Sie sind auf Seite 1von 26

Transactions of the Institute of Measurement and Control 22,1 (2000) pp.

327

Intelligent instrumentation
D.W. Clarke
Invensys UTC for Advanced Instrumentation, Department of Engineering Science, Parks Road, Oxford OX1 3PJ, UK

With the falling cost of microelectronics, intelligent instruments are increasingly used in process control. They are designed to integrate into data networks such as Fieldbus, so that remote users can access internal values and recongure their function. It is argued that such instruments should exploit models of the plants data generation to achieve optimal ltering, of their own internal behaviour as veried by self-testing, and of the users needs. A gure-of-merit for sensor signal processing is dened, and some fundamental results indicating the benet of appropriate signal models are deduced. The value of validation, in which diagnosed faults affect the measurement status and its uncertainty, is emphasized. Examples of a self-validating thermocouple, a dissolved-oxygen sensor and a ow-control valve are described together with some guidelines about practical implementation. Key words: actuator validation; sensor validation; signal processing; smart sensors.

1. Introduction The control-theoretic view of a sensor is that it provides the exact value y( t ) of a process variable, as shown in Figure 1(a). Sometimes to provide a degree of realism white noise is added and it is left to a Kalman lter or observer in the control algorithm to estimate the underlying state. In industry there is much more emphasis on measurement technology, for without good sensing, good control is impossible. For process applications, analogue 4/20 mA transmitters were developed for communicating measurements to a remote direct digital control (DDC) computer (see Figure 1b). Such a transmitter has several advantages: the use of current loops allows for long cable runs without signicant interference,

Address for correspondence: David Clarke, Invensys UTC for Advanced Instrumentation, Department of Engineering Science, Parks Road, Oxford OX1 3PJ, UK. E-mail: David.Clarkeeng.ox.ac.uk

2000 The Institute of Measurement and Control

0142-3312(00)TM003OA

4 Intelligent instrumentation

Figure 1 Inside an analogue and a smart transmitter

transmitters can be powered down the wires from the control room, and the standard enables the interworking of transmitters from different vendors. The measurement industry was amongst the rst to adopt microprocessors, leading to so-called smart transmitters (Figure 1c). Whilst for compatibility with existing control practice the 4/20 mA current output is retained, there is usually additional digital communication, either via separate cabling or modulated onto the 4/20 mA loop. With the new data link, the value of more than one process variable can be transmitted together with status information. The many benets provided by microelectronics include: 1) the ability progressively to reduce costs by increasing integration; 2) reduction of the number of noise/uncertainty sources by early conversion to the digital domain; 3) noting that process penetration is expensive, the simultaneous use of trans-

D.W. Clarke

ducers Ti for several process variables; for example T1 could be for differential pressure, T2 for temperature, and the combined signals used for mass-ow computations; 4) internal compensation for nonidealities in the transduction (expensive in analogue systems); 5) more sophisticated signal processing, e.g., adaptive ltering and spectrumbased calculations for condition monitoring; 6) the provision of useful housekeeping duties, including conversion to engineering units, reconguration, self-test, alarms, maintenance of a historical database. Whilst Barney (1985) denes an intelligent instrument (II) simply to be where some further processing . . . is carried out to rene the data, it is perhaps useful to be more prescriptive. For example, the book by Ohba (1992) quotes the denition of Brekenridge and Husson (1978):
The sensor itself has a data processing function and automatic calibration/automatic compensation function, in which the sensor itself detects and eliminates abnormal values or exceptional values. It incorporates an algorithm, which is capable of being altered, and a certain degree of memory function. Further desirable characteristics are that the sensor is coupled to other sensors, adapts to changes in environmental conditions, and has a discrimination function.

At the risk of anthropomorphism we therefore argue that an II should be aware of: (1) the process general ideas of communication theory (Schwartz, 1990; McDonough and Whalen, 1995) indicate that optimum ltering and discrimination require good models of the message. A simple example is that slowly moving signals x( t ) permit the use of narrow-band lters to improve the signal/noise ratio; (2) itself built-in self test (BIST) of the whole process of data collection and transmission should be provided so that the data sent to the user can be validated (Yung and Clarke, 1989; Cassar et al., 1992; Henry and Clarke, 1993; Clarke, 1995; Yang and Clarke, 1999); (3) the user there are several possible destinations of data, as in Figure 2, each having different needs. The use of the network of Figure 2 inspires the vision of the remote use of webbrowsing tools to interrogate the instrument and to transmit Java code for changing its behaviour, say by implanting a new BIST. From BIST, faults can be agged for the maintenance agent; indeed a historical record of the environmental stress suffered by the instrument could be used to predict its expected time to failure (Elsayed, 1996). For many users the new function required of an intelligent instrument is the on-line generation of uncertainty (Kline and McClintock, 1953; Coleman and Steele, 1999). This is a number, in the same units as the measured data y( t ), giving the range of values that the true signal is expected to lie at some probability level, say 95%. The expected uncertainty is required for legal accounting of a metered product being passed to a customer, and is useful for reconciliation of data from

6 Intelligent instrumentation

Figure 2 Typical users of an intelligent instrument

several sources using say an overall plant mass-balance (Crowe et al., 1983; Bagajewicz, 1997). Uncertainty is also helpful in deciding optimal sensor location (Madron and Veverska, 1992) and in sustaining operation during certain detected faults (Ali and Narasimhan, 1993, 1995). The uncertainty of a measurement has two components: random effects due to noise (treatable by conventional ltering ideas), and systematic uncertainty due to bias: a xed error that can be reduced by calibration. Detailed analysis of an instrument is required to deduce its overall uncertainty: see Yang (1993), where it is noted that this uncertainty might vary throughout the sensors operating region. As a simple example, suppose that the measurement process involves the use of some functional from y = f ( u, v ); then the chain rule gives: y = f f u + v u v

Considering u (etc.) to be the uncertainty in u, and assuming independence, it is conventional to write: y2 =

f u

u2 +

f v

v2

It is also conventional to add the two components of uncertainty of an individual 2 variable in quadrature: y2 = y2 bias + ynoise. Hence a full uncertainty treatment

D.W. Clarke

of an II requires a physical model of the instrument (to evaluate bias effects) and knowledge of noise sources and the corresponding data ltering. The design of an intelligent instrument applies a wide range of technologies (Soloman, 1999). Transduction methods based on silicon (Middelhoek and Audet, 1989) or optical bres are increasingly used; low-noise front-ends and digital data transmission adopt ideas from communications practice; optimal ltering exploits control and digital signal processing theories; change- and feature-detection use neural nets (Tarassenko, 1998) or probabilistic concepts such as Bayesian inference (Garthwaite et al., 1995); the implementation is via application-specic integrated circuits often specied in a high-level compilable language; and overall validation brings in uncertainty and reliability engineering. It is argued that many of these ideas are required anyway by the user for optimum exploitation of the measurement. The key reason why they are best integrated within the instrument is purely economic: the many sales that a given II will achieve means that the cost of development is amortized across all users. Hence there is the potential to reduce dramatically the life-cycle cost of its application by providing an intelligent instrument which has the highest achievable performance, allows for exibility in generating data suitable for each destination, and involves BIST and validation for credibility of operation. 2. Performance 2.1 A gure-of-merit As an intelligent instrument processes sampled data within a microprocessor, the simplest model of its behaviour is as shown in Figure 3. It is assumed that the true continuous-time process variable x( t ) is transduced subject to additive noise n*( t ) and that subsequent digital processing is noise-free (i.e., consideration of round-off errors in xed-point implementation is excluded). The analogue data is anti-aliased by a suitable Nyquist lter H( j ) [taken here to be a brickwall lter: H( f ) = 1, f Bn Hz] and sampled at a rate fs Hz. The objective of the signal ( k ) of the plant variable x. processing is to obtain a good estimate x The transduced noise n*( t ) (again for simplicity) is assumed to have a doublesided constant power spectral density nn( f ) = A, f B* Hz. The effect of Nyquist ltering is to reduce the noise bandwidth to Bn, where Bn = fs /2 Hz, as shown

Figure 3 Discrete-time model of sensor processing

8 Intelligent instrumentation

in Figure 4. Hence the effective additive noise on the sampled signal x( k ) is n( k ), where {n} is an uncorrelated random sequence (URS) with variance: 2 n =
2 max

fs /2

fs /2

A df = Afs =

2 maxfs 2B*

and is the variance obtained by rapid sampling with fs = 2B*; sampling at a still faster rate means that the noise n( k ) becomes autocorrelated. Slow sampling, with a correspondingly narrow lter bandwidth Bn, can be used to minimize the effective noise on the data. Suppose that the plant signal x( t ) is constant and that the task of the II is to in a given time T. The standard method is provide the most accurate estimate x to be the average of the m samples acquired, where m = fsT, which has to take x a variance: } = 2 = Var{x 2 2 1 2 n maxfs max = = m 2B* fsT 2TB*

As the data has been averaged over a period T, the effective update rate to the user of the data is fu = 1/ T Hz. Hence consider the following gure-of-merit for the II: ) = F( x fu update rate of x = 2 variance of x

It is seen that F combines both the users sample rate and the accuracy of the received data: a low F implies that fast-changing data cannot be captured with high precision. For the averager, the gure-of-merit becomes: ) = F( x 2B* 1 = 2 A max

Figure 4 Assumed power spectral density of the noise

D.W. Clarke

It is seen that for this simple example, F depends only on the spectral density of the additive noise, reafrming the importance of ensuring minimal noise throughout the analogue chain preceding the analogue/digital converter (ADC). It is noted that there is the usual trade-off between update rate and variance and that the duty of an II is to attain a performance F as close as possible to the ultimate. With only one user, provision of data at a rate fu Hz can, at one extreme, be achieved by a single sample combined with a Nyquist lter of bandwidth fu /2 Hz or, more likely, by the II sampling at some internally-chosen higher rate fs Hz and then averaging. With the assumptions about x( t ) being constant and the noise having a at spectrum, these two approaches have the same value of F.

2.2 Noise with serial correlation A common approximation is that the autocorrelated noise often found in process plant can be modelled by white noise passing through a time-constant Tp, such that the autocorrelation function (ACF) of n( t ) is:
/ Tp Rnn( ) = 2 ne

Taking the process variable x to be a constant, , the sensed output is y( t ) = + n( t ), and an efcient continuous estimator over an experiment time T is: = u 1 T

y( t ) dt

It is shown by Bendat and Piersol (1966) that the variance of this estimate is given by: Var{ } = 1 T

Rnn( ) d T

which for large T and our example process becomes: Var{ } = 2 T

/ Tp 2 d = 22 ne n

Tp T

Hence the ultimate (fast-sampled) gure-of-merit F for this process is 1/(22 n Tp ) and hence is small for sluggish noise. As often an II has limited computing power, it is of interest to deduce how using a non-zero sampling interval h = 1/ fs affects the accuracy of the estimator. The equivalent discrete-time process y( k ) is given by: y( k ) = ( k ) + 1 q1

where = exp( h / Tp ) is the autoregressive parameter, q the forward-shift operator and, to match the variances of y( t ) and y( k ), the variance of the URS ( k ) is 2 2 n(1 ). Using a pre-whitening lter to transform the sequence gives:

10 Intelligent instrumentation

z( k ) = (1 q1 ) y( k ) = ( k ) + (1 ) where the new sequence {z( k )} is N((1 ), 2 n ). Moreover, the sequence is now uncorrelated, so the estimate of the mean of z using N samples has a variance: Var{ z} =
2 2 2 n (1 ) = N N

Hence as = y = z /(1 ) this gives: Var{ } = (1 2 ) 2 1 + 2 n n = (1 )2 N 1 N

Substituting for the parameter from the continuous-time model: Var{ } = 2 n h Tp Tp h = 22 coth x coth x, which approaches 22 as x 0 n n T 2Tp T T

where x = h/2Tp. Hence the sampled gure of merit F is such that Fmax/ F = x coth x as plotted in Figure 5. It is seen that there is not much loss of performance for x in the range 0.5 1, i.e., h = Tp 2Tp. Considering the generation process for n( t ) as white noise passing through a simple lag Hl, the noise bandwidth Bn, dened by Bendat and

Figure 5 Increase in variance due to sampling

D.W. Clarke 11

Piersol (1966) as that of the brickwall lter generating the same mean square value as n( t ), gives: Bn =

Hl( f )2 df /maxHl2 = 1/(4Tp ) Hz

Given noise band-limited to Bn, sampling faster than 2Bn does not affect the variance, so that an appropriate value of h is 2Tp, consistent with the values derived above.

2.3 Tracking a varying signal For control in particular the assumption of a constant x( t ) is not useful. Conven ( k ) is deduced via a Kalman lter of the whole plant, where tionally the estimate x x is now just one component of a state vector. However, in many applications it is not economically feasible to derive an overall dynamic model, but an II can still produce a reasonable estimate provided that a simple signal-generation description is used. For example, a common specication is that there is a typical maximum rate-of-change of the signal x( t ), say, so that a suitable form is Brownian motion: proces variable: x( k + 1) = x( k ) + e( k ) sampled data: y( k ) = x( k ) = n( k ) where Var{e( k )} = 2 e determines the drift rate. Consider the backward-difference ( x( k ) x( k 1))/ h, which from our model is e( k 1)/ h = approximation x fse( k ). Hence we can (loosely) equate with fse. By using the correct Nyquist lter we obtain 2 n = fsA, so that the scaling of variances by sample rate gives e = / fs and n = Afs. In the following, it is assumed that sampling is fast enough so that n e; this is achieved if fs 2/3 / A1/3. As a running example, suppose 2 that = 10 V/s and A = 0.01 V /Hz; then fs 21.5 Hz, and we will select fs = 100 Hz, in which case n = 1 V and e = 0.1 V. An observer (see, e.g., Middleton and Goodwin, 1990) for the above model is: ( k + 1k ) = x ( kk 1) + ( y( k ) x ( kk 1)), x , is where the gain is given by = p /( p2 + 2 n ) and p, giving the variance of x the positive root of the algebraic Riccati equation: p = p + 2 e Hence 2 e + en 2 p2 2 2 p2 p2 e e n = 0 p + 2 n

p=

1+

2 e , where = 4 n

12 Intelligent instrumentation

2 2 e , for small + en 1 + 2 8 e, giving en = en + 2 1 + n 1 1+

en for n h=

The observer can be written in the standard rst-order lter form: ( k + 1) = x ( k ) + (1 )y( k ), where = 1 = x

Such a lter, with xed , is often used for smoothing sampled noisy data; as 1 we obtain a greater effect with smaller output variance. The transfer function, (1 )z1 /(1 z1 ), is the sampled representation of a rst-order unit-gain continuous-time lter with an s-pole at a = ln / h / h = fs for small . The noise bandwidth Be of this lter, using the analysis of the previous subsection, is a /4 = fs /4 Hz. With the running example, = 0.1, p = 0.1 V2 (compared with the accurate value of 0.105), and the lter bandwidth is 2.5 Hz. ) for this observer is: The gure-of-merit F( x F=
3/2 fs fs fs fs = = = p en / f Af A s s

which for the running example comes to 1000 Hz/V2. It is noted, from the assumption concerning the sampling rate fs, that this value of F considerably exceeds the value F = 1/ A ( = 100 Hz/V2 in our example) for the simpler averager; this consequence of the Bayesian assumptions inherent in the formulation is well worth contemplating. A more common model implicit in most instruments is that the signal x( t ) is a sinewave bandlimited by Be, i.e., that the frequency can be anywhere in [0, Be ] Hz. All that can be done is to use a brickwall lter with bandwidth Be, sample at a rate fs = 2Be, and accept that the noise has a variance 2 n = Afs. As no information is passed between samples (unlike the observer where it is assumed that the new value of x( k ) is near the old value), the gure-of-merit is just the standard F = 1/ A much lower than that achievable by using the Brownian motion model. 2.4 Frequency as data Many transducers produce a signal y( t ) whose frequency depends on the process variable x( t ), i.e., y( t ) = a sin ( t ), where ( t ) = 2

[ f0 + x( )] d

A typical example is the Coriolis massow meter, for which the vibration frequency depends on the density of the process uid in the ow tube; another is

D.W. Clarke 13

that of the toothed wheel used for speed and crank-position sensing in car engines. A temperature sensor can be based on a quartz oscillator, as the process temperature affects the crystal dimensions and hence its tuned frequency: as such sensors are straightforward to make and offer a simple interface to digital electronics without the need of an ADC, their use is increasing. If x( t ) is the constant X, the frequency output fm will also be constant at fm = f0 + X Hz. A simple way of measuring fm is as shown in Figure 6: the time between zerocrossings in deduced by the count N of a clock at frequency fc fm. Clearly N = = (2fc / N f0 )/ . There is an implicit uncer2fc / fm, and the estimated value is x tainty of 1 count in the method, as the input signal might cross zero just before the 2fc / N f0. Hence the uncertainty in counter is updated: i.e., 2fc /( N + 1) f0 x is: x = x 2f c f2 2fc 2fc 0 if f f0 N ( N + 1) N2 2fc m

2 2 The uncertainty has a rectangular distribution with variance 2 = f4 0 /48 fc and as the update rate is about 2f0 (counts can be over two successive half-cycles), the 3 gure-of-merit is F = 48 2f2 c / f0, provided there is no other source of uncertainty. This shows the importance of having (a) fc f0, and (b) a high gain factor . Further improvement is attained by reducing the effective value of f0 (the counter is gated by the signal of frequency f0 + x, but it is only the x term that matters). This can be achieved by using synchronous demodulation methods, in which the signal is multiplied by a generated sinewave at frequency f1 Hz, and by suitable ltering the difference signal of frequency f0 + x f1 is passed into the gating circuit, implying that f* 0 = f0 f1 and hence a much large value of F.

3. Validation The computer control of a car engine has a range of self-diagnostic functions, as described in White (1997):
The modern engine management system (EMS) has a self-test capability that regularly examines the signals from the engine sensors and in some instances the actuators. In addition to the self-test capability, the modern EMS normally contains a limp-home

Figure 6 Simple frequency measurement

14 Intelligent instrumentation

facility that is usually termed LOS or Limited Operating Strategy. This means that in the event of a serious fault in one or more of the sensor circuits, and the system perceives that a sensor is operating outside its design parameters, the EMS will substitute a xed default value in place of the defective sensor.

As the engine/EMS design is integrated, it is relatively straightforward to plan for limp-home; the challenge for the engineer is to build similar features into a process control system, hence improving the availability of production. It is argued that limp-home control is assisted by using standard metrics to describe the behaviour of a self-validating (SEVA) instrument. Consider the operation of a Coriolis massow meter deriving its measurement * using an algorithm of the form m *=m o + aT, where m o is the raw calculation m * is: and aT a temperature compensation term. Then the relative uncertainty in m * 2 o 2 m o 2 m m aT 2 T 2 = + * m* m* mo m T

As the correction factor aT is small, the squaring of the terms makes the contributing effect of relative temperature errors on the overall uncertainty smaller still. A fault in the temperature sensor, such as open-circuit, is easily detected by BIST. Without correction, the fault would propagate through the compensation equation to produce a large offset. Having agged the fault, what should the instrument do? It is argued that the correct approach is to use the last good value of T, but to increase the corresponding uncertainty at some rate so that ultimately the uncertainty corresponds to the historically available process temperature range. * t for duty? The fault has changed the meters accuracy Is the new value of m from (say) 0.1% to 1%; this might not be sufcient for legal accounting, but is perhaps enough to enable limp-home control. A SEVA instrument therefore generates a set of standard metrics, as shown in Figure 7. The raw measurement value RMV and its relative uncertainty RU are affected by the diagnostic state resulting from BIST to produce the validated measurement and uncertainty VMV, VU. The MV status has a set of discrete states reecting the measurement validity, for example clear indicates that the RMV is ne and VMV = RMV, whilst blurred means that a noncritical fault has been detected and that the VMV has been compensated as described above. A blind

Figure 7 Generating SEVA data types

D.W. Clarke 15

status indicates device failure: the VMV and VU reect history (VMV = last-good value; VU = historical [minimum maximum]). If a PID regulator is SEVA-aware it can be recongured according to the measurement status, the validated uncertainty VU, and the criticality of the loop. Table 1 shows one possible switching algorithm. Yang and Clarke (1996) discuss other forms of controller response to a non-clear status, including loop retuning. 4. Examples 4.1 A thermocouple example Yang and Clarke (1997) provide a detailed treatment of a validation system for thermocouples, including a full uncertainty analysis. Common faults include open- and short-circuits and loss of thermal contact between the thermocouple and the process. This latter fault is detected by injecting a pulse of current into the thermocouple, inducing ohmic heating, and inspecting the subsequent sensed temperature transient. The idea is that loss of heat transfer is two-way, and good thermal contact will direct away the injected heat and the temperature rise during the test will be small. However, poor thermal contact results in a larger increase in temperature which takes a correspondingly long time to decay back to the surrounding value. In one experiment the sensor was placed in a beaker of water at around 25C. The loss-of-contact fault was simulated by taking the junction out of the water (for t = [15. . .56]s) and the measured temperature drops to around 13C. The initial slow response (after t = 15 s) seen in Figure 8 is regarded as normal since both the rate-of-change and the reported temperature are within their expected operating ranges. The loss-of-contact test was invoked at t = 28 s: Figure 8 shows the rise in electromotive force (EMF) to be signicant and the presence of the fault is exposed. The test was again invoked again at t = 68 s, after the thermocouple has been replaced in the water: the smaller rise in EMF indicates that the fault condition no longer exists. The validated sensor behaviour is shown in Figure 9. During the test the MV Status is set to dazzled and the VU is increased. When the test data is being analysed, the raw data is still affected by the BIST, so MV Status is set to blurred. However live data is available and VU is decreased. After the fault is detected
Table 1 PID reconguration with a SEVA sensor Status clear dazzled blurred blind PID mode normal freeze output detune according to VU manual control

16 Intelligent instrumentation

Figure 8 Loss of contact: thermocouple EMF

Figure 9 Loss of contact: validated sensor behaviour

the MV status remains at blurred. Replacing at the sensing junction in the water t = 56 s causes a large jump in the temperature measurement. This is diagnosed as a spike fault, signalled by the MV Status being set temporarily to dazzled. The MV Status remains at blurred until the next test. Since the subsequent test result at t = 68 s suggests that the fault condition no longer exists, the sensor enters a recovery period and MV Status is nally reset to clear.

D.W. Clarke 17

4.2 Validating a DOx sensor Dissolved oxygen (DOx) sensors are widely used in both industrial and environmental monitoring applications, particularly for indicating the levels of oxygen dissolved in river water. The most common type of sensor is based on the Clark cell (1959), which is an electrochemical cell for generating an electric current by chemical means, i.e., by providing electrons to, and taking them from, molecules and ions in a solution. Figure 10 shows a typical Clark-type DOx sensor, the Foxboro 871DO. The steady measurement current iss, related to the overall rate of oxygen consumption, is given by iss = nAFJ, where n is the number of electrons (four) released per molecule, A is the area of the electrode, F is Faradays constant and J is the molar ux density of oxygen at the electrode surface. There are several possible faults (Fraher and Clarke, 1998) associated with a DOx sensor based on the Clark cell, for example fouling, membrane rupture, and electrolyte depletion. Stagnation, which occurs when sensing a still uid which then becomes locally depleted of oxygen, is a related installation fault. We concentrate here on fouling, where the surface of the membrane becomes coated by foreign agents. This event is quite likely under adverse conditions, and its effect is to render inaccurate the calibration constant G0 of the sensor. Figure 11 indicates fouling, which changes the effective values of thickness b (increase) or diffusivity Dm (reduction) and hence reduces iss for a given value of dissolved oxygen. Moreover, fouling signicantly increases the effective timeconstant of the sensors transient response, which is proportional to = b2 / Dm (typically about 1 min at room temperature). Clearly the steady-state current cannot indicate the fouling level as distinct from a reduction in the dissolved oxygen content, so we apply transient analysis. The basic idea is that before the electrodes are energized the system has reached an equilibrium in which the partial pressures are the same throughout ( pe = pm = pw ), i.e., the total charge of O2 in the electrolyte is independent of the degree of fouling. After switch-on, the existing oxygen molecules in the electrolyte are rapidly removed so that there is a large initial current; i( t ) then drops until equilibrium at iss is reached with steady diffusion through the membrane. This transient current can be used for validation. Ficks second law of diffusion applied to the membrane gives: p 2p = Dm 2 t y with boundary conditions p(0, t ) = 0 and p( b, t ) = pw, and with initial conditions: p( y, 0) = pw. Applying these conditions leads to an expression for the transient current: i( t ) = iss [1 + 2

exp ( n22t / )]

2h Integrating between times t = h and t = 2 h gives I = h i( t ) dt, so:

18 Intelligent instrumentation

Figure 10 Cross-section of the Foxboro 871DO

D.W. Clarke 19

Figure 11 A fouled DOx sensor

I 2 =h 2 iss i.e., I 2 =h 1+ iss lnx

1 2n22h / 2 2 (e en h / ) n2

1 2n2 2 ( x xn ) , where x = exp( 2h / ) 2 n

For a given value of h this function is simply that of x, g( x ) say, which can be precalculated; this is assisted by the rapid convergence of the terms so that only a few are required ( x is typically rather less than 1 for realistic values of h ). Hence from experimental values of I and iss the corresponding x can be determined by table look-up, giving the value of : = b2 2h = Dm lnx

A corrected oxygen reading [O2 ]c w is then deduced from that determined from the calibrated value [O2 ]w using: [O2 ]c w =

[O2 ]w

20 Intelligent instrumentation

where 0 is the value of obtained for the calibrated, clean, membrane. In an experiment two DOx sensors were placed in a stirred water tank and one membrane was deliberately fouled. Figure 12 shows the corrected output of this sensor (dotted line) in comparison with the uncorrected (dashed) value. Also shown are the computed uncertainty bounds (the two solid lines bounding the corrected value) and the output of the parallel clean sensor. Note that uncertainty increases after the detection of fouling, that at all times the corrected and true values are very close, and that the uncertainty bounds are good indicators of the likely accuracy. Hence the DOx sensor has the required features of a SEVA device: 1) detection of the most probable faults; 2) generation of uncertainty even under fault conditions; 3) correction of faulty data to give a credible best estimate of the true measurand. Note in particular the time axis of Figure 12. The SEVA sensor allows for justin-time maintenance: this increase in effective availability is a major benet of true sensor validation.

Figure 12 Typical compensation results

D.W. Clarke 21

4.3 Actuator validation An actuator enables a process variable (such as ow, position or force) to be manipulated according to the demand in a control signal u. The most common actuator is a ow-control valve; valves suffer from nonlinearities (stiction, hysteresis, etc.) that degrade performance. For example, Bialowski (1992) reported that 30% of all loops in Canadian paper mills were oscillating because of valve problems. The intelligent actuator (Isermann and Raab, 1993), depicted in Figure 13, quanties and compensates for its undesirable characteristics, detects and corrects fault conditions, and reports its behaviour to the Next Level Up using standard metrics. Changes in dynamics can then lead to automatic retuning of the loops PI regulator, whilst its current saturation limits can be taken into account by constrained model-predictive controllers. Discussions of such a self-validating or SEVA actuator are in Alsop (1995), Clarke (1995) and Yang and Clarke (1999), where the improvements that can be gained by internal compensation of nonlinearities are demonstrated. As an example of the utility of internal static and dynamic compensation, a careworn 1-inch pneumatic ow-control valve was tested by injecting a staircase current u( t ) into its input I/P converter. The fractional change in the output shaft position X without any compensation is shown in Figure 14. The gure indicates several nonidealities: 1) hysteresis the steady-states differ when X is decreasing compared with their earlier positions; 2) nonlinear dynamics as X approaches 1 pu, the time-constant reduces; 3) velocity limiting when X is falling, its initial slope is roughly constant. Normal practice is to add some form of positioner onto the valve; this improves performance but the valve still exhibits hysteresis. Full compensation involved estimating a nonlinear static model, which was found to exhibit backlash (and so readily handled) and ZieglerNichols tuning of a PI controller using shaft position feedback. The corresponding results are shown in Figure 15: consistent transients throughout the entire range.

Figure 13 Internals of an intelligent actuator

22 Intelligent instrumentation

Figure 14 Valve position without compensation

D.W. Clarke 23

Figure 15 With static compensation and dynamic feedback

24 Intelligent instrumentation

5. Implementation An intelligent instrument can be considered to be a special-purpose control system, in which a set of signals x( t ) is processed to provide the measurement data and a further set u( t ) is used to control the local environment. For instruments such as those based on vibrating transducers, u( t ) is required to sustain operation; in others such as the self-validating thermocouple u( t ) is there for BIST. Hence for the development of II algorithms the appropriate tools are those for blockdiagram simulation and real-time mathematical operation perhaps using a realtime PC as is normal practice in control and signal processing environments. Given a credible set of algorithms for the instrument the next task is to embed the operation into specic hardware. The extent of integration required depends crucially on the market and on the life-cycle prediction of the product, as there are many possibilities. At one (simple) extreme the development environment would generate C-code suitable for a real-time PC module having standard I/O interfaces. The other extreme, valid for large expected markets, is to implement the entire functionality into one or more application-specic integrated circuits (ASICs). More commonly, the manufacturer would have a standard set of integrated modules suitable for data acquisition, processing and communications. It is likely that (for cost and power consumption reasons) the throughput of calculations with these modules restricts the algorithmic complexity allowable. Hence a key task will be to shoe-horn the method such that there is only a minimal loss in performance. One typical approach is to use xed- rather than oatingpoint for arithmetic operations. A C-code algorithm requires a digital processor. Hence if the nal implementation uses an existing processor it may become obsolete and difcult to maintain over its later years of life. It is important, therefore, to consider the next level of implementation: the provision of a hardware description in a high-level language such as VHDL (Rushton, 1995) suitable for compilation into the current level of ASIC technology. The use of a particular language (Handel) for sensor validation architectures is described by Henry (1995) and Henry et al. (1996), where the advantage of eld-programmable gate arrays (FPGAs) for rapid prototyping is discussed. A typical example of the development of an intelligent instrument is that of a digital Coriolis massow meter (Henry et al., 2000). The uid passes along two parallel straight pipe sections (Figure 16) which are vibrating with a differential yawing motion. Voice-coil actuators drive the oscillations and similar sensors measure the relative velocities of the end-points of the pipes. The frequency of vibration provides a uid density measurement, and the small change in phase (up to 4 at full ow) due to the Coriolis acceleration provides the measured mass ow. The existing commercial transmitter uses analogue phase-locked loops (Egan, 1998) for control and measurements, and there is a small digital processor for housekeeping and communications. Whilst providing high accuracy (better than 0.1%) the meter (like others) is prone to stall when subjected to the damping effects of highly aerated uids, and the limitations of the processor means that full validation cannot be applied. The key to good measurements is accurate vibration control; the key to over-

D.W. Clarke 25

Figure 16 Outline of the Coriolis meters ow tubes

coming stalling due to damping is to reduce the amplitude set-point. Hence a nonlinear controller of vibration amplitude was developed (Clarke, 1998) whose response is as fast and accurate as possible over the two-decade range of setpoints. Accurate data acquisition for the measurement function led to the adoption of high-speed sigma-delta converters (the method of choice for II, as then the design of Nyquist lters is easy and there is a useful trade-off between accuracy and sample rate) which deliver 18 bits of accuracy at 44 MHz. The zero-crossing method was found to be insufciently accurate to deduce frequency and phase; instead a Fourier-based approach was adopted which uses data over entire halfcycles. The nal version was found to give a repeatability, in ow-laboratory conditions, having a standard deviation of 0.006%. Moreover, the ability to reduce vibration amplitude means that the meter no longer stalls, despite aeration, and hence can be used in product-transfer operations which involve a batch process starting with an empty tube. The transmitter includes a range of validation functions and generates the appropriate SEVA metrics, including an estimate of uncertainty even during aerated conditions. 6. Conclusions Intelligent instruments are designed with deep knowledge of the measurement process, their own internal workings, and with responsiveness to the needs of the

26 Intelligent instrumentation

variety of potential destinations of data. Their internal parameters can be interrogated over a network, their failure modes are diagnosed via BIST, and there is remote reprogammability of function. Via optimal data processing, their performance can have a high and guaranteed gure-of-merit. In particular the output measured data is made credible by self-validation. These features reduce the lifecycle cost of ownership and offer the prospect of greater plant availability, provided that the process control system can respond to changes in the SEVA metrics broadcast by the validation algorithms. Acknowledgements The support of Invensys plc and of EPSRC grant L/41547 (Rapid implementation of Fieldbus-complaint, standalone, self-validating instruments) is gratefully acknowledged. References
Ali, Y. and Narasimhan, S. 1993: Sensor network design for maximising reliability of linear processes. AIChE Journal 39, 82029. 1995: Redundant sensor network design for linear processes. AIChE Journal 41, 223749. Alsop, P.J. 1995: The validated actuator in process control. PhD thesis. Department of Engineering Science, University of Oxford. Bagajewicz, M.J. 1997: Design and retrot of sensor networks in process plants. AIChE Journal 43, 230006. Barney, G.C. 1985: Intelligent instrumentation. London: Prentice-Hall. Bendat, J.S. and Piersol, A.G. 1966: Measurement and analysis of random data. New York: Wiley. Bialowski, W.L. 1992: Dreams vs reality: a view from both sides of the gap. In: Control Systems 92. BC Canada: Whistler. Brekenridge, R.A. and Husson, C. 1978: Smart sensors in spacecraft: the impact and trends. In: AIAA/NASA Conference on Smart Sensors, Hampton. Cassar, J.P., Bayart, M. and Staroswiecki, M. 1992: Hierarchical data validation in control systems using smart actuators and sensors. In: Intelligent components and instruments for control applications. Malaga, Spain. Clarke, D.W. 1995: Sensor, actuator and loop validation. IEEE Control Systems Magazine 15, 3945. 1998: Nonlinear control of the oscillation amplitude of a Coriolis mass-ow meter. European Journal of Control 4, 196207. Coleman, H.W. and Steele, W.G. 1999: Experimentation and uncertainty analysis for engineers. New York: Wiley. Crowe, C.M., Garcia Campos, Y.A. and Hrymak, A. 1983: Reconciliation of process ow rates by matrix projection. AIChE Journal 29, 88188. Egan, W.F. 1998: Phase-lock basics. New York: Wiley. Elsayed, E.A. 1996: Reliability engineering. Reading, MA: Addison Wesley Longman. Fraher, P.M.A. and Clarke, D.W. 1998: Fault detection and compensation in Clark-type Dox sensors. IEEE Transactions in Instrumentation and Measurement 47, 68691. Garthwaite, P.H., Jolliffe, I.T. and Jones, B. 1995: Statistical inference. Hemel Hempstead: Prentice-Hall. Henry, M.P. 1995: Hardware compilation a new technique for rapid prototyping of digital systems applied to sensor validation. Control Engineering Practice 3, 90724. Henry, M.P. and Clarke, D.W. 1993: The selfvalidating sensor: rationale, denitions and examples. Control Engineering Practice 1, 585610. Henry, M.R., Archer, N., Atia, M.R.A., Bowles, J., Clarke, D.W., Fraher, P.M.A., Page, I., Randall, G. and Yang, J.C.Y. 1996:

D.W. Clarke 27
Programmable hardware architectures for sensor validation. Control Engineering Practice 4, 133953. Henry, M.R., Clarke, D.W., Archer, N., Bowles, J., Leahy, M.J., Vignos, J. and Zhou, F.B. 2000: A self-validating digital Coriolis meter: an overview. Accepted for publication in Control Engineering Practice. Isermann, R. and Raab, U. 1993: Intelligent actuators ways to autonomous actuating systems. Automatica 29, 131531. Kline, S.J. and McClintock, F.A. 1953: Describing uncertainties in single-sample experiments. ASME Mechanical Engineering, 38. Madron, F. and Veverska, V. 1992: Optimal selection of measuring points in complex plants by linear models. AIChE Journal 38, 22736. McDonough, R.N. and Whalen, A.D. 1995: Detection of signals in noise. San Diego, CA: Academic Press. Middelhoek, S. and Audet, S.A. 1989: Silicon sensors. London: Academic Press. Middleton, R.H. and Goodwin, G.C. 1990: Digital control and estimation. Englewood Cliffs, NJ: Prentice-Hall. Ohba, R. 1992: Intelligent sensor technology. Chichester: Wiley. Rushton, A. 1995: VHDL for logic synthesis. Maidenhead: McGraw-Hill. Schwartz, M. 1990: Information transmission, modulation and noise. New York: McGrawHill. Soloman, S. 1999: Sensors handbook. New York: McGraw-Hill. Tarassenko, L. 1998: A guide to neural computing applications. London: Arnold. White, C. 1997: Automotive engine management systems and fuel injection handbook. Sparkford: Haynes. Yang, J.C.Y. 1993: Self-validating sensors. PhD thesis. Department of Engineering Science, University of Oxford. Yang, J.C.Y. and Clarke, D.W. 1996: Control using self-validating sensors. Transactions of the Institute of Measurement and Control 18, 1523. 1997: A self-validating thermocouple. IEEE Transactions in Control Systems Technology 5, 23953. 1999: The self-validating actuator. Control Engineering Practice 7, 24960. Yang, S.K. and Clarke, D.W. 1989: Local sensor validation. Measurement and Control 22, 13241.

Das könnte Ihnen auch gefallen