Next Article in Journal
Previous Article in Journal

Sensors 2008, 8(2), 897-909;

Full Paper
Networked Estimation with an Area-Triggered Transmission Method
Vinh Hao Nguyen and Young Soo Suh *
Department of Electrical Engineering, University of Ulsan, Namgu, Ulsan 680-749, Korea E-mail: vinhhao@hcmut.edu.vn. E-mail: suh@ieee.org
*
Author to whom correspondence should be addressed. Email: suh@ieee.org
Received: 18 January 2008 / Accepted: 7 February 2008 / Published: 15 February 2008

Abstract

: This paper is concerned with the networked estimation problem in which sensor data are transmitted over the network. In the event-driven sampling scheme known as level-crossing or send-on-delta, sensor data are transmitted to the estimator node if the difference between the current sensor value and the last transmitted one is greater than a given threshold. The event-driven sampling generally requires less transmission than the time-driven one. However, the transmission rate of the send-on-delta method becomes large when the sensor noise is large since sensor data variation becomes large due to the sensor noise. Motivated by this issue, we propose another event-driven sampling method called area-triggered in which sensor data are sent only when the integral of differences between the current sensor value and the last transmitted one is greater than a given threshold. Through theoretical analysis and simulation results, we show that in the certain cases the proposed method not only reduces data transmission rate but also improves estimation performance in comparison with the conventional event-driven method.
Keywords:
Networked estimation; event-driven; level-crossing; send-on-delta

1. Introduction

The traditional way to transmit sensor data in networked control systems is to sample sensor signals equidistant in time. This approach is widely used since analysis and design are simple and inherited from well-established control system theories. If the network speed is high and the traffic is sparse, the periodic sampling approach has many merits because network-induced delay is small and can be ignored. But when the network bandwidth is limited due to executing tasks of several nodes, time delay becomes large and randomly varying. To avoid these problems, the sensor data transmission rate should be reduced.

Recent works have discussed event-driven alternatives to traditional time-triggered sampling scheme. It has been shown to be more efficient than time-triggered one in some situations, especially in network bandwidth improvement. For instance in [1], the authors provided a comparison of time-triggered impulse control and level-triggered one where the level-triggered scheme gave lower average error for the same average rate of impulse. In [2, 3], an adjustable deadband was defined on each node to reduce network traffic. The node does not broadcast a new message if its signal is within the deadband. A level-crossing sampling scheme in association with an 1-bit coding and decoding strategy to reduce data size was introduced in [4]. This coding sampling scheme becomes highly efficient under data-rate constraints since the nodes only transmit one bit, 0 or 1, per sample. In [5], an optimal level-triggered sampling design problem was proposed in order to minimize the distortion of a filter over a finite horizon. A similar sampling scheme named send-on-delta (SOD) transmission was explored in [6-8]. By adjusting the threshold value at each sensor node, data transmission rate is reduced in order for the network bandwidth to increase and can be used for other traffic. Although the event-driven schemes mentioned above have several names, the basic principle of sampling is the same. That is, in framework of networked control system, a sensor node transmits data only if the difference Δ between the current sensor value and the last transmitted one is greater than a given threshold. Hereafter, for ease in representation we refer to them as SOD sampling.

One major issue of SOD sampling method as pointed out in [9] is that it does not detect signal oscillations or steady-state error if the difference Δ remains within the threshold range during a long time. Furthermore, in highly noisy systems, the sensor value could exceed the threshold not because the true value exceeds the threshold but because sensor noise exceeds the threshold. In that case, there could be many unnecessary sensor data transmissions.

Motivated by the perspective results of the event-based integral sampling in [9] and the networked estimation problem using SOD transmission method in [7], we propose a novel sampling scheme called area-triggered or send-on-area (SOA) in which sensor data are sent only when the integral of Δ is greater than a given threshold. The proposed SOA sampling scheme is slightly different from [9] to improve bandwidth network in case of large noise. Then, a networked estimator based on Kalman filter is formulated to estimate states of the system periodically even when the sensor nodes do not transmit data. Through theoretical analysis and simulation results, we show that in the certain cases the proposed method gives better estimation performance than the SOD method in [7].

2. Area-triggered sampling scheme

Consider a networked control system described by the linear continuous-time model:

x ˙ ( t ) = A x ( t ) + B u ( t ) + w ( t ) y ( t ) = C x ( t ) + v ( t )
where xRn is the state of the plant, u is the deterministic input signal, yRp is the measurement output which is sent to the estimator node by the sensor nodes. w(t) is the process noise with covariance Q, and v(t) is the measurement noise with covariance R. We assume that w(t) and v(t) are uncorrelated, zero mean white Gaussian random processes.

The area-triggered sampling scheme illustrated in Fig.1b is stated as follows:

  • Let ylast,i (1 ≤ ip) be the last transmitted value of the i-th sensor output at instant tlast,i. The area formed by the current output yi(t) and ylast,i along time:

    Π i ( t , t last , i ) t last , i t ( y i ( t ) y last , i ) d t .

  • A new sensor value will be sent to the estimator node if the following condition is satisfied:

    | Π i ( t , t last , i ) | > α i
    where αi is the given threshold value at the i-th sensor node.

Note that in the SOD sampling scheme [7] shown in Fig.1a, a new sensor value will be sent if |Δi(t, tlast,i)| > δi, where δi is the given threshold value.

2.1. Effect of noise on sensor data transmission rate

In this section, we assume u(t) = 0 to investigate effect of noise on the system. The difference between the current sensor value and the last transmitted one is defined by:

Δ i ( t , t last , i ) y i ( t ) y last , i = C i [ Φ ( t , t last , i ) I ] x ( t last , i ) + C i t last , i t Φ ( t , r ) w ( r ) d r + ν i ( t ) ν i ( t last , i )
where Φ(t, tlast,i) = exp(A(ttlast,i)). From (2) we have:
Π i ( t , t last , i ) = t last , i t Δ i ( t , t last , i ) d r .

Inserting (4) into (5), we obtain (6):

Π i ( t , t last , i ) = t last , i t Δ i ( t , t last , i ) d r = t last , i t C i [ Φ ( t , t last , i ) I ] x ( t last , i ) d r + t last , i t C i t last , i u Φ ( t , r ) w ( r ) drdu + t last , i t [ ν i ( t ) ν i ( t last , i ) ] d r = t last , i t C i [ Φ ( t , t last , i ) I ] x ( t last , i ) d r + t last , i t C i [ r t Φ ( u , r ) d u ] w ( r ) d r + t last , i t ν i ( t ) d r t last , i t ν i ( t last , i ) d r .

Assuming x(tlast,i) = 0, we can derive the variance of Πi(t, tlast,i) as follows

E { Π i Π i } = C i t last , i t θ ( t , r ) w ( r ) w ( r ) θ ( t , r ) d r C i + t last , i t 2 R ( i , i ) d r = C i t last , i t θ ( t , r ) Q θ ( t , r ) d r C i + 2 R ( i , i ) ( t t last , i )
where R(i,i) is the (i,i) -th element of R and θ(t,r) is defined by
θ ( t , r ) r t Φ ( u , r ) d u .

Note that the variance of Δi(t, tlast,i) in the SOD sampling scheme [7] is given by

E { Δ i Δ i } = C i t last , i t Φ ( t , r ) Q Φ ( t , r ) d r C i + 2 R ( i , i ) .

From (7) and (8) we see that both EiΠi′} and EiΔi} consists of two terms. The first term depends on the process noise w(t) and the second one depends on the measurement noise v(t). Therefore, we believe EiΠi′} and (in the case of SOA) plays the same role as EiΔi} (in the case of SOD) in investigation of data transmission rate. This assumption will be verified in the next section 2.3. In the analysis in [7], we know that data transmission rate is proportional to EiΔi}. In the highly noisy systems, large R value makes EiΔi} in (8) large. It leads to increase data transmission rate when applying the SOD sampling scheme. But with the SOA sampling scheme, we can constrain the effect of R on EiΠi′} in (7) by lowering the factor (ttlast,i). This situation can be achieved by choosing the αi threshold value such that (ttlast,i) ≪ 1s. In other words, if the time interval of two consecutive sensor data packets is less than one second, then the SOA method can be applied. This is not a so strict condition in networked control systems.

2.2. Πi computation and SOA sampling in discrete time at the sensor nodes

Let T be sampling time of the signals yi(t), yk,i(k = 1,2,3,…) be the k-th sampled value yi(kT) from time the i-th sensor node transmits ylast,i. If the sensor output has no noise, yk,i lies on the smooth curve Cix(t). But under the effect of measurement noise, yk,i will be in the vicinity of Cix(t) as illustrated in Fig.2. Supposing that condition (3) is satisfied at t = t3,i, Πi(t, tlast,i) shown approximately by the slashed area is calculated:

Π i = t last , i t 1 , i ( y i ( t ) y last , i ) d t + t 1 , i t 2 , i ( y i ( t ) y last , i ) d t + t 2 , i t 3 , i ( y i ( t ) y last , i ) d t ( y 1 , i y last , i ) T / 2 + ( y 1 , i + y 2 , i 2 y last , i ) T / 2 + ( y 2 , i + y 3 , i 2 y last , i ) T / 2

The principle of computing the slashed area in (9) is to divide it into several small parts. Each part is a trapezoid with the height T and two sides (yk,iylast,i), (yk+1,iylast,i) except for the first part which is a triangle. However, using (9) is inconvenient if condition (3) is not satisfied for a long time interval because the sensor node has to spend much memory on storing yk,i. An algorithm to calculate Πi and transmit sensor data in which we only use 3 memory units to store (Πi, ylast,i, y0,i) at every period T is proposed as follows:

Π i = 0 ; k = 0 ; y 0 , i = y last , i ; while | Π i | < α i k = k + 1 Π i = Π i + ( y k , i + y 0 , i 2 y last , i ) T / 2 y 0 , i = y k , i end while y last , i = y k , i ; Transmit y last , i

2.3. Effect of noise on signal distortion

We define a performance index called squared error distortion Di for each sensor node:

D i k = 1 N ( y r , i ( k T ) y last , i ( k T ) ) 2
where yr,i(kT) = Cix(kT) is the true output value of the i-th sensor without measurement noise and ylast,i(kT) is the i-th sensor value received at the estimator node.

It is very difficult to derive an explicit expression of signal distortion in two methods SOD and SOA by theoretical analysis because it depends not only on the given threshold but also the system model and noise. Thus, in this section we evaluate the effect of measurement noise on distortion by considering an example instead.

Consider two sensor outputs y(t) = 0.1t + v(t) and y(t) = 5(1 − e−0.1t) + v(t), where R is varying. We will see how data transmission rate and distortion vary as R varies in both methods SOD and SOA. The evaluation is implemented with T = 0.01s in 50 seconds. The threshold values δ and α are given in 3 cases:

  • Case 1. δ = 0.005: small threshold

  • Case 2. δ = 0.01: medium threshold

  • Case 3. δ = 0.05: large threshold

In the above example, α is chosen according to δ such that number of data transmissions in two methods is identical as v(t) = 0 (without noise). From the results in the Fig. 3 and Fig. 4, the effect of R on data transmission rate and distortion is summarized as follows:

-

When the threshold value is small as in case 1, distortion of SOA is equivalent to that of SOD but data transmission rate of SOA is smaller than that of SOD. In this case, using SOA will reduce data transmission rate.

-

When the threshold value increases as in case 2, distortion of SOA is slightly smaller than that of SOD and data transmission rate of SOA is much smaller than that of SOD. In other words, we get benefit of data transmission rate reduction and a little distortion reduction from SOA method.

-

When the threshold value is large as in case 3, SOA reduces not only data transmission rate but also distortion. The larger R value is, the more reduction is. This result is remarkable! It means that in the SOA method, signal distortion is not degraded along with the reduction of data transmission rate even when system noise is large. Therefore, we hope that estimation performance will be significantly improved when applying a filter.

3. State estimation with SOA Transmission method

The networked estimation problem applying SOA transmission method is depicted as follows:

  • Measurement outputs yi (1 ≤ ip) are sampled at the period T but their data are only transmitted to the estimator node when (3) is satisfied.

  • For simplicity in the problem formulation, transmission delay from the sensor nodes to the estimator node is ignored.

  • The estimator node estimates states of the plant regularly at the period T regardless of whether or not sensor data arrive. If there is no i-th sensor data received for t > tlast,i, the estimator node considers that the measurement value of the i-th sensor output yi(t) is still equal to ylast,i but the measurement noise increases from vi(t) to vn,i(t) = vi(t) + Δi(t, tlast,i).

To formulate a state estimation problem, the bound of Δi(t, tlast,i) needs to be determined. In the next section, we will compute the bound of Δi(t, tlast,i) and then a modified Kalman filter is applied for state estimation.

3.1. Bound of Δi (t, tlast,i)

At time t = t1,i, if the estimator node does not receive i-th sensor data as illustrated in Fig.2, from (3) and (9) we know that:

| Π i ( t 1 , i , t last , i ) | = | ( y 1 , i y last , i ) T / 2 | < α i ,
therefore
2 α i / T < ( y 1 , i y last , i ) < 2 α i / T .

At time t = t2,i, if the estimator node does not receive i-th sensor data yet then:

| ( y 1 , i y last , i ) T / 2 + ( y 1 , i + y 2 , i 2 y last , i ) T / 2 | < α i ,
or
2 α i / T 2 ( y 1 , i y last , i ) < ( y 2 , i y last , i ) < 2 α i / T 2 ( y 1 , i y last , i ) .

Inserting (13) into (15) we obtain:

6 α i / T < ( y 2 , i y last , i ) < 6 α i / T .

Computing similarly for instants t = tk,i(k = 3,4,…), we always have the following inequality:

6 α i / T < ( y k , i y last , i ) < 6 α i / T .

In other words, the bound of Δi(t, tlast,i) is given by

| Δ i ( t , t last , i ) | < 6 α i / T .

Note that Δi(t, tlast,i) = 0 at time the estimator node receives i-th sensor data. Then, the process is repeated as (12)-(17) as long as the estimator node does not receive data from the i-th sensor node yet.

3.2. State estimation

Assuming that Δi(t, tlast,i) has a uniform distribution with (18), variance of Δi(t, tlast,i) will be (6αi/T)2 /3. Thus, if there is no i-th sensor data received for t > tlast,i, variance of measurement noise is increased from R(i, i) to R(i, i) + (6αi/T)2 /3.

A modified Kalman filter for state estimation k at step k, where there is a change in the measurement update part of the discrete Kalman filter algorithm [10], is given as in the Fig.5. We use the discretized system model sampled at period T :

A d = e A T , B d = 0 T e A r Bdr ,

Qd is the process noise covariance of the discretized system:

Q d = 0 T e A r Q e A r d r ,
and ylast is the vector of p last received sensor values:
y last = [ y last , 1 y last , 2 y last , p ] .

In the modified Kalman filter in Fig.5, the states of the plant are estimated regularly at every period T regardless of whether or not sensor data arrive. If i-th sensor data arrive, Δi(t,tlast,i) = 0, the modified Kalman filter acts like the conventional Kalman filter. Otherwise, it uses ylast,i as the measurement value and (i,i) = R(i,i) + (6αi/ T)2/ 3 as measurement noise covariance for state estimation.

4. Simulation

To verify the proposed filter, we consider an example of the second-order system with step input where the output is sampled by the SOD and SOA methods:

x ˙ ( t ) = [ 0 1 1 / a b / a ] x ( t ) + [ 0 M / a ] u ( t ) + w ( t ) y ( t ) = [ 1 0 ] x ( t ) + u ( t ) Q = 0.01 , R = 0.01 , T = 10 m s

The system parameters are given in the following two cases for performance evaluation:

  • Case 1. (underdampled system) M = 30, a = 5, b = 1.

  • Case 2. (undamped system) M = 30, a = 5, b = 0.

The simulation process is implemented for 50 seconds. In each case, we use two methods (SOD and SOA) for performance comparison. Estimation performance is evaluated by the average distortion:

D ¯ i = 1 N k = 1 N e k , i 2 = 1 N k = 1 N ( x k , i x ^ k , i ) 2
where ei (i = 1,2) is the estimation error, xi is the reference state, and N = 5000.

The simulation results with different threshold values for the two cases are shown in Table 1 and Table 2, where α value is chosen according to δ value such that n (number of sensor data transmissions) is identical in the two methods. In both cases, we see that when δ is small (i.e. δ = 0.1, 0.3), estimation performance of SOD and SOA is almost the same. Rigorously speaking, SOD is slightly better than SOA. However, when δ is increasing, SOA method shows to be outperform significantly. For example as δ = 0.9, SOA outperforms SOD by the 1 reduction of 5 times (i.e. 0.0116 vs. 0.0536 in case 1, and 0.0017 vs. 0.0081 in case 2). As illustrated in Fig. 6 and Fig. 7, the estimation error of SOA is much smaller than that of SOD.

Through simulation results, we see that the SOA method provides better estimation performance than the SOD method. The key reason is that in the SOD method, noise-containing sensor data not only make transmission rate increase but also degrade estimation performance. Whereas, thanks to the integral block which acts as a noise filter in the SOA sampling scheme, the sensor node does not send highly noisy data, but it transmits data only when the output indeed changes value. For that reason, the transmitted sensor data in the SOA method is more reliable. This helps estimation performance better.

5. Conclusion

In this paper, the state estimation problem with SOA transmission method over network has been considered. We have shown that when increasing the threshold value to improve bandwidth network, the SOA method gives much better estimation performance than the SOD method. This is very useful in the realistic applications where sensor data transmission rate needs to be lowered due to joining of many sensor nodes or for power saving in wireless networks. The SOA method has been also proven to be more efficient than the SOD method in the highly noisy systems.

This work was supported by the Korea Science and Engineering Foundation (KOSEF) grant funded by the Korea government (MOST) (No. R01-2006-000-11334-0). The authors also would like to thank Ministry of Commerce, Industry and Energy and Ulsan Metropolitan City which partly supported this research through the Network-based Automation Research Center (NARC) at University of Ulsan.

References and Notes

  1. Astrom, K.J.; Bernhardsson, B.M. Comparison of Riemann and Lebesgue sampling for first order stochastic systems. Proceedings 41st Conference on Decision and Control, Las Vegas, U.S.A. 2002; pp. 2011–2016. [Google Scholar]
  2. Otanez, P.G.; Moyne, J.R.; Tilbury, D.M. Using deadbands to reduce communications in networked control systems. Proceedings of the American Control Conference, Anchorage, U.S.A. 2002; pp. 3015–3020. [Google Scholar]
  3. Sandra, H.; Peter, H.; Eckehard, S.; Martin, B. Network Traffic Reduction in Haptic Telepresence Systems by Deadband Control. In Proceedings IFAC World Congress; International Federation of Automatic Control: Prague, Czech Republic, 2005. [Google Scholar]
  4. Kofman, E.; Braslavsky, J.H. Level Crossing Sampling in Feedback Stabilization under Data-Rate Constraints. Proceedings 45th Conference on Decision and Control, San Diego, U.S.A. 2006; pp. 4423–4428. [Google Scholar]
  5. Rabi, M.; Moustakides, G.V.; Baras, J.S. Multiple Sampling for Estimation on a Finite Horizon. Proceedings 45th Conference on Decision and Control, San Diego, U.S.A. 2006; pp. 1351–1357. [Google Scholar]
  6. Mikowicz, M. Send-On-Delta Concept: An Event-Based Data Reporting Strategy. Sensors 2006, 65, 49–63. [Google Scholar]
  7. Suh, Y.S.; Nguyen, V.H.; Ro, Y.S. Modified Kalman filter for networked monitoring systems employing a send-on-delta method. Automatica 2007, 43(2), 332–338. [Google Scholar]
  8. Suh, Y.S. Send-On-Delta sensor data transmission with a linear predictor. Sensors 2007, 7, 537–547. [Google Scholar]
  9. Mikowicz, M. Asymptotic Effectiveness of the Event-Based Sampling According to the Integral Criterion. Sensors 2007, 7, 16–37. [Google Scholar]
  10. Brown, R.G.; Hwang, P.Y.C. Introduction to Random Signals and Applied Kalman Filtering; John Wiley & Sons: New York, 1997. [Google Scholar]
  11. Astrom, K. J. Introduction to Stochastic Control Theory; Academic Press: New York, 1970. [Google Scholar]
Sensors 08 00897f1 1024
Figure 1. SOD and SOA sampling schemes

Click here to enlarge figure

Figure 1. SOD and SOA sampling schemes
Sensors 08 00897f1 1024
Sensors 08 00897f2 1024
Figure 2. Sensor output with noise in discrete time.

Click here to enlarge figure

Figure 2. Sensor output with noise in discrete time.
Sensors 08 00897f2 1024
Sensors 08 00897f3 1024
Figure 3. Effect of R on data transmission rate and distortion for y(t) = 0.1t + v(t).

Click here to enlarge figure

Figure 3. Effect of R on data transmission rate and distortion for y(t) = 0.1t + v(t).
Sensors 08 00897f3 1024
Sensors 08 00897f4 1024
Figure 4. Effect of R on data transmission rate and distortion for y(t) = 5(1 − e−0.1t) + v(t).

Click here to enlarge figure

Figure 4. Effect of R on data transmission rate and distortion for y(t) = 5(1 − e−0.1t) + v(t).
Sensors 08 00897f4 1024
Sensors 08 00897f5 1024
Figure 5. Structure of the modified Kalman filter.

Click here to enlarge figure

Figure 5. Structure of the modified Kalman filter.
Sensors 08 00897f5 1024
Sensors 08 00897f6 1024
Figure 6. Estimation error as δ = 0.9, α = 0.0881 in case 1

Click here to enlarge figure

Figure 6. Estimation error as δ = 0.9, α = 0.0881 in case 1
Sensors 08 00897f6 1024
Sensors 08 00897f7 1024
Figure 7. Estimation error as δ = 0.9, α = 0.0303 in case 2.

Click here to enlarge figure

Figure 7. Estimation error as δ = 0.9, α = 0.0303 in case 2.
Sensors 08 00897f7 1024
Table Table 1. Estimation performance of 2 methods with different threshold values in case 1

Click here to display table

Table 1. Estimation performance of 2 methods with different threshold values in case 1
δ0.10.30.50.70.9
α0.00060.00810.03020.05340.0881
n2465490240175140
1 by SOD6.08e-40.00260.01010.02680.0536
1 by SOA6.40e-40.00300.00610.01030.0116
2 by SOD0.00800.01060.01760.02050.0313
2 by SOA0.00820.01100.01390.01440.0148
Table Table 2. Estimation performance of 2 methods with different threshold values in case 2

Click here to display table

Table 2. Estimation performance of 2 methods with different threshold values in case 2
δ0.10.30.50.70.9
α0.00050.00380.00940.01810.0303
n334716011094822658
1 by SOD3.49e-48.94e-40.00220.00430.0081
1 by SOA3.65e-48.32e-40.00150.00160.0017
2 by SOD0.00520.00670.00870.01090.0131
2 by SOA0.00530.00700.00790.00800.0080
Sensors EISSN 1424-8220 Published by MDPI AG, Basel, Switzerland RSS E-Mail Table of Contents Alert