Delta-Modulation Coding Redesign for Feedback-Controlled Systems

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/29603714 Delta-Modulation Coding Redesign for Feedback-Controlled Systems ArticleinIEEE Transactions on Industrial Electronics · July 2009 DOI: 10.1109/TIE.2009.2020079·Source: OAI CITATIONS 34 READS 355 3 authors, including: Fabio Gómez-Estern Universidad Loyola Andalucía 64 PUBLICATIONS1,302 CITATIONS SEE PROFILE Francisco R. Rubio Universidad de Sevilla 123 PUBLICATIONS3,077 CITATIONS SEE PROFILE All content following this page was uploaded by Francisco R. Rubio on 12 March 2015. The user has requested enhancement of the downloaded file. 1 Delta-Modulation Coding Redesign for Feedback Controlled Systems Carlos Canudas-de-Wit, Fabio G´omez-Estern, and Francisco R. Rubio Abstract—This paper investigates the closed–loop properties of the differential coding scheme known as Delta–Modulation (∆- M) when used in feedback loops within the context of feedback controlled systems. We propose a new modified scheme of the original form of the ∆-Malgorithm which is better suited for applications where the sensed information is used in feedback. A state feedback controller is implemented with the state estimated by a predictor–based differential decoder. Stability of the resulting closed–loop systems (controller–coder–decoder) are studied. These properties (stability and performance) depend on the quantization parameter ∆, which is assumed constant in the first part of our work. In a further step, parameter ∆is made adaptive, by defining an adaptation law exclusively in terms of information available at both the transmitter and receiver side. With this approach both stability and performance is improved. Index Terms—Differential coding, Delta modulation, Networked controlled systems, NCS, Wireless Sensor Networks. I. INTRODUCTION DELTA modulation (∆-M) is a well-known differential coding technique used for reducing the data rate required for voice communication, see [20]. The standard technique is based on synchronizing a state predictor on emitter and receiver and just sending a one–bit error signal corresponding to the innovation of the sampled data with respect to the predictor. The prediction is then updated by adding a positive or negative quantity (determined by the bit that has been transmitted) of absolute value ∆, a known parameter shared between emitter and receiver. This paper studies different aspects of the use of differential coding as a means for transmitting sensing signals in feedback controlled linear systems. In particular we focus on the stability issues that appear when the ∆-modulation scheme is embedded in a control system between the device measuring the state and the control signal. The problem is of interest in the area of Networked Controlled Systems (NCS) calling for data-compression algorithms, but also in digital control applications where sensed information is converted using analog-to-digital converters (ADC) with few bits (i.e. 1-bit ADC). A. Delta-modulation as a coding mechanism in NCS Networked Control Systems is a multidisciplinary field where Control and Communications technologies and Information theory meet. Whereas the advantages of using networks C. Canudas-de-Wit is with CNRS, GIPSA-Lab, UMR-5528, Grenoble, FRANCE. Email: [email protected]. Fabio G´omez-Estern, Francisco R. Rubio and Jos´e Forn´es are with the Department of Automatic Control and Systems Engineering at the University of Sevilla, Spain. Email: [email protected],[email protected], [email protected], as a means of connecting plants and actuators is easily understood, the set of problems arisen by the technological convergence is complex. In one sense, stability of controllers may be compromised by the inclusion of a network, and new closed–loop analysis must be done, with specific quantization– modulation considerations, (see, for instance, [7] and the references therein). Another approach is to monitor the network quality of service (QoS) parameters, to consequentially tune the controller gains for maximum stable performance, as is presented in [22]–[23]. Finally, co–design approach (see [18]) addresses simultaneously the modification of the control strategy and the message transmission schedule for best overall results. Applications concerned with real-time wireless networked controlled systems, as shown in Figure 1, seek for datacompression algorithms aiming at reducing the amount of information that may be transmitted throughout the communication channel, and therefore permitting a better resource allocation and/or an improvement of the permissible closed–loop system bandwidth (data-rate). Moreover, recent advances in this field are strongly focused on energy consumption, which is probably the key to cheap distributed sensing. Vendors of technologies like ZigBee and Bluetooth now claim that devices may be powered on the same battery for years, leaving a way to non replaceable battery devices into the market. Energy consumption is thus a goal together with bandwidth when dealing with modern sensor networks, and new communications technologies. The search for minimal data transfer devices in control has also been motivated by underwater applications, where data rates are normally bounded below 100 b.p.s. due to the strongly dissipative medium. differential encoder differential decoder Communication Channel Adaptive mechanism Adaptive mechanism xk+1 =Axk+Buk uk Kˆxk δk δk ˆxk xk Fig. 1. Block diagram of the problem set up studied in this paper. The Delta modulation (∆-M) algorithm can also be understood as the coarsest two-level (1-bit) quantizer. Optimization of the quantization levels is mandatory in large-scale systems, and may be of great interest while designing low cost transmitter/receiver components. Differential coding in feedback system is thus a simple alternative to reported control schemes hal-00394981, version 1 - 14 Jun 2009 Author manuscript, published in "IEEE Transactions on Industrial Electronics 56, 7 (2009) 1-20" 2 concerning the use of quantizers in the context of NCS, i.e. [12], [7], [17], [14], [27], [16]. More precisely, the use of delta–modulators for networked control has been reported in [15], as a practical implementation without theoretical stability analysis. The modified form of the (∆-M) algorithm proposed here, has the ability of enforce the separation principle between the control law and the estimation process (coding strategy). As a result, the (linear) feedback law is first designed disregarding the coding algorithm (x≡ˆx). Then, the coding scheme is designed in order to preserve stability when embedded in the feedback loop. It is also worth to mention that this modified form of the ∆-Mcoding structure explicitly contains information about the model plant and the controller feedback gain. Others works leading with some variants of the Delta modulation coding in feedback systems can be found in [4] for energy-aware coding, [3] for the combination of Delta modulation with entropy variable length coding, [11] for the use of more than one bit, and [5] for multivariable coding in the fixed quantization case. B. Delta-modulation as 1-bit converter The other possible scenario where our results will be of interest, concerns embedded control systems where the sensors and digital controller are embedded in the same electronics (system on chip, SoC). With the aim of saving energy, cost, and space in the silicon layers, low-resolution analog- to-digital converters are preferable to high-resolution ones. Numerous examples of the use of Delta–modulators embedded in general microelectronic devices have been reported in the literature. For instance, [1] and [24] use FPGAs and ASICs for implementing Delta and Sigma–Delta–modulation A/D converters for current and voltage measurement. The need for oversampling the analog signals, is compensated by the technological advantage of reducing the number of ADC channels and comparators. Interesting applications using Delta-modulation inside control loops are also found in gyroscopic and acceleration sensors in MEMS, which are configured as shown in Fig. 2. The principle, as described in [13]–[19], is based on feedback– sensing the capacitive mass displacement produced by an acceleration on the device. In [13] and [9], this analogous mass displacement is converted to a digital signal by using 1- bit Sigma–Delta modulation encoders. Then, the 1-bit signal is decoded and used in feedback to “cancel” the external acceleration that disturbs the small mass motion. If the control is successfully designed, then the control signal equals the external acceleration providing the desired measure. Note that this example is similar to the framework displayed in Figure 1 with the difference that the encoder signals are transmitted by wire. C. Main paper contribution The paper contributions are presented in the following order. First we study the case of the Delta-modulation algorithm under fixed-gain ∆. To begin with, we address the continuoustime formulation case, which is relevant because it captures the limiting case of the discrete-time formulation (finite data rate transmission), and provides a better understanding of the maximal stability properties that can be attained. It also shows the substantial simplification that can be reached in the stability analysis, as well as the separation properties between states and estimation obtained by using the proposed ∆-Mmodified form. In this part, we study the closed–loop stability properties of the original form of the ∆-Malgorithm. Then, we propose a new coding law that improves over the original form of the ∆-Malgorithm, and it allows for the separation principle mentioned previously. Based in this modified form, we then study the discretetime case, which permits a more clear assessment of the data rate constraints. It is shown that the system states converge to a ball enclosing the origin, and that the result is semi–global. The size of the attraction region and the system precision, depends on the coding gain ∆linearly, and is also a function of the location of the unstable discrete–time open–loop poles. These results show that the stability properties improve as the sampling time is reduced, or equivalently as the transmission data rate is increased. Finally, the last part of the paper illustrates an adaptation mechanism, consisting of making time variant the quantization interval ∆. In our adaptive approach, closed–loop global asymptotic stability is proved for the noiseless case. The effect of random noise in the performance of the adaptive quantization scheme is also analyzed via numerical simulations. D. Comparison with other approaches The problem of quantization with time–varying resolution in feedback loops has been addressed in [10] and [2]. The first of these works presents, in the case of fixed resolution, a scheme similar to [6], in the sense that the state estimation is computed trough a filter built upon the closed–loop system matrix. However, the extension to variable–scale quantization is only defined in the zooming–in direction, and hence the initial states are upper bounded by the initial choice of the zoom factor. As a consequence, only semi–global stabilization is achieved. Here we propose a ∆–update law that works well for both in–and–out zooming directions, thus providing a means to capture unbounded initial states in the zooming–out stage and guaranteeing global asymptotic stability. Moreover, by defining an explicit ∆–update law in both directions, if the state is driven temporarily out of the domain of attraction at any time due to unattended disturbances, the system will recover stability, unlike the case of [10]. On the other hand, the work of [2] also guarantees global asymptotic stability with time–varying full–state quantization in two subsequent stages. Although our approach can be roughly understood as a particular case of the Theorems provided there, an important new feature introduced here is a state predictor that reduces the amount of data transmitted per sample by only sending the quantized prediction error. This reduces the data–rate to a minimum of 1 bit per sample, while the data–rate in Theorem 3 of [2] is log2(2M), a value that cannot be taken arbitrarily low. Mis actually defined as a function of the matrices A,B,Kof the control system, and hal-00394981, version 1 - 14 Jun 2009 3 Capacitor plate displacement m·a(t) Capacitance to Digital conversion (1-bit) Digital controller 1-bit DAC converter Acceleration a(t) Electrostatic feedback force Digital output Measured Acceleration + + Fig. 2. Block diagram of MEMS accelerometer using ∆-Mboth in the inner analog capacitance–to–digital conversion and in the outer force–feedback loop. it must be large enough to make the (thereby defined) scale factor small. In Theorem 4 of [2] a 1-bit per sample data–rate transmission scheme is also discussed, but here the separation principle is not present, as the feedback u=H(q(x)) is no longer a linear feedback of the state estimated on the receiver side. But probably the most significant difference between that approach and the one presented here is the zooming factor. In [2], it is calculated in terms of the convergence of the state to an attractive ellipsoid, whose geometry depends in a complex way on the system matrices. In our approach the zoom factor is updated at all times with a simple law that only depends on the last two state estimations, irrespective of the system matrices (and hence not subject to the side effects of bad identification). E. Definitions and notation Let x∈Rn,A∈Rn×n. The following definitions and notation are in order: •||x||2=Pn i=1 x2 i •|x|=Pn i=1 |xi|, where |x| ≥ ||x|| •sgn(x) = [sgn(x1),sgn(x2),··· ,sgn(xn)]T, with sgn(0) = ±1arbitrarily1, and •||A|| is the induced Euclidean norm of A. F. Control setup The control setup under study is shown in Figure 3, which describes a closed–loop system under one-way communication channel. The control computations are assumed to be done at the plant side, whereas the sensor is remotely located. Information from sensor to controller is then transmitted through a data communication network, and coded by the differential encoder/decoder maps ΦE:x7→ δ, and ΦD:δ7→ ˆx, where δ(t)is the encoded signal vector of dimension n, with only two-valued elements δi(t)∈ {−1,1},∀n= 1,2...,n, and ˆx(t)∈Rnis the estimated value of xas obtained from the decoding map ΦE. This architecture is commonly found in the NCS literature where a common actuator is used for controlling a system with distributed sensors (probably a large number of low cost wireless devices). Some practical applications where its use is clearly justified are: •Energy–efficiency control in buildings. In such applications, it is quite frequent to distribute a set of low–power temperature and humidity sensors all over the premises, in order to gather distributed information and use it together 1When our encoding algorithm requires that value, we will allow it to issue a +1 or -1 arbitrarily. differential encoder differential decoder Communication Channel ˙x=Ax +Bu u Kˆx δ ˆx x Fig. 3. Block diagram of the continuous-time version of the problem. with heat flow or energy consumption models to design optimal control strategies. In those cases, the actuators are normally few and large in size (air conditioning units, window openers...) and they are wired to the controller (when not directly built–in). •Another application where a great research effort is being done is in underwater vehicle control. In a scenario where a set of underwater explorer robots are remotely controlled and coordinated by a surface vessel, the limitations on data rate appear exclusively on the upward link, as the surface ship enjoys a significantly larger source of energy for transmitting. G. Assumptions The hypotheses used all along the paper, are the following: •The coding strategy is assumed to be scalar, i.e. each component xi kof the vector signal xkis coded independently of each other. Therefore, the encoding (respectively decoding ΦD) process is defined as a ndimensional map, ΦE= [Φ1 E,Φ2 E,...Φn E]T, with elements Φi E:xi7→ δi, •The encoded system outputs are binary, i.e. δi k∈ {−1,1} •The maximum bit rate per unit of time is R[b.p.s.], •Transmission and ADC delays are neglected. Aspects related to transmission delays have been studied elsewhere, see for instance [26], and [25]. II. CONTINUOUS-TIME FORMULATION To simplify matters, we first consider in this section a fully continuous-time formulation. The more realistic discretetime case, including data-rate constraints, will be studied in subsequent sections. hal-00394981, version 1 - 14 Jun 2009 4 +- R R ∆ ∆ δ δ ˆx ˆx ˜xx × × Encoder Decoder Fig. 4. Block diagram of standard continuous-time version of the ∆- Mcoding scheme. The encoder ΦEis depicted at the left, where the decoder ΦDis shown at the right. For n > 1,δis defined as δ= [δ1, δ2, . . . , δn]T, with δi∈ −1,1. We consider linear systems, with a linear feedback of the following form: ˙x(t) = Ax(t) + Bu(t)(1a) u(t) = −Kˆx(t)(1b) where x(t)∈Rn,A∈Rn×n,B∈Rn×mare controllable matrices, u(t)is the input vector. A. Differential coding There exists a large class of source coding algorithms aiming at compressing information for a more efficient data transmission. Differential coding schemes belong to the temporal waveform coding class of algorithms. In differential coding, differences between successive samples are encoded rather than the samples themselves. Since the difference between samples is expected to be smaller than the actual sample amplitude, fewer bits are required to represent the differences. Delta modulation (∆-M), shown in Figure 4, is the simplest form of differential coding (see [20]), in which a two-level (1-bit) quantizer is used in conjunction with a first-order predictor. The continuous-time version of the ∆-Mcoding algorithm is: •Encoder map ΦE:x7→ δ ˙ ˆx(t) = ∆(t)·δ(t),ˆx(0) = 0 (2a) δ(t) = sgn(x(t)−ˆx(t)) .(2b) where sgn(ζ) = [sgn(ζ1),sgn(ζ2),··· ,sgn(ζn)]T, •Transmitted information: δi(t)∈ {−1,1}, •Decoder map ΦD:δ7→ ˆx ˆx(t) = Zt 0 ∆(τ)·δ(τ)dτ, ˆx(0) = 0.(3) ∆(t)∈Rn×nis the step size matrix, which in general is chosen to be constant (time-varying, or adaptive gains can also be considered, as shown later in this paper). For designing simplicity, ∆(t)is frequently chosen to be diagonal. A mandatory choice is that the gain ∆(t)must be strictly the same on both sides of the coding process, in order for the decoder to work properly. To simplify notation, time-arguments will be dropped in this section in all variables when its explicit mention is not necessary. B. Coding separation principle The complete coding process can be seen as an estimation process defined by the coding map ΦC:x7→ ˆx, for the problem of state feedback stabilization In the case of reliable noiseless channel transmission ΦCis simply defined as ΦC= ΦE◦ΦD. Behind the separation principle that will be advocated here, lies the idea of first designing the feedback gain K disregarding the coding algorithm; assuming that x≡ˆx, and then designing the coding map ΦCin such a way to preserve stability when ΦCis embedded in the feedback loop. In fact, the feedback law design is separated from the design of the coding algorithm, while the converse is not true, as it will be demonstrated later. With this in mind, we first consider here that the linear feedback gain Kis designed so that the matrix Ac=A−BK is strictly stable under the controllability assumption of the pair (A, B). Then, we will see that some modifications on the standard ∆-Malgorithm are in order, for this separation principle to be possible. Before introducing the modified form of the ∆-Malgorithm, we first discuss the stability properties achievable with the standard form. C. Standard ∆-Mcoding algorithm in feedback Consider the problem of finding a suitable value for ∆so as to stabilize system (1), under the coding law (2)-(3). It is easy to see that the error equations read as: ˙x=Acx+BK˜x(4a) ˙ ˜x=Acx+BK ˜x |{z } ˙x −∆sgn(˜x)(4b) with ˜x=x−ˆx, and sgn(˜x)as defined in Section I-E. A first observation is that this system is fully coupled, and that the control and coding law can not be designed independently of each other. In particular we note that the gain ∆should, somehow, be chosen large enough to locally dominate the rate of change of the system state, ˙x. Clearly, this can be done only under relatively “high-values” for ∆. Putting aside technicalities in the sense of existence of solutions (i.e. Filippov’s type, etc.) due to the discontinuous right-hand side of Equation (4), which are not central for the discussion here, the computation details given in Appendix VIII-A, show that if ∆is constant, and of the form ∆ = ∆0·In×n,∆0>0 with ∆0being a scalar, one can select a sufficiently large ∆0, such as to ensure that the state and error trajectories ζ= (xT,˜xT)Ttend to zero. That is, ∆0must satisfy a condition of the type: ∆0≥c· ||ζ(0)|| where c > 0is a constant depending on the system and controller parameters. hal-00394981, version 1 - 14 Jun 2009 5 +- + + ++ R R∆ ∆ δ=±1δˆx ˆx ˜xx × × Ac Ac Encoder Decoder Fig. 5. Block diagram of standard continuous–time version of the ∆-Mcoder with the proposed modification. D. A new structure for the ∆-Mcoding The previous discussion (summarized in Proposition 5) shows that relatively large gains for ∆0will be required to stabilize the system, as the estimation of the bound on ∆0 would tend to be conservative, and the size of the region of attraction is suited to be large. Besides, in this continuous–time framework, high gains will result in an important “chattering” effect, thus increasing the estimation error variance when the algorithm is discretized, or adding noise to the measured signal. Although the analysis is conservative, and probably lower values ∆0may still be possible, it is interesting to study new structures of the ∆-Mcoding such as to reduce the necessary gain for stabilization. The algorithm presented next is a modification of the original form of the ∆-Malgorithm, ad most importantly, it yields an estimation error equation decoupled from the system state x. Consider now the following modified ∆-Malgorithm ˙ ˆx(t) = Acˆx+ ∆(t)·δ(t)(5) together with system (1), and δ(t)as defined before. Note that a new term Acˆxhas been included in this case. As a consequence, the coding algorithm depends explicitly on the system model and the feedback gain used. This equation describes both: the encoder and the decoder, as shown in the block diagram of Figure 5. It can be shown that the new error equation resulting from this new structure is described as two systems in cascade interconnection, i.e. ˙x=Acx+BK˜x(6a) ˙ ˜x=A˜x−∆·sgn(˜x)(6b) note that equation (6b) describes an autonomous system whose solution is the input of the stable linear system (6a). The stability properties of this algorithm are simpler to analyze than the one presented previously. They are also more tractable, and certainly less conservative. They are given next, and also result in a system (6) which is semi-globally asymptotically stable. Proposition 1: MODIFIED ∆-MCODING. Consider system (1) together with the modified ∆-Mcoding scheme (5). As before, assume that ∆∈Rn×nis constant and has the following form: ∆ = ∆0·In×n, and that ∆0fulfills the following inequality, ∆0> a · ||˜x(0)||, a =1 2λsup A+AT, then, ζ(t) = xT(t),˜xT(t)Tis bounded and tends to zero as t→ ∞. Proof: The proof is straightforward. Let V= ˜xT˜x/2, and from equation (6b), we have that ˙ V=1 2˜xT(A˜x−∆sgn(˜x)) + 1 2(A˜x−∆sgn(˜x))T˜x ≤ −∆0|˜x|+1 2˜xTA+AT˜x ≤ −∆0||˜x|| +1 2˜xTA+AT˜x ≤ −∆0||˜x|| +a||˜x||2=−||˜x|| (∆0−a||˜x||) where we have used the relation −|x| ≤ −||x||. From here it can be seen that the condition ∆0> a · ||˜x(0)||, with a given as in the last Proposition, makes Vdecrease, and hence ensures that ||˜x(t)|| <||˜x(0)||, and that ˜x(t)remains bounded and tends to zero in finite time. Finite-time convergence is typical in switching systems of the form (6b), and will not be demonstrated here. From this analysis we can also conclude that ˜x(t)∈ L2∩ L∞. To complete the proof note that equation (6a) describes an strictly stable linear system with input2˜x(t)∈ L2∩L∞, hence we can conclude that x(t)is also bounded and tends to zero, as the Proposition states. Remark 1: Note that the proposed new coding form, in addition to simplify the stability conditions by making them only depend on the estimation initial error, the new structure introduces a side effect of a low pass filtering action that will improve the filtering properties of the decoding dynamic equations. III. DISCRETE-TIME ALGORITHM In this section we present the extension of the continuous– time differential coding to the more realistic case of a discrete– time framework. We first introduce a simple case of a one– dimensional unstable system for which the optimal gain ∆and the attraction domain are easily found. Then, we extend this result to systems of higher dimension. A. One-dimensional system example Consider the following one–dimensional discrete–time system, together with the control law, and the differential coding modified law: •Open–loop system, and encoder: xk+1 =axk+buk(7) ˆxk+1 = [a−bK]ˆxk+ ∆ ·δk(8) δk=sgn(˜xk)(9) •Transmitted information: δk∈ {−1,1}, 2Because ˜x(t)is bounded and tends to zero with Lyapunov function ˙ V < βkxk2for some β, locally. hal-00394981, version 1 - 14 Jun 2009 6 •Decoder and control law: ˆxk+1 = [a−bK]ˆxk+ ∆ ·δk(10) uk=−Kˆxk(11) with, K∈R,a≥1,ac= (a−bK); |ac|<1, and ˜xk= xk−ˆxk. The modified differential coder (8)-(10) differs from its standard form in that the term within the square brackets depends on the system model parameters a, and b, and on the control gain K. In the standard form this term is equal to one, i.e. the encoder is described by a delayed integral term. As mentioned before, the advantages of this modification is that the coding error equations become decoupled from the system state equations, making the design of the feedback gain Kindependent of the coding structure. This algorithm gives the following error equations, in cascade form: xk+1 =acxk+bK˜xk(12) ˜xk+1 =a˜xk−∆sgn(˜xk)(13) stability of the whole system can thus be tackled by only studying the stability of the coding error equation (13). Let Vk= ˜x2 k, and ∇Vk=Vk+1 −Vk, then ∇Vk= ˜x2 k+1 −˜x2 k= (a2−1)˜x2 k−2a∆|˜xk|+ ∆2(14) The right hand side of the equality defines a second order polynomial of the form αr2+βr + ∆2= 0, with r=|˜xk| and roots r1, r2given by: r1=a−1 a2−1∆, r2=a+ 1 a2−1∆ Note that, for unstable systems3, these roots are always real and positive since a > 1, and that these values define three zones, where ∇Vkchanges sign, i.e. ∇Vk=   ≥0if |˜xk| ≤ r1 <0if r1<|˜xk|< r2 ≥0if r2≤ |˜xk| For some constant ∆, this means that if the initial condition |˜x0|is taken within the region where ∇Vkis negative, then the function Vkwill decrease until |˜xk|enters in the region |˜xk|< r1, where ∇Vkchanges sign. At the next sampling time, the system is pushed away from the region r1≤ |˜xk|. Due to the discrete nature of the problem, special care must be taken in order to check that this repulsive jump (at k+ 1) does not lead the system state out of |˜xk+1|< r2, and hence guarantee that the region |˜x|< r2is indeed an invariant set. The condition ensuring such a property depends on aand ∆ as follows. From (14) we have that the maximum possible jump of xk+1 obtained from any value in |˜xk|< r1happens at ˜xk∗= 0, and has a magnitude equal to ∆. Therefore, it is straightforward to see that the state remains within the region |˜xk|< r2for all k > k∗+ 1, as long as the following relation holds: ∆< r2=a+ 1 a2−1∆ 3The case of stable systems is simpler. We will only discuss here the more involved case of unstable ones. This inequality is satisfied if and only if a < 2, which is the well known necessary and sufficient condition for stabilization (see [21]). Summarizing, we have that if the initial error coding variable verifies |˜x0|< r2, then the error coding state |˜xk|is locally attracted to a threshold delimited by the value of r1. After that, the state |˜xk|is ultimately bounded by ∆, which is below r2as long as a < 2. It is important to remark that the value of ∆plays an important role in both; the system stability and the system precision: •|˜xk|<a+1 a2−1∆defined the domain of attraction (DOA) of the trajectories delimiting the system stability domain. It is enlarged by increasing ∆. •|˜xk|<∆is an invariant set (included in the previous DOA) defining the system precision. Better precision is reach with small values for. Therefore, larger values for ∆will make the system more stable but less precise; conversely, the reduction of the gain ∆will lead to small estimation errors, but at the same time it will reduce the domain where the system remains stable. This analysis is summarized below. Proposition 2: DISCRETE ONE-DIMENSIONAL SYSTEM. Consider system (7)-(11), with constant ∆and a < 2. Then if the initial conditions of the coding error are such that |˜x0|< r2 then, the following holds: •|˜xk|< r2,∀k≥0, •∃k0:|˜xk| ≤ ∆,∀k≥k0, and •limk→∞ d(xk,Bγ) = 0. where r2= ∆(a+1)/(a2−1) and d(xk,Bγ)is the minimum distance from xkto any point within the interval Bγ:= {x∈R:|x|< γ}, γ =Kb 1−ac ∆ Proof: See Appendix VIII-B Remark 2: This result displays an inherent trade–off between stability and precision for discrete–time differential coding when the gain ∆is fixed. This suggests the search of other coding strategies with varying gains. Note also that, as the sampling time4is chosen small, aapproaches 1and the precision is increased. Indeed, lima→1r1(a) = 0, thus by making Tsinfinitely small, the limit case of the continuous– time precision is approached. B. Noisy one-dimensional system The above analysis will be extended to the case where the open–loop model is affected by bounded noise. In this case, the state equations turn into xk+1 =axk+buk+wk uk=−Kˆxk ˆxk+1 =acˆxk+ ∆sgn(˜xk) where wkis a bounded noise (|wk|< W). With the proposed feedback, the dynamics of the error variable ˜xkturns into ˜xk+1 =a˜xk−∆sgn(˜xk) + wk 4The pole of the discrete-time system ais related the open–loop continuous one as a=eωolTs. hal-00394981, version 1 - 14 Jun 2009 7 Using again the Lyapunov function Vk= ˜x2 k, and ∇Vk= Vk+1 −Vk, then ∇Vk= (a2−1)˜x2 k−2(a˜xk+wk)∆sgn(˜xk) + 2a˜xkwk+ ∆2+w2 k ≤(a2−1)˜x2 k−2a|˜xk|(∆ −W) + (∆ + W)2(15) There is an interval of |˜xk|such that ∇Vkis negative as long as the last second order polynomial on |˜xk|has real roots (otherwise it would be an everywhere positive parabola). This condition is implied by 4a2(∆ −W)2−4(a2−1)(∆ + W)2>0 for which it is necessary that (∆ −W)2 (∆ + W)2>a2−1 a2= 1 −1 a2.(16) It is clear that the right hand side of the inequality is less than one for all a∈IR. Now we will consider that ∆is a tuning parameter and lim∆→∞ (∆−W)2 (∆+W)2= 1 Hence, the left hand side of (16) can be made arbitrarily close to 1, and greater than any value of the RHS. Rearranging terms, we have that the set of values of ∆such that the polynomial (15) has real roots is ∆> W 1 + q1 + 1 a2 1 + q1−1 a2 (17) This inequality makes sense for unstable open–loop plants (a > 1). Otherwise, (16) is trivially satisfied for all ∆. The roots of polynomial (15) determine the region of attraction in ˜xof the proposed scheme, and its steady state error. These roots are r1,2= a(∆ −W)±pa2((∆ −W)2−(∆ + W)2) + (∆ + W)2 2(a2−1) From the analysis of this expression, the following observations are in order: i) considering that the variable of polynomial (15) is the absolute value of |˜x|, the existence of positive real roots guarantees the existence of an interval on ˜xsuch that ∆V < 0; ii) a necessary condition for the polynomial to have positive real roots is (17); iii) if the latter holds, there are real values r1, r2such that r1<|˜x|< r2implies ∆V < 0, i.e. the domain of attraction is defined by |˜x|< r2, and the estimation error is ultimately bounded by |˜x|< r1. Regarding the variable xk, an analysis analogous to the previous subsection can be made by observing the following expression, xk=1 z−ac(Kb˜xk+wk) = 1 z−ac˜wk.Now considering that the new input ˜wkis ultimately bounded as |˜wk|< r1+W(from the upper bounds on |wk|and |˜xk|), the same arguments of the previous subsection lead to the conclusion that the state xasymptotically approaches the interval |xk| ≤ ˜γwith ˜γ=Kbr1+W 1−ac. C. n-dimensional systems The previous study can be generalized to systems with higher dimension of the form xk+1 =Axk+Buk(18) uk=−Kxk(19) where xk∈Rn,A∈Rn×n,B∈Rn×1.(A, B)are stabilizable pair. For the sake of space, the noise considerations will be spared for simplicity. We consider systems whose matrix A has distinct and real eigenvalues, i.e. there exists a transform matrix Tsuch that Λ = T AT −1=diag{λi}, with λi6= 0,∀i. The differential encoding modified law is: ˆxk+1 =Acˆxk+ ∆ sgn(T˜xk)(20) with, Ac= (A−BK); |λi{Ac}| <1, and ˜xk=xk−ˆxk,∈ Rn,∆∈Rn×n. Note that the sign function depends on the transform coordinates ˜zk=T˜xk, and sgn(˜zk) = sgn(T˜xk) = [sgn(˜z1,k),sgn(˜z2,k),···sgn(˜zn,k)]T The error equations in the xk, and ˜zkcoordinates are: xk+1 =Acxk+BKT −1˜zk(21) ˜zk+1 = Λ˜zk−T∆sgn(˜zk)(22) Proposition 3: DISCRETE N-DIMENSIONAL SYSTEM. Let λmand λMbe the smaller and the larger eigenvalues of Λ, respectively. Consider the differential encoding law (20), in feedback with system (18)-(19), with the gain ∆given as ∆ = ∆0·T−1·˜ ∆ ˜ ∆ = diag {sgn(λ1),sgn(λ2),··· ,sgn(λn)} where ∆0is a positive scalar constant. Then, if λ2 m>(λ2 M− 1), and the initial condition of the coded state is such that ||˜z0|| < r2the following holds: •||˜zk|| ≤ r2,∀k≥0, •∃k0:|˜zk| ≤ r1,∀k≥k0, and •limk→∞ d(xk,Bβ) = 0. where 0< r1< r2are given as: r1,2=∆0 λ2 M−1λm±qλ2 m−λ2 M+ 1 and d(xk,Bβ)is the minimum Euclidean distance from xkto any point within the ball Bβ:= {x∈Rn:kxk< β}, and βis a constant that can be computed as in the proof of Proposition 2. Proof: The proof follows along similar steps to the previous proof. Introduce Vk= ˜zT k˜zk, and ∇Vk=Vk+1 −Vk, then ∇Vk= ˜zT k+1 ˜zk+1 −˜zT k˜zk= ˜zT kΛ2− I˜zk −2sgn(˜zk)˜ ∆Λ˜zk+ ∆2 0sgn(˜zk)T˜ ∆2sgn(˜zk) = ˜zT kΛ2− I˜zk−2sgn(˜zk)˜ ∆Λ˜zk+ ∆2 0 hal-00394981, version 1 - 14 Jun 2009 8 Note that −sgn(˜zk)˜ ∆Λ˜zk=−X i ∆0|˜zi,k| · |λi| ≤ −∆0λmX i |˜zi,k|=−∆0λm|˜zk| ≤ −∆0λm||˜zk|| Then, ∇Vk≤(λ2 M−1)||˜zk||2−∆0λm||˜zk|| + ∆2 0≡ϕ(||˜zk||) ϕ(||˜zk||) = 0 defines a second order polynomial with roots, r1, and r2, as defined previously. A necessary condition for this polynomial to have real roots, or equivalently, for the minimum of ϕ(||˜zk||)to be negative is that λ2 m>(λ2 M−1). However, this condition is only sufficient for making ∇Vknegative in the domain r1≤ ||˜zk|| ≤ r2. Other less conservative conditions may be found. We have, as before, three regions: ∇Vk=   ≥0if ||˜zk|| ≤ r1 <0if r1<||˜zk|| < r2 ≥0if r2≤ ||˜zk|| and hence, the first two properties invoked in Proposition 3 follow exactly the same arguments than the ones used in previous sections. They are in consequence omitted here. The last statement follows from the relation xk=G(z)˜zk, and assuming that the x-subsystem (21) may also be diagonalized, the diagonalization leads to a set of nperturbed systems of the form (12) in new coordinates, from which we conclude boundedness of the state ||xk|| based on the same arguments as in Proposition 2. D. Data-Rates It is assumed here that samples of x(t), are digitalized with large enough resolution such that xkdigitalization errors are neglected. As the transmission is done by only one bit, the data transmission rate (number of bits transmitted per unit of time) associated to this scheme, is Rδ=one bit Ts=fs,in[Bits/sec] where fs= 1/Tsis the sampling frequency, and the system precision is determined by the size of ∆, the minimum quantization step in the decoding process. Let R, in [Bits/sec], be the network data-rate. Then if the maximum network capabilities are used, i.e. Rδ=R=fs, the system precision and the domain of stability can be rewritten as a function of R. For instance, in the one-dimensional case, the expression for r1and r2is rewritten as: r1=eα/R −1 e2α/R −1∆, r2=eα/R + 1 e2α/R −1∆ where αis the unstable pole of the continuous-time original system. As expected, increasing the transmission rate Rimproves precision (limR→∞ r1(R) = 0), and stability (limR→∞ r2(R) = ∞). IV. ADAPTIVE CODING The proposed scheme can be improved by making the quantization factor ∆variable. We will show that proper choice of a ∆–adaptation mechanism relating the quantization +- Gain selection 1 -1 1 z−ac ∆k ∆0 δk=±1 ∆k+1 =φk∆k ˆxk ˜xk xk × φk Fig. 6. Adaptive coding scheme (only the encoder structure is depicted). The figure shows the case of one–dimensional systems. The selection gain block toggles the value of φkaccording to Equation (24). step to the amplitude of the state results in global asymptotic convergence of the estimation error and the system states to zero. This is a significant achievement with respect to the fixed-gain scheme presented before, which was limited to finite domains of attraction and convergence to finite balls5. The main ideas of the adaptive scheme will be introduced here, while the details and the stability proof has been reported in [8]. One possibility is to make the adaptation law for ∆be explicitly state depended. However, the state xis not available at the receiver side. Therefore, the adaptation law must defined exclusively in terms of the information available both at the receiver and transmitter; which is not the state itself, but the coding signal δk, and also the estimation of the state ˆxk. This is an additional difficulty resulting for the hypotheses made in this work. A reasonable approach is to enlarge ∆kfor large values of the estimated error prediction, and decrease it for smaller values, in the same way as used in the standard adaptive delta modulation for open-loop scenarios [20], i.e. ∆k+1 =K(δkδk−1)∆k, K > 1. With this adaptation law, the value of ∆kincreases when two consecutive signals δk, δk−1have the same sign. Kis then selected to minimize the total distortion. There exist also other variations to this law, like the continuously variable slope delta modulation (CVSD). However, it is important to stress out that in the context of feedback systems, instabilities may occur if rate of variation of ∆kdoes not accommodate to the maximum (or minimum) rate of variation of the dynamics of the system to be controlled. This is the main difficulty in designing adaptive Delta-modulation laws for control. Under these considerations, a ∆–adaptive mechanism, with minimal storage and computation power requirements, can be designed under the following assumptions: 1) If δk=δk−1, the error prediction is assumed to be growing, thus ∆kmust be increased. 2) If δk6=δk−1, the error prediction is assumed to converge (oscillating close to zero) and ∆kmust be decreased. 5The attraction domain can be enlarged arbitrarily (by increasing ∆), at the price of increasing also the amplitude of the remaining oscillation (granularity). hal-00394981, version 1 - 14 Jun 2009