PID controller parameters.
In this chapter, we propose several Nyquist-like stability criteria for linear dynamical systems that are described by fractional commensurate order linear time-invariant (FCO-LTI) state-space equations (thus endowed with fractional-order transfer functions) by means of the argument principle for complex analysis. Based on the standard Cauchy integral contour or its shifting ones, the stability conditions are necessary and sufficient, independent of any intermediate poles computation, domain transformation, and distribution investigation, which can be implemented graphically with locus plotting or numerically without any locus plotting. The proposed criteria apply to both single and multiple fractional cases as well and can be exploited in regular-order systems without any modification. Case study is included.
- argument principle
- Cauchy integral contour
Fractional-order calculus possesses a long history in pure mathematics. In recent decades, its involvements in systems, control, and engineering have attracted great attention; in the latest years, its significant extensions in various aspects of systems and control are frequently encountered [1, 2, 3, 4, 5, 6, 7, 8]. It turns out that phenomena modeled with fractional-order calculus much more widely exist than those based on regular-order ones. It has been shown that fractional-order calculus describes real-world dynamics and behaviors more accurately than the regular-order counterparts and embraces many more analytical features and numerical properties of the observed things; indeed, many practical plants and objects are essentially fractional-order. Without exhausting the literature, typical examples include the so-called non-integer-order system of the voltage–current relation of semi-infinite lossy transmission line  and diffusion of the heat through a semi-infinite solid, where heat flow is equal to the half-derivative of the temperature .
One of the major difficulties for us to exploit the fractional-order models is the absence of solution formulas for fractional-order differential equations. Lately, lots of numerical methods for approximate solution of fractional-order derivative and integral are suggested such that fractional-order calculus can be solved numerically. As far as fractional-order systems and their control are concerned, there are mainly three schools related to fractional-order calculus in terms of system configuration: (i) integer-order plant with fractional-order controller, (ii) fractional-order plant with integer-order controller, and (iii) fractional-order plant with fractional-order controller. The principal reason for us to bother with fractional-order controllers is that fractional-order controllers can outperform the integer-order counterparts in many aspects. For example, it has been confirmed that fractional-order PID can provide better performances and equip designers with more parametrization freedoms (due to its distributed parameter features [4, 11, 12, 13]).
An important and unavoidable problem about fractional-order systems is stability [13, 14, 15]. As is well known, stability in integer-order LTI systems is determined by the eigenvalues distribution; namely, whether or not there are eigenvalues on the close right-half complex plane. The situation changes greatly in fractional-order LTI systems, due to its specific eigenvalue distribution patterns. More precisely, on the one hand, eigenvalues of fractional-order LTI systems cannot generally be computed in analytical and closed formulas; on the other hand, stability of the fractional-order LTI systems is reflected by the eigenvalue distribution in some case-sensitive complex sectors [13, 15], rather than simply the close right-half complex plane for regular-order LTI systems. In this paper, we revisit stability analysis in fractional commensurate order LTI (FCO-LTI) systems by exploiting the complex scaling methodology, together with the well-known argument principle for complex analysis . This work is inspired by the study for structural and spectral characteristics of LTI systems that is also developed by means of the argument principle [17, 18, 19]. The complex scaling technique is a powerful tool in stability analysis and stabilization for classes of linear and/or nonlinear systems; the relevant results by the author and his colleagues can be found in [20, 21, 22, 23, 24, 25]. Also around fractional-order systems, the main results of this chapter are several Nyquist-like criteria for stability with necessary and sufficient conditions , which can be interpreted and implemented either graphically with loci plotting or numerically without loci plotting, independent of any prior pole distribution and complex/frequency-domain facts.
Outline of the paper. Section 2 reviews basic concepts and propositions about stability in FCO-LTI systems that are depicted by fractional commensurate order differential equations or state-space equations. The main results of the study are explicated in Section 3. Numerical examples are sketched in Section 4, whereas conclusions are given in Section 5.
2. Preliminaries and properties in FCO-LTI systems
2.1 Preliminaries to fractional-order calculus
Based on [13, 15], fractional-order calculus can be viewed as a generalization of the regular (integer-order) calculus, including integration and differentiation. The basic idea of fractional-order calculus is as old as the regular one and can be traced back to 1695 when Leibniz and L’Hôpital discussed what they termed the half-order derivative. The exact definition formula for the so-called -order calculus was well established then by Riemann and Liouville in the form of
where and are real numbers while is an integer; more precisely, and is the smallest integer that is strictly larger than . is the gamma function at ; by (, p. 160), and it is convergent for each .
Basic facts about fractional-order calculus are given as follows :
• If is analytical in , then is analytical in and .
• If is an integer and , then reduces to the th-order derivative of with respect to ; namely, .
• If and thus , the definition formula for yields the identity relation .
• Fractional-order differentiation and integration are linear operations. Thus
• Under some additional assumptions about , the following additive index relation (or the semigroup property) holds true:
• If for with being a positive integer, then fractional-order derivative commutes with integer-order derivative:
The fractional-order calculus (1) and its properties are essentially claimed in the time domain. Therefore, it is generally difficult to handle these relations directly and explicitly. To surmount such difficulties, the Laplace transform of (1) is frequently used, which is given by
where and is the Laplace transform variable. Under the assumption that the initial conditions involved are zeros, it follows that . To simplify our notations, we denote by in the following if nothing otherwise is meant.
2.2 Definition and features of FCO-LTI state-space equations
A scalar fractional-order linear time-invariant system can be described with a fractional-order state-space equation in the form of
where , , and are the state, input, and output vectors, respectively. In accordance with , and , , , , and are constant matrices. We denote:
For simplicity, we employ to stand for the fractional-order indices set with with a little abuse of notations. The corresponding transfer function follows as
which is the fractional-order transfer function defined from to . In (4), stands for a diagonal matrix. Also, and are constants, while and are nonnegative real numbers satisfying
In the following, the fractional-order polynomial
is called the characteristic polynomial of the state-space equation (3).
We note by complex analysis (, p. 100) that is well-defined and satisfies
where is the principal branch of the complex logarithm of or the principal sheet of the Riemann surface in the sense of . In view of (6), we see that as fractional-order polynomial and as fractional-order fraction are well-defined only on for all , whenever at least one of , , and is a fraction number. Both are well-defined on the whole complex plane , if all , , and are integers.
Bearing (6) in mind, our questions are (i) under what conditions is holomorphic and has only isolated zeros and (ii) under what conditions is meromorphic and has only isolated zeros and poles?
To address (i), let us return to (6) and observe for any and that
To see under what conditions (respectively, ) possesses only isolated zeros and poles, let us assume that there exists such that
where , , and are integers. Then, if we set , then and can be re-expressed as
Obviously, is a so-called rational function on the -complex plane and possesses only finitely many isolated zeros and poles. More precisely, is a special meromorphic function that is determined up to a multiplicative constant by prescribing the locations and multiplicities of its zeros and poles (, p. 87); or equivalently it is in the form of a complex fraction with regular-order polynomials as its nominator and denominator. Also, is a regular-order polynomial that is holomorphic on the whole -complex plane and has finitely many isolated zeros. Thus, and can be written in the form of
Under the assumption that (7) and suppose that and have no zero, zeros, and zero singularities, it holds that . In addition, are integers such that and .
By (10), it is not hard to see that as (or equivalently by (6), ) and as (or equivalently, ). The latter says specifically by Corollary 3.2 of  that is a pole of . Then, is holomorphic on and has only finitely many isolated zeros and poles. It follows by (, pp. 86–87) that is meromorphic on . Confined to the discussion of this paper, may or may not be rational.
In the sequel, when the assumption (7) is true and and have no zeros and poles at the origin, then and are well-defined on with respect to fractional commensurate order . Only fractional commensurate systems are considered in this study.
2.3 Closed-loop configuration with FCO-LTI systems
Consider the feedback system illustrated in Figure 1, in which we denote by and , respectively, an FCO-LTI plant and an FCO-LTI feedback subsystem that possess the following fractional-order state-space equations.
where , , , and , respectively, are constant matrices, while , , , and are constant.
Fractional-order transfer functions for and are given as follows:
where , , , and are obvious and is the adjoint. Also
Now we construct the state-space equations for the open- and closed-loop systems of Figure 1. The open-loop system can be expressed by the fractional-order state-space equation:
In the closed-loop system, we can write the closed-loop state-space equation as
where . To explicate the Nyquist approach, we begin with the conventional return difference equation in the feedback configuration .
By definition, the characteristic polynomial for the closed-loop system is
where means the determinant of and
with and . Clearly, is the characteristic polynomial for , while and are the characteristic polynomials for the subsystems and , respectively, in the feedback configuration of Figure 1.
Let us return to (15) and continue to observe that
which is nothing but the return difference relationship for the fractional-order feedback system . By the definitions, it is clear that , , , and are all fractional-order. It is based on (18) that Nyquist-like criteria will be worked out. However, in order to get rid of any open-loop structure and spectrum, let us instead work with
3. Main results
3.1 Nyquist contours in the -/-domains
As another preparation for stability analysis in fractional-order systems by means of the argument principle for meromorphic functions, we need to choose appropriate Nyquist contours.
Firstly, the simply closed curve defined on the -domain as illustrated by the dashed-line in Figure 2 is the standard contour for a Nyquist-like stability criterion in terms of . The contour portions of along the two slopes actually overlap the slope lines. Clearly, the contour is actually the boundary of the sector region encircled by the dashed-line. The radiuses of the two arcs in the sector are sufficiently small and large, respectively, or simply and . The sector is symmetric with respect to the real axis, whose half angle is .
Secondly, the simply closed curve defined on the -domain as illustrated by Figure 3 presents the standard contour used for a Nyquist-like stability criterion in terms of . Again the contour is plotted with dashed-lines; in particular, the contour portion along the imaginary axis is actually overlapping the imaginary axis. More precisely, is the boundary of the open right-half complex plane in the sense that
where denotes the interior of a closed set. Similar to the contour , the radiuses of the two half-circles in are sufficiently small and large, respectively, namely, and .
Remarks about the contours and :
• In both cases, the origin of the complex plane is excluded from the contours themselves and their interiors. The reason for these specific contours is that and (respectively, and ) are well-defined merely on due to the relation (6) and .
• One might suggest that in order to detour the origin, the small arc in and the small half-circle in can also be taken from the left-hand side around the origin so that possible sufficiency deficiency in the subsequent stability conditions can be dropped. In fact, if the origin is in the interior of , and have no definition at the origin. Therefore, the argument principle does not apply rigorously.
3.2 Stability conditions related to FCO-LTI systems
where is the principal branch argument of on the Riemann surface.
3.3 Stability criterion in FCO-LTI systems
In what follows, a fractional-order polynomial is said to be commensurately Hurwitz if is fractionally commensurate order with some in the sense that all the roots of possess negative real parts. More specifically, we write
where are integers and for any . It is straightforward to see by definition that a fractionally commensurate order Hurwitz polynomial must be holomorphic on the whole complex plane.
vanishes nowhere over , namely, for all ; and the number of its clockwise encirclement around the origin is equal to that of its counterclockwise ones, namely,
In the above, the clockwise/counterclockwise orientation of can be self-defined.
Under the given assumption about the concerned characteristic polynomial and the fact that is holomorphic, we can assert that is well-defined in the sense that it is meromorphic and without singularities at the origin. This says in particular that the argument principle applies to (22) as long as over . Apparently, Eq. (22) holds even if there are any factor cancelations between and . This says that all unstable poles in , if any, remain in the left-hand side of (22).
More precisely, since vanishes nowhere over , we conclude readily that
where denotes the zero number of in and denotes the net number of the locus encirclements around the origin.
Note that all the roots of are beyond . It follows that . Together with , it follows by (17) that
The above equation says that if and only if . The desired results are verified if we mention that is equivalent to the assertion that has no roots in . This says exactly that has no roots in , and thus the concerned system is stable.
Several remarks about Theorem 1.
• Theorem 1 is independent of the contour and locus orientations; or the locus orientations can be alternatively defined after the locus is already drawn. The fractionally commensurate Hurwitz polynomial can be arbitrarily prescribed so that no existence issues exist. In addition, the stability locus is not unique. The polynomial actually provides us additional freedom in frequency-domain analysis and synthesis.
• When the stability locus with respect to the infinite portion of is concerned, it is most appropriate to let the commensurate degree of , namely, c-, satisfy , although can be arbitrary as long as it is fractionally commensurate Hurwitz. For example, in the sense of (25), if c-, the stability locus approaches the origin as . It may be graphically hard to discern possible encirclements around ; if c-, the stability locus contains portions that are plotted infinitely far from . When c-and let with be constant and monic, it follows from (25) that
• Each and all the conditions in Theorem 1 can be implemented only by numerically integrating for computing the argument incremental along the Cauchy integral contour , and then checking if holds. In this way, numerically implementing Theorem 1 entails no graphical locus plotting. This is also the case for the following results.
• The clockwise/counterclockwise orientation of can be self-defined. This is also the case in all the subsequent results.
Next, a regular-order polynomial is said to be -sector Hurwitz if all the roots of satisfy (20). More specifically, we write
where are integers and for any . It is easy to see that a -sector Hurwitz polynomial is holomorphic.
vanishes nowhere over , namely, for all ; and the number of its clockwise encirclement around the origin is equal to that of its counterclockwise ones, namely, .
3.4 Stability criteria for closed-loop FCO-LTI systems
Based on the return difference equation (31) claimed in the feedback configuration of Figure 1, together with the argument principle, the following -domain criterion follows readily.
satisfies: (i) for all ; (ii) . Here, stands for the contour by shifting to its left with distance .
To complete the proof, it remains to only show why we must work with the contour in general, rather than the standard contour itself directly. To see this, we notice that the shifting factor is introduced for detouring possible open-loop zeros and poles on the imaginary axis but excluding the origin. Since all zeros and poles, if any, are isolated, is always available. Furthermore, we have by (12) that
which says clearly that if there are imaginary zeros of , then factor cancelation will happen between and . Such factor cancelation, if any, can be revealed analytically as in the above algebras, whereas it does bring us trouble in numerically computing the locus (obviously, numerical computation cannot reflect any existence of factor cancelation rigorously). Consequently, when the standard contour is adopted, if there do exist imaginary open-loop poles, the corresponding stability locus cannot be well-defined with respect to . When this happens, one need to know the exact positions of imaginary zeros and/or poles and modify the contour accordingly to detour them before computing the locus; in other words, working with brings sufficiency deficiency when imaginary open-loop poles exist.
On the contrary, if there exist no imaginary open-loop poles, it is not hard to see that working with in numerically computing the locus yields no sufficiency redundance as , noting that always exists.
As a -domain counterpart to Theorem 3, we have.
satisfies: (i) for all ; (ii) . Here, we have
In the above, is the contour by shifting to its left with distance .
Several remarks about Theorems 3 and 4:
• The shifted contour reduces to the standard contour when . This is also the case for the shifted contour and .
• Clearly, the detouring treatments in Theorems 3 and 4 do not exist in Theorems 1 and 2, since the stability conditions in the latter ones are claimed directly on the fractional-order characteristic polynomials, in which transfer functions are not involved.
4. Numerical illustrations
4.1 Example description for Theorems 1 and 2
Consider a single fractional-order commensurate system  with the characteristic polynomial
where the commensurate order with , , and being positive integers as appropriately. In all the numerical simulations based on Theorem 1, the fractionally commensurate Hurwitz polynomial is employed. In all the numerical simulations based on Theorem 2, the -sector Hurwitz polynomial is adopted.
In what follows, the -domain contour is defined with and , while the -domain contour is defined with and .
4.2 Numerical results for Theorems 1 and 2
The following cases are considered in terms of and . In each figure, the left-hand sub-figure plots the stability locus in terms of , or simply the -locus, for some fixed and , while the right-hand sub-figure presents the stability locus in terms of , or simply the -locus.
• , , and . By examining the -loci of Figure 4 graphically, no encirclements around the origin are counted in each case of ; indeed, for each can be verified numerically without locus plotting. Therefore, the system is stable in each case.
The same conclusions can be drawn by examining the -loci of Figure 4. More precisely, we have for each numerically.
• , , and . By the -loci of Figure 5, no encirclements around the origin are counted in each case of graphically; or for each numerically. Therefore, the system is stable in each case.
The same conclusions can be drawn by examining the -loci of Figure 5. More precisely, we have for each numerically.
• , , and thus holds always. By the -loci of Figure 6, one net encirclement around the origin is counted in each case of graphically; alternatively, for each numerically. Therefore, the system is unstable in each case.
The instability conclusions in each can be revealed by means of the -loci of Figure 6. More precisely, we have for each numerically.
• , , and thus holds always. By the -loci of Figure 7, one net encirclement around the origin is counted in each case of graphically; alternatively, for each numerically. Therefore, the system is unstable in each case.
The instability conclusions in each can be drawn again by means of the -loci of Figure 7. More precisely, we have for each numerically.
• , , and . By the -loci of Figure 8, one net encirclement around the origin is counted in the case of graphically, or for numerically; two net encirclements around the origin are counted in each case of , or for each numerically. Therefore, the system is unstable in each case of .
The instability conclusions in each can be drawn again by means of the -loci of Figure 8. More specifically, we have for and for each numerically.
• , , and . By the -loci of Figure 9, no encirclements around the origin are counted in each case of graphically, or for each numerically. Therefore, the system is stable in each case of . However, two net encirclements around the origin are counted in each case of graphically or for each numerically. Therefore, the system is unstable in either case of .
Stability in each case of and instability for either case of can be verified by the -loci of Figure 9 as appropriately. Indeed, we have for and for each numerically.
Based on the numerical results, the stability/instability conclusions based on the -loci completely coincide with those drawn based on the -loci. This reflects the fact of Remark 2. These numerical results are also in accordance with those by  about the same example, which are summarized by working with solving polynomial roots. It is worth mentioning that polynomial roots are not always solvable in general. Fortunately, the suggested Nyquist-like criteria can be implemented graphically and numerically, independent of any polynomial root solution and inter-complex-plane transformation. Hence, the suggested technique is applicable more generally.
4.3 Example description for Theorem 3
where is a regular-order sensor model and is a cascading model consisting of a fractional-order PID in the form of 1, an amplifier modeled as , an exciter modeled as , and a generator with model . Fractional-order PID parametrization is addressed in  by means of particle swarm optimization.
In the following, we focus merely on verifying the closed-loop stability based on Theorem 3, based on the parametrization results therein. To this end, the -domain shifting contour is defined with , , and . To utilize Theorem 3, the fractionally commensurate Hurwitz polynomial is employed.
The so-called optimal controller parameters are listed in Table 1.
4.4 Numerical results for Theorem 3
Based on Table 1, the stability loci in the two cases are plotted in Figure 10. The stability loci for the two cases cannot be distinguished from each other graphically. By counting the outmost circle as one clockwise encirclement around the origin, then one can count another counterclockwise encirclement after zooming into the local region around the origin; it follows that the net encirclements number is zero. Indeed, our numerical phase increment computations in either case yields that . From these facts, Theorem 3 ensures that the closed-loop fractional-order system is stable. This coincides with the results in .
Stability is one of the imperative and thorny issues in analysis and synthesis of various types of fractional-order systems. By the literature [28, 29, 30], the frequently adopted approaches are through single/multiple complex transformation such that fractional-order characteristic polynomials are transformed into standard regular-order polynomials, and then stability testing of the concerned fractional-order systems is completed by the root distribution of the corresponding regular-order polynomials. In view of the root computation feature, such existing approaches are direct in testing methodology.
In this paper, we claimed and proved an indirect approach that is meant also in the -complex domain but involves no root computation at all. What is more, the main results can be interpreted and implemented graphically with locus plotting as we do in the conventional Nyquist criteria, as well as numerically without any locus plotting (or simply via complex function argument integration). This implies that the complex scaling approach is numerically tractable so that is much more applicable in fractional-order control design and parametrization. This point is significant for practical control applications involving fractional-order plants, which are our perspective topics in the future.
The study is completed under the support of the National Natural Science Foundation of China under Grant No. 61573001.
- The fractional-order integral portion in the PID is approximated by K I / s λ + 0.0001 in order to avoid definition problem at the origin when it is in the form of K I / s λ .