the more frequently required lowpass real-izations

sig-The excellent stability of active filter networks in the dc to 100 kHz tion frequency range makes these circuits especially useful When combined withwell-behaved Bessel or Butterworth filter approximations, nominal error band-limiting functions are realizable Filter error analysis is accordingly developed tooptimize the implementation of these filters for input signal conditioning, aliasingprevention, and output interpolation purposes associated with data conversion sys-tems for dc, sinusoidal, and harmonic signal types A final section develops maxi-mally flat bandpass filters for application in instrumentation systems

Lowpass filters are frequently required to band-limit measurement signals in mentation applications to achieve a frequency-selective function of interest The ap-plication of an arbitrary signal set to a lowpass filter can result in a significant atten-

instru-Copyright © 2002 by John Wiley & Sons, Inc ISBNs: 0-471-20506-0 (Print); 0-471-22155-4 (Electronic)

uation of higher frequency components, thereby defining a stopband whose ary is influenced by the choice of filter cutoff frequency, with the unattenuated fre-quency components defining the filter passband For instrumentation purposes, ap-

bound-proximating the ideal lowpass filter amplitude A( f ) and phase B( f ) responses

described by Figure 3-1 is beneficial in order to achieve signal band-limiting out alteration or the addition of errors to a passband signal of interest In fact, pre-serving the accuracy of measurement signals is of sufficient importance that consid-eration of filter characterizations that correspond to well-behaved functions such asButterworth and Bessel polynomials are especially useful However, an ideal filter

with-is physically unrealizable because practical filters are represented by ratios of nomials that cannot possess the discontinuities required for sharply defined filterboundaries

poly-Figure 3-2 describes the Butterworth lowpass amplitude response A( f ) and ure 3-3 its phase response B( f ), where n denotes the filter order or number of poles.

Fig-Butterworth filters are characterized by a maximally flat amplitude response in the

vicinity of dc, which extends toward its –3 dB cutoff frequency f c as n increases.

This characteristic is defined by equations (3-1) and (3-2) and Table 3-1

Butter-worth attenuation is rapid beyond f cas filter order increases with a slightly ear phase response that provides a good approximation to an ideal lowpass filter

nonlin-An analysis of the error attributable to this approximation is derived in Section 3-3.Figure 3-4 presents the Butterworth highpass response

FIGURE 3-2 Butterworth lowpass amplitude.

FIGURE 3-3 Butterworth lowpass phase.

A( f ) = (3-2)

=

Bessel filters are all-pole filters, like their Butterworth counterparts with anamplitude response described by equations (3-3) and (3-4) and Table 3-2 Bessellowpass filters are characterized by a more linear phase delay extending to their

cutoff frequency f c and beyond as a function of filter order n shown in Figure

3-5 However, this linear-phase property applies only to lowpass filters Unlike the

1ᎏᎏ

兹1苶苶+苶(f/苶f苶c)苶2n苶

b0

ᎏᎏ

兹B苶(s苶)B苶(–苶s)苶

FIGURE 3-4 Butterworth highpass amplitude.

TABLE 3-1 Butterworth Polynomial Coefficients

flat passband of Butterworth lowpass filters, the Bessel passband has no value thatdoes not exhibit amplitude attenuation with a Gaussian amplitude response de-scribed by Figure 3-6 It is also useful to compare the overshoot of Bessel andButterworth filters in Table 3-3, which reveals the Bessel to be much better be-haved for bandlimiting pulse-type instrumentation signals and where phase linear-ity is essential.

3-2 ACTIVE FILTER NETWORKS

In 1955, Sallen and Key of MIT published a description of 18 active filter networksfor the realization of various filter approximations However, a rigorous sensitivityanalysis by Geffe and others disclosed by 1967 that only four of the original net-works exhibited low sensitivity to component drift Of these, the unity-gain andmultiple-feedback networks are of particular value for implementing lowpass and

bandpass filters, respectively, to Q values of 10 Work by others resulted in the sensitivity biquad resonator, which can provide stable Q values to 200, and the sta-

low-ble gyrator band-reject filter These four networks are shown in Figure 3-7 with keysensitivity parameters The sensitivity of a network can be determined, for example,

when the change in its Q for a change in its passive-element values is evaluated Equation (3-5) describes the change in the Q of a network by multiplying the ther-

mal coefficient of the component of interest by its sensitivity coefficient Normally,

50 to 100 ppm/°C components yield good performance

FIGURE 3-6 Bessel lowpass amplitude.

TABLE 3-3 Filter Overshoot Pulse Response

FIGURE 3-7 Recommended active filter networks: (a) unity gain, (b) multiple feedback,

(c) biquad, and (d) gyrator

real-although the biquad network must be used for high Q bandpass filters However, the stability of the biquad at higher Q values depends upon the availability of ade-

quate amplifier loop gain at the filter center frequency Both bandpass networks can

be stagger-tuned for a maximally flat passband response when required The

princi-ple of operation of the gyrator is that a conductance –G gyrates a capacitive current

to an effective inductive current Frequency stability is very good, and a band-rejectfilter notch depth to about – 40 dB is generally available It should be appreciatedthat the principal capability of the active filter network is to synthesize a com-plex–conjugate pole pair This achievement, as described below, permits the real-ization of any mathematically definable filter approximation

Kirchoff’s current law provides that the sum of the currents into any node iszero A nodal analysis of the unity-gain lowpass network yields equations (3-6)

through (3-9) It includes the assumption that current in C2is equal to current in R2;the realization of this requires the use of a low-input-bias-current operational ampli-fier for accurate performance The transfer function is obtained upon substituting

for V xin equation (3-6) its independent expression obtained from equation (3-7).Filter pole positions are defined by equation (3-9) Figure 3-8 shows these nodalequations and the complex-plane pole positions mathematically described by equa-tion (3-9) This second-order network has two denominator roots (two poles) and issometimes referred to as a resonator

= (3-7)Rearranging,

␻1= and ␻2= and ␦= (R1+ R2)

s1,2= –␦兹␻苶1苶␻2苶± j兹␻苶1苶␻2苶·兹1苶苶–苶␦2苶 (3-9)

A recent technique using MOS technology has made possible the realization ofmultipole unity-gain network active filters in total integrated circuit form withoutthe requirement for external components Small-value MOS capacitors are utilizedwith MOS switches in a switched-capacitor circuit for simulating large-value re-sistors under control of a multiphase clock With reference to Figure 3-9 the rate

f s at which the capacitor is toggled determines its charging to V and discharging to V⬘ Consequently, the average current flow I from V to V⬘ defines an equivalent resistor R that would provide the same average current shown by the identity of

C2

ᎏ2

1ᎏ

equation (3-10) The switching rate f sis normally much higher than the signal quencies of interest so that the time sampling of the signal can be ignored in asimplified analysis Filter accuracy is primarily determined by the stability of the

fre-frequency of f s and the accuracy of implementation of the monolithic MOS pacitor ratios

The most important parameter in the selection of operational amplifiers for tive filter service is open-loop gain The ratio of open-loop to closed-loop gain, orloop gain, must be 102or greater for stable and well-behaved performance at thehighest signal frequencies present This is critical in the application of bandpass fil-ters to ensure a realization that accurately follows the design calculations Amplifi-

ac-er input and output impedances are normally sufficiently close to the ideal infiniteinput and zero output values to be inconsequential for impedances in active filternetworks Metal film resistors having a temperature coefficient of 50 ppm/°C arerecommended for active filter design

Selection of capacitor type is the most difficult decision because of many acting factors For most applications, polystyrene capacitors are recommended be-cause of their reliable –120 ppm/°C temperature coefficient and 0.05% capacitanceretrace deviation with temperature cycling Where capacitance values above 0.1 ␮Fare required, however, polycarbonate capacitors are available in values to 1 ␮F with

inter-a ±50 ppm/°C temperinter-ature coefficient inter-and 0.25% retrinter-ace Micinter-a cinter-apinter-acitors inter-are the

V – V⬘ᎏ

I

FIGURE 3-9 Switched capacitor unity-gain network.

most stable devices with ± 50 ppm/°C tempco and 0.1% retrace, but practical pacitance availability is typically only 100 pF to 5000 pF Mylar capacitors areavailable in values to 10 ␮F with 0.3% retrace, but their tempco averages 400ppm/°C.

ca-The choice of resistor and capacitor tolerance determines the accuracy of thefilter implementation such as its cutoff frequency and passband flatness Cost con-siderations normally dictate the choice of 1% tolerance resistors and 2–5% toler-ance capacitors However, it is usual practice to pair larger and smaller capacitorvalues to achieve required filter network values to within 1%, which results in fil-ter parameters accurate to 1 or 2% with low tempco and retrace components.Filter response is typically displaced inversely to passive-component tolerance,such as lowering of cutoff frequency for component values on the high side of

their tolerance band For more critical realizations, such as high-Q bandpass

fil-ters, some provision for adjustment provides flexibility needed for an accurate plementation

im-Table 3-4 provides the capacitor values in farads for unity-gain networks

tabulat-ed according to the number of filter poles Higher-order filters are formtabulat-ed by a cade of the second- and third-order networks shown in Figure 3-10, each of which

cas-is different For example, a sixth-order filter will have six different capacitor valuesand not consist of a cascade of identical two-pole or three-pole networks Figures3-11 and 3-12 illustrate the design procedure with 1 kHz cutoff, two-pole Butter-worth lowpass and highpass filters including the frequency and impedance scalingsteps The three-pole filter design procedure is identical with observation of the ap-

TABLE 3-4 Unity-Gain Network Capacitor Values in Farads

propriate network capacitor locations, but should be driven from a low point impedance such as an operational amplifier A design guide for unity-gain ac-tive filters is summarized in the following steps:

driving-1 Select an appropriate filter approximation and number of poles required toprovide the necessary response from the curves of Figures 3-2 through 3-6

2 Choose the filter network appropriate for the required realization from Figure3-10 and perform the necessary component frequency and impedance scal-ing

3 Implement the filter components by selecting 1% standard-value resistorsand then pairing a larger and smaller capacitor to realize each capacitor value

to within 1%

FIGURE 3-10 Two- and three-pole unity-gain networks.

FIGURE 3-11 Butterworth unity-gain lowpass filter example.

Component values from Table 3-4 are normalized to

1 rad/s with resistors taken

at 1 ⍀ and capacitors in farads.

3-3 FILTER ERROR ANALYSIS

Requirements for signal band-limiting in data acquisition and conversion systemsinclude signal quality upgrading by signal conditioning circuits, aliasing preventionassociated with sampled-data operations, and intersample error smoothing in outputsignal reconstruction The accuracy, stability, and efficiency of lowpass active filternetworks satisfy most of these requirements with the realization of filter character-

FIGURE 3-12 Butterworth unity-gain highpass filter example.

Component values from Table 3-4 are normalized

to 1 rad/s with capacitors taken at 1 F and resistors the inverse capacitor values from the table in ohms.

istics appropriate for specific applications However, when a filter is superimposed

on a signal of interest, filter gain and phase deviations from the ideal result in a nal amplitude error that constitutes component error It is therefore useful to evalu-ate filter parameters in order to select filter functions appropriate for signals of in-terest It will be shown that applying this approach results in a nominal filter erroradded to the total system error budget Since dc, sinusoidal, and harmonic signalsare encountered in practice, analysis is performed for these signal types to identifyoptimum filter parameters for achieving minimum error

sig-Both dc and sinusoidal signals exhibit a single spectral term Filter gain error isthus the primary source of error because single line spectra are unaffected by filterphase nonlinearities Figure 3-13 describes the passband gain deviation, with refer-ence to 0 Hz and expressed as average percent error of full scale, for three lowpass

filters The filter error attributable to gain deviation [1.0 – A( f )] is shown to be

min-FIGURE 3-13 Plot of filter errors for dc and sinusoidal signals as a function of passband.

imum for the Butterworth characteristic, which is an expected result considering thepassband flatness provided by Butterworth filters Of significance is that small filtercomponent errors can be achieved by restricting signal spectral occupancy to a frac-tion of the filter cutoff frequency.

Table 3-5 presents a tabulation of the example filters evaluated with dc and soidal signals defining mean amplitude errors for signal bandwidth occupancy to

sinu-specified filter passband fractions of the cutoff frequency f c Equation (3-11) vides an approximate mean error evaluation for RC, Bessel, and Butterworth filtercharacteristics Most applications are better served by the three-pole Butterworthfilter, which offers a component error of 0苶.苶l苶%FS for signal passband occupancy to40% of the filter cutoff, plus good stopband attenuation While it may appear ineffi-cient not to utilize a filter passband up to its cutoff frequency, the total bandwidthsacrificed is usually small Higher filter orders may also be evaluated when greaterstopband attenuation is of interest, with substitution of their amplitude response

pro-A( f ) in equation (3-11).

␧

苶%苶F苶S苶 = BW/fc冱0 [1.0 – A( f )] · 100% (dc and sinusoidal signals) (3-11)The consequence of nonlinear phase delay with harmonic signals is described byFigure 3-14 The application of a harmonic signal just within the passband of a six-pole Butterworth filter provides the distorted output waveform shown The varia-tion in time delay between signal components at their specific frequencies results in

a signal time displacement and the amplitude alteration described This time tion is apparent from evaluation of equation (3-12), where linear phase provides aconstant time delay A comprehensive method for evaluating passband filter error

One-Pole Three-Pole Three-Pole One-Pole Three-Pole Three-Pole

RC Bessel Butterworth RC Bessel Butterworth0.05 0.999 0.999 1.000 0.1 0.1 00.1 0.997 0.998 1.000 0.3 0.2 00.2 0.985 0.988 1.000 0.9 0.7 00.3 0.958 0.972 1.000 1.9 1.4 00.4 0.928 0.951 0.998 3.3 2.3 0.10.5 0.894 0.924 0.992 4.7 3.3 0.20.6 0.857 0.891 0.977 6.3 4.6 0.70.7 0.819 0.852 0.946 8.0 6.0 1.40.8 0.781 0.808 0.890 9.7 7.7 2.60.9 0.743 0.760 0.808 11.5 9.5 4.41.0 0.707 0.707 0.707 13.3 11.1 6.9

BW

ᎏ

f c