Preconditioning Techniques for Sparse Linear SystemsView this Special Issue
A Graph Approach to Observability in Physical Sparse Linear Systems
A sparse linear system constitutes a valid model for a broad range of physical systems, such as electric power networks, industrial processes, control systems or traffic models. The physical magnitudes in those systems may be directly measured by means of sensor networks that, in conjunction with data obtained from contextual and boundary constraints, allow the estimation of the state of the systems. The term observability refers to the capability of estimating the state variables of a system based on the available information. In the case of linear systems, diffierent graphical approaches were developed to address this issue. In this paper a new unified graph based technique is proposed in order to determine the observability of a sparse linear physical system or, at least, a system that can be linearized after a first order derivative, using a given sensor set. A network associated to a linear equation system is introduced, which allows addressing and solving three related problems: the characterization of those cases for which algebraic and topological observability analysis return contradictory results; the characterization of a necessary and sufficient condition for topological observability; the determination of the maximum observable subsystem in case of unobservability. Two examples illustrate the developed techniques.
The state variables that characterize a physical system are estimated by means of the data available at any given time. This data can be generated from a sensor network spread out over an area or from contextual and boundary constraints. In general, the known system variables are said to be sensed or measured variables whether they are sensed with a real device or their magnitudes are obtained in a sort of virtual sensors. The remaining variables are considered nonsensed or unmeasured variables. In such a context, the observability issue arises when we would like to know if the sensing system is enough to be able to determine the state of the system, that is, the system state variables.
This paper deals with a scenario where a well-known model describes the behavior of a physical system in terms of relationships between system variables and parameters. The system must be linear or linearized after a first-order derivative. In this context, a given sensing network is considered, and the system observability analysis is addressed.
The term observability was introduced in the realm of linear dynamical control systems . It stems from the capability of estimating the state of a system based on the information available. Although observability is essentially a numerical and algebraic problem, some techniques based on topology and graph theory have been developed to provide solutions in this area.
Due to the fact that observability and the problems related to it were studied in different engineering disciplines, the technical terminology is not totally uniform. As a result, some terms are more widely used in some areas and not in others and, in a few cases, different terms describe the same thing in different fields.
Five examples are described below pursuing the following aims: on one hand, illustrating how observability and other related problems constitute research topics in different physical, engineering, and industrial areas, where a sensor network is designed in order to analyze a given system; on the other, showing the multiple points of view from which these issues can be addressed and, in particular, how topological and graph-based approaches were developed in some cases.
The term sensor network comprises a broad spectrum of engineering and physical systems and, in particular, the topic of wireless sensor networks has led to issues that, in one way or another, are related to observability. This is the case of coverage, optimal node placement, and the minimum number of nodes required to achieve connectivity. In , it is shown that a graph model can be used to describe those systems, and some graph approaches have been developed in order to provide an answer to the challenges posed.
Whithin the sphere of linear control systems, the controllability problem was addressed from a graph-theoretic approach. A graph associated to a system was defined in  and conclusions related to several system properties are derived from the analysis of such a graph. A survey of the techniques proposed in the literature for structured linear systems can be found in . More recently, a graph approach to observability analysis is proposed in [5, 6].
High-voltage electric power networks constitute another field, where observability has been an important issue in system analysis for decades [7, 8]. It is worth mentioning that the approach to the problem in  where the authors characterize what they call topological observability through the existence of certain graphs that, defined in the electrical network, obey constraints derived from the sensing network. However, these graph techniques do not allow the inclusion of measurements that are currently being considered, such as current and phasor measures.
Observability has also been a motivation for research in traffic models in topics related to the origin/destination trip matrix estimation challenge. This is the case of , where the authors adapt topological techniques developed for electric power networks to this new context. Although this issue is more complex than the description made by the authors in their paper, it has been taken as an example to illustrate the techniques proposed in the present work as will be shown in a later section.
Material and energy balances that must take place in industrial processes are analyzed in . There, its authors distinguish up to four categories of balancing equations, depending on whether they consider or not materials, chemical reactions, energy, and entropy. They study the solvability of the resulting equations, for which a set of sensed variables is taken into account. The observability and redundancy of measurements as well as the errors in the measured values are included in the dissertation. Statistical techniques are used to estimate the state of the system by reconciliation. In the case of linear systems, a parallelism is established between system and sensing observability conditions and the existence of certain graphs defined from the process balancing flowsheet.
The common topic of the aforementioned scenarios, with regards to graph theory, is that certain graph techniques were developed in all the cases because of the existence of graphs or networks that characterized the systems with a given sensor set. Furthermore, the equations that describe the networks are linear or linearized. In this paper, a new graph technique is presented in order to characterize the observability of any linear physical system. The implementation of such a technique imposes constraints on the problem, summarized by the fact that the systems must be sparse and of large dimension. For any sparse and large dimensional physical system, an associated network will be defined based, exclusively, on structural considerations, that is, the topology of the equation system in its matrix form that relates the sensed variables with the state variables. It will be demonstrated that the system can be said to be topologically observable if there exists a certain graph within the associated network.
Krumpholz et al. developed in  a topological approach for the observability issue in the scope of electric power systems. Nevertheless, the problem related to the characterization of those cases for which algebraic and structural techniques return contradictory results is not studied. In this paper, the latter problem is solved, which has allowed carrying out a more general demonstration of the necessary and sufficient condition for topological observability than the one proposed by Krumpholz. Numerous techniques have been developed and widely and successfully tested for decades [12–15] in the scope of topological observability analysis in electric power systems. In this paper, a new graph approach is presented, which allows addressing the observability of any linear physical system or, at least, a system linearized after a first-order derivative, and not exclusively electric power systems. Boukhobza et al. had already developed a graph-theoretic technique in order to determine the state and input observability in structured linear systems . Unlike that proposal, the approach presented in this paper makes it possible to exploit techniques like those mentioned above [12–15] to characterize concepts like parametric unobservability and to easily determine the maximum observable subsystem.
The rest of the paper is organized as follows. Starting from a mathematical model, some terms will be introduced concerning observability and sparse physical systems in the next section. Section 3 is devoted to the bases of graph theory and the concepts used throughout the paper. Once the theoretical assumptions have been described, an analogy between linear equation systems and graph theory is established by means of a network associated to the physical system and a given sensor set. Section 5 introduces the concept of topological observability, which is characterized through the existence of a constrained graph in the associated network. The following section is devoted to the cases where the system is not observable and how the search for the maximum observable subsystem is addressed by means of the same graph techniques. Section 7 includes two examples in order to illustrate the techniques proposed in this paper, and how they can be implemented in absolutely different real engineering scenarios. Finally, some conclusions are presented in Section 8.
2. Mathematical Model
In order to determine the state of a system, , consider a set of variables that are sensed. These variables can be expressed in terms of the system state variables, : where represents a vector of errors due to the measurement acquisition process. In what follows, this error vector will be ignored because of its irrelevance regarding observability issues. Two different cases might be considered at this point, depending on the linearity of the above equations. On one hand, assume those equations are linear. Then, is a linear system, and a matrix formulation can be proposed instead of (2.1): where is a characterization matrix of the system. On the other hand, consider that is a nonlinear system that can be linearized around a certain state and let be the jacobian matrix, thus: where and . Summarizing, both cases resemble an equation system of the form: where is a constant term vector that results from the magnitudes sensed throughout the system, is the unknown vector that is directly related to the state variables, and is a coefficient matrix. In what follows, and in order to simplify the explanation, we will refer to and as the measurement and state variable column vectors, respectively. Also, will denote a generic measured variable, and will be a generic state variable. The observability issue arises when we would like to know if the variables considered in the sensor set are enough to determine the state of the system. It depends not only on how large the number of measurements is but also on their nature, and how they are spread out over the system. From an algebraic point of view, a system is said to be observable if the system given by (2.4) is solvable, that is, the equation system is consistent, and there exist at least linear independent equations. As Krumpholz et al. define in , the system is said to be algebraically observable if and only if the rank of is equal to . A well-known problem comes up when the system is ill-conditioned  and (2.4) must be solved or matrix is manipulated. For such cases, different numerical algorithms are proposed in the literature [17, 18]. In order to avoid this problem, other authors  take advantage of symbolic methods for sparse matrices . What this paper is related to are the cases, where the observability of a system such as the one defined above can be addressed in terms of structural considerations, what is called topological observability . In order to introduce this topic, let us define some concepts and hypotheses.
Let be an -dimensional physical system that is going to be the object of our study, and let a sensed variables set be defined, where magnitudes are measured over . Furthermore, let be the matrix associated to , as defined in (2.4). We will say that is a sparse system if the behavior of at any point can be justified exclusively by means of the knowledge of the variables in an area based on a certain neighborhood relationship. This is the case of a traffic model system where flow fluctuations in a certain region are strongly dependent on what happens in that area, whereas the events that take place in other parts show a weak dependence or absolute independence from them. One of the features that characterize a sparse system is that matrix is a sparse matrix. Then, some conclusions can be established in terms of structural considerations of , when the matrix dimensions and the degree of sparsity are large enough. For this purpose, Bunch and Rose  define a graph associated to a matrix , where a nonzero element of represents an edge that joins vertices and . Based on this, some properties can be studied in terms of graph theory because of the duality between sparse linear systems and graphs.
The obvious solution of calculating the rank of matrix may present problems and may not be even possible in the case of ill-conditioned systems, as mentioned above. In these cases, a topological-based approach becomes a good choice that presents a series of additional advantages derived from the capability of graphs to answer questions related to observability analysis, including the identification of the maximum observable subsystem and optimal additional sensor placement. In short, in this paper we will introduce new topological analysis techniques by means of certain graphs associated with sparse systems in order to determine the topological observability of such systems.
3. Graph Theory
A graph is defined as a collection of nodes or vertices that are joined through the so-called edges or branches. For the sake of homogeneity here we will use the term branches both for general graphs and for the case of trees, which are basically graphs without loops. In the scope of this work we are interested in defining graphs within a given network, which is also a collection of nodes and branches. In other words, a network must be interpreted as the context where any given graph is declared, in such a way that nodes and branches belonging to a graph are also present in the network for which the graph is defined. Nevertheless, not all the nodes and branches of the network are always present in a graph.
Definition 3.1. Let be a network, where and are the sets of nodes and branches, respectively; a graph of is defined as a set of nodes, , and a set of branches, .
Thus, as in the case of networks, a graph can be denoted by a couple, as follows: In what follows, it is assumed that is a connected network, that is, a network where there exists a path in between every pair of nodes of . In the same way, a connected or unconnected graph of can be defined. If a graph of is not connected, each connected subgraph that makes it up is known as a connected component. When a connected graph contains no loops, it is called a tree of .
Definition 3.2. A graph of is said to be a spanning tree if contains no loops, and .
A directed graph results from the assignment of a direction to each branch in such a way that a node is known as the source, while another node is the target of a directed link.
The matrix representation of any graph is the node-to-branch incidence matrix, . This is a matrix with as many rows as nodes are in the graph, and where the number of columns is equal to the number of branches in the graph. The elements of , in the case of directed graphs, are defined as follows:
The rank of a graph of is defined as: where size denotes the number of nodes in , while indicates the number of connected components of .
Definition 3.3. Let be a connected network, a graph of is said to be of full rank if its rank equals the maximum possible value, .
The rank of a graph of is, by definition, equal to the rank of its associated incidence matrix . If is of full rank, the rank of equals the number of rows minus one. In other words, one row of is linearly dependent on the others. That is the reason why a reduced node-to-branch incidence matrix is defined, resulting from the elimination of a row from . The following expression summarizes all of the above: The selection of one node among others for which the associated row is erased is arbitrary. In what follows, this node is going to be known as the reference node.
Definition 3.4. The closure  of a connected graph in is defined as a graph , where and is composed by all the branches in that join pairs of nodes in .
4. Network Flow Analogy
Consider a set of linear independent variables, , that determine the state of a system . Let be a system variable whose magnitude may be expressed as a linear relationship between the state variables, as follows (notation: in what follows, subscripts and , are used to refer to generic state variables and network nodes; subscript , refers to a node that is known as source node; subscript , denotes measurements and equations; subscript , refers to generic network branches; subscript under arrays or vectors denotes that those structures contain exclusively branch parameters or variables): where there is at least one value of for which . Consider that , where , is the first nonzero coefficient in the above expression. In other words, . Then, the expression can be rewritten as: which is consistent with the analogy to a flow network as shown in Figure 1. In it, a current is injected into the network through node and flows to the remaining nodes, to , according to the admittance values and potential differences of the branches connecting them. Therefore, the following equality must hold: where, for a generic node , represents the potential level of the node with respect to a zero potential reference node, GND in the figure; represents the admittance that characterizes a branch connecting node to node , so that is the current that flows from node to node due to the potential difference observed from node to node ; similarly, is the admittance between node and GND; hence, a current flows from node to the reference node. The network in Figure 1 is defined as the elementary network associated to the linear (4.3), which is known as the network nodal equation at node . The elementary network is a tree, and is defined as the source node of that tree, while the remaining nodes are considered target nodes.
Note that, on one hand, the elementary network in Figure 1 is characterized by the nodal (4.3), where the flow injected in node equals the magnitude; on the other hand, a solution to the elementary network in Figure 1 is consistent with (4.1).
Let be a system, where is a set of variables whose magnitudes may be expressed as linear relationships of the form: A network associated to a linear equation system such as the one shown above is defined as the result of the superposition of the elementary networks associated to each for all. Then, the solvability of the linear equation system (4.4) is equivalent to that of its associated network, since a particular solution to the equation system is consistent with the associated network. Figure 2 shows an example of an associated network as a result of considering all the elementary networks in their entirety, denoted by . Let us take a look at a generic node in the figure, such as node number 4. It is easy to see how the incident branches to node 4 are due to elementary networks associated to variables for which 4 is the source node, such as , and those elementary networks including 4 as a flow target node, such as and .
Let be the elementary network associated to a generic variable defined in as shown in (4.4), where is the source node. A branch admittance matrix of is defined as a diagonal matrix as follows: In what follows, it is assumed that all coefficients considered in the construction of a matrix , such as the one defined above, are nonzero. In other words, null coefficients, , are removed from (4.4). Note that this constraint does not guarantee that all diagonal elements in are nonzero because there might exist a case in which, for a certain , the sum equals zero. Those cases are related to the concept of parametric unobservability, and it will be introduced later.
Taking into account the contribution of all the variables in to the whole associated network , a branch admittance matrix of is defined as a block diagonal matrix: Then, the following equality is satisfied: where is the reduced node to branch incidence matrix of ; is the nodal potential vector, that is, the system state variable column vector; if is the number of branches in , and they are numbered from 1 to , is the branch flow vector, that is, a column vector of magnitudes that flow through branches in ; is a matrix that relates potentials at nodes in with branch flows .
Equations (4.4) can be expressed in matrix form as follows: where is defined as a coefficient matrix, and where and are column vectors. Note that each row of , that is, each variable considered in the system, will result in an elementary network of that is a tree because of the lack of loops. Therefore, as any branch in arises from the existence of a nonzero element in , a equation to branch incidence matrix, , associated to can also be defined as follows:
The following equality holds:
Equation (4.7) characterizes network as well as (4.8) characterizes system from a set of variables and, therefore, from equalities (4.7) and (4.10) it can be concluded that the study of the determinism of is equivalent to the observability of under constraints related to the variables taken into account.
5. Topological Observability
Krumpholz et al. introduced in  the term parametric unobservability as a vague notion needed to justify the concept of topological observability in electric power networks under certain assumptions. In this section, we present a formal description that allows defining and characterizing parametric unobservability and demonstrates how topological observability can be addressed by means of the existence of certain graphs under constraints.
Let be a large -dimensional sparse physical system, where a sensing system is defined by means of measured variables, . Let be the coefficient matrix, as defined in (4.8), associated to and the sensing system, and let be the associated network. It is important to note that characterizes only those parts of the system related to measurements, but not the whole physical system. In particular, it shows the relationship between the sensor set considered and the state variables. Therefore, might be a diagonal or block diagonal matrix, without implying either the existence or nonexistence of decoupled subsystems in . Obviously, the observability analysis of decoupled subsystems, if they exist, can be carried out independently.
The necessary and sufficient condition for algebraic observability of a system and a sensing configuration , as proposed above, is
Let us consider an algebraically observable system with respect to a sensor set . As is an matrix and , from (5.1), it follows that a collection of linearly independent rows of can be found. Let be the subset of corresponding to those linearly independent rows of . Therefore, an equation subsystem might be defined in with respect to , that should be characterized using an coefficient matrix and its associated network in such a way that: where , , and the determinant . is known as a critical sensing configuration in the sense that the loss of any measurement in should derive in the loss of the observability condition with respect to . For the same reason, system is said to be critically observable with respect to . The determinant is calculated as a sum of products, each coming from elements in , and no two coming from the same row or column. Since is a nonsingular matrix, at least one of these products must be nonzero. Thus, without loss of generality, in what follows let a permutation of rows be considered such that all the factors of the aforementioned nonnull product lie on the principal diagonal of . Note that any row permutation in does not alter the associated network .
It is clear that the first entry in in the first row is nonnull and, therefore, there exists in a branch joining node 1 and the reference node. In the second row, there are two possible cases: on one hand, if the diagonal element is the first nonzero element in that row, there exists a branch in joining node 2 and the reference node and, indirectly, the first node too; on the other hand, if the diagonal element is not the first nonzero one, there is a link in between nodes 2 and 1. This argument can be repeated for the next row and up to the last one. Eventually, a spanning tree of full-rank of is completed because of the lack of loops and the inclusion of the totality of the nodes in the network. Furthermore, the previous analysis leads exclusively to one branch in from each row in . In other words, the branches in are derived from different measurements in . Since results from the superposition of elementary networks, one for each sensed value, each branch in belongs to a different elementary network.
In order to demonstrate that the existence of such a spanning tree is sufficient, under certain conditions, for the observability of a system with respect to a sensing configuration, a reverse path is considered in which branches are added recursively to a starting spanning tree until the entire network is encompassed.
Consider a spanning tree of , where each one of the branches of belongs to a different elementary network out of the that form . That is, each elementary network in has a branch and only one that belongs to . From (4.10), it follows that a matrix can be defined as: where is a selection of columns from , while is a selection of rows from corresponding to the branches of . Thus, is the identity matrix because the branches of belong to different elementary networks and it follows that: Note that as has the same sparse pattern as , and is a spanning tree of of full rank, . In other words, because is nonsingular. Let be a generic row of . The first nonzero entry in row is in the same column, generically represented by , as the first nonzero element in row of . At this point, two cases might take place: one in which column is the only nonzero entry in row , and another for which there exists a second nonzero element in column of row in . Since the determinant of a square matrix can be calculated, according to Laplace’s formula, as a weighed sum of cofactors or adjuncts along a row or a column, it follows that: where and are the elements of in row and columns and , respectively, and and are their cofactors. Taking into account the same notation as used in (4.4), if is the only nonzero entry in row , ; otherwise, . In both cases, the determinant must be different from zero.
Let be a graph of that results from the union of and one-branch of not in . Consider that the additional branch belongs to an elementary network that corresponds to row of and whose source node is denoted by . If matrix is defined from in the same way as matrix was from , then, two different cases may follow(1)the additional branch joins node and the reference one. Therefore, as the admittance of this branch is equal to , the only entry that makes matrices and different is: and, from (5.5), the determinant: that is equal to zero when: (2)the additional branch joins node and a node , where . In this case, the branch admittance equals and both matrices , and are equal but for two entries in row : and, again, the determinant: that vanishes when:
Consider to be a graph of that results from the addition to of a number of branches of not in , and let the matrix be defined such that . Let be a graph of formed after the inclusion in of a branch of not in , and consider that the additional branch belongs to an elementary network that corresponds to row of for which the source node is denoted by . One of the next two cases will follow:(1)the additional branch joins nodes and the reference one. The admittance of branch is equal to and the determinant of is estimated by: where is the cofactor of . The above determinant becomes null when: (2)the additional branch joins node and a node , where . Then, the branch admittance is equal to , and the determinant of is given by: where and are the cofactors of and , respectively. The determinant will be null if:
New branches can be added to the given graph, one by one, until the entire network is completed, after the inclusion of all the branches in . Therefore, it is concluded by induction that the determinant of matrix is nonzero if its entries , , do not meet any equality such as (5.8) and (5.13) for branches that join the reference node and (5.11) and (5.15) otherwise.
Consider an example in which a collection of four sensed magnitudes are acquired from a four-dimensional physical system. As a result, an equal number of linear equations that relate and the state variables are established, and a matrix of coefficients is given by:
Figure 3(a) shows the resulting associated network, where the branch admittance values are indicated as well as the sensed variables to which each branch is associated. Figure 3(b) shows a spanning tree of full rank, in which it can be noted that the four branches that conform the tree are associated to four different measured variables. Then, it follows that: where . The graphs in Figure 4 show how the entire network can be reached from by the addition of each branch of not belonging to , and how, at each step , a new is defined from the previous one after modifying one or two matrix entries, depending on the case. It can be seen that it always follows that except for the exception cases defined in (5.8), (5.11), (5.13), and (5.15).
|(a) associated network|
|(b) spanning tree|
Note that this result was reached from the consideration of, on one hand, the network topology and the number, nature, and location of sensors in the network and, on the other, the network parameters. To deal with these two approaches, the concept of parametric unobservability is introduced.
Definition 5.1. A large dimensional and sparse physical system , for which a sensing system is defined, is said to be parametrically unobservable with respect to if, in spite of the fact that the ranks of matrices and are equal to , the rank of is less than due to the value of one or more coefficients of .
The relevance of this concept lies in the fact that, in large dimensional sparse physical systems where the parameters are roughly estimated from empirical data or are subject to environmental distortion, it is unlikely for parametric unobservability to occur . In other fields, such as structured linear systems, it is often necessary to work under the assumption of a lack of knowledge of system parameters . In these scenarios, the parametrically unobservability should be associated with a particular set of parameter values. Thus, the observability of a system is said to be true when it is so for almost all parameter values, that is, for all of them except for a set of particular cases in the parameter space. Even though not all physical systems may meet this requirements, there exist evidences that are true for real cases. For example, electric power network analysis involves hundreds or even thousands of state variables that are usually related to the voltage at network nodes. The system state  can be estimated by means of the measurement of the power that flows into and through the electric network and which is influenced only by neighboring node states. Thus, the resulting system is clearly sparse, and circuit parameters are affected by environmental conditions such as temperature and humidity as well as by the unreliability of parameter estimation. Another example is the case of traffic model analysis . As explained later in the example in this paper, vehicles usually move along a geographical area according to a set of established origin/destination pairs. Traffic flows are sensed at routes in the network in order to estimate the state of the system, that is, origin/destination pair traffic flows. As the network grows, the sparsity becomes more plausible. Additionally, system coefficients are estimated, among other factors, from probabilistic considerations related to the ability of people to opt for one route or another. In brief, parametric unobservability is, in these two cases, highly improbable despite being mathematically possible.
On the basis of the large dimension, sparsity and parameterization uncertainty of such systems, in order to address the observability issue a new strategy is proposed involving exclusively structural and not numerical considerations. For this, a new observability definition should be provided.
Definition 5.2. Let be a large dimensional sparse physical system, where a sensor network is considered; with is said to be topologically observable if is algebraically observable or parametrically unobservable with respect to .
Summarizing, it has been demonstrated that the existence of a spanning tree of full-rank of where the branches of belong to different elementary networks of constitutes a necessary and sufficient condition for topological observability. In what follows, any graph of with a number of branches that belong to different elementary networks, that is, are associated to different measurements of , is known as a measured graph.
Theorem 5.3. Let be a linear and large -dimensional sparse physical system, where a sensing system is defined by means of a number of measured variables, ; is said to be topologically observable with respect to if and only if there exists a measured spanning tree of .
The analysis of the observability of a large dimensional sparse physical system with respect to a sensing system from a topological point of view involves searching for a measured spanning tree of full rank among all possible graphs of constructed in such a way that each elementary network that forms contributes with and only with one branch to . If the number of sensed values considered is larger than the dimension of system , will be included, if it exists, as part of a measured spanning graph of . In what follows, it is assumed that any graph of is a measured graph.
There could be different ways to construct a spanning tree, and any one of them would be valid [12–15]. However what is important here is the fact that the existence of a measured spanning tree is a sufficient condition for the topological observability of a linear system.
Summarizing, taking as a basis the experience in observability analysis in electric power systems, a generalization of the topological approach was developed to address this issue in the scope of other linear, or linearized after a first order derivative, real engineering physical systems. A necessary and sufficient condition for topological observability was established by means of a graph theoretic approach. Finally, thanks to this approach, the characterization of the cases where algebraic strategies do not lead to the same results as those derived from structural analysis was carried out.
6. Maximum Observable Subsystem and Observability Islands
If the observability system test fails for a sensing configuration, it is said that the system is not observable or unobservable. In such cases, the knowledge that might be acquired about one or more parts of the system by all the measures considered should not be underestimated. If a system is not observable, it may be possible to identify a subsystem for which the state can be estimated, it is said that the subsystem is observable. A nondivisible observable subsystem is known as an observability island. The number of observability islands may vary and depends on the associated network topology, the sensors considered and their location in the network.
Consider an -dimensional sparse physical system and a sensing configuration for which an associated network is defined. Let be an observable island of and , and let be its associated subnetwork; is known as an observable subnetwork.
A node belonging to is said to be an observable node, and a branch belonging to is an observable branch. A measured spanning graph of is known as an observable graph. Nodes and branches that do not belong to any observable subnetwork are said to be unobservable.
Let be a measured graph of ; a measure associated to a branch of is said to be wholly contained  in if the elementary network associated to is contained in the closure of in . By extension, a measure is said to be wholly contained in a subnetwork if its associated elementary network is included in the closure of in .
Any measurement considered in is wholly contained in . Hence, the state of an observability island may be estimated by means of a wholly contained sensor set.
The union of all the observability islands in a system derives in a maximum observable subsystem while the union of all their associated observable subnetworks in results in the maximum observable subnetwork. This subnetwork is maximum because it comprises the largest possible number of observable nodes and, if it exists, it is unique .
Consider a system that is not observable for a sensor set . Then, no measured spanning tree will be found, as derived from Theorem 5.3. Instead, consider a measured graph of as one of the largest connected graphs that can be formed according to the sensing system and the constraints described earlier. Then, is known as a maximum measured graph of but not spanning. The next theorem was extracted from , where relevant properties concerning maximum measured graphs and observable subsystems are described.
Theorem 6.1. Let be a system and the associated network from a given sensor set. If is any maximum measured graph of , the maximum observable subnetwork is contained in the closure of in .
Therefore, based on one of the maximum measured graphs, an iterative process can take place by which the not wholly contained measurements, and their elementary networks are removed from the system until the maximum observable subnetwork is obtained. Additionally, other strategies concerning the search for the maximum observable subsystem can be found in [13, 14] in the scope of electric power networks.
Two examples are presented in this section in order to illustrate the techniques developed in this paper, focusing the attention on the fact that these techniques are valid for different real engineering problems, where a collection of linear equations or equations linearized after a first order derivative describe the behavior of the system from a measurement acquisition system viewpoint.
7.1. Traffic Model Example
One of the fundamental problems in traffic models concerns the estimation of the origin/destination (OD) trip matrix. Traffic flows are measured by means of sensors spread out at different locations in a study area. These data, in conjunction with other available information, are used to estimate the target matrix, that is, the traffic derived from any OD movement. For each OD pair there exist, in general, more than one alternative to complete the trip that are usually expressed in terms of percentages or probabilities based on contextual factors. In addition, the flow magnitudes at a link in a traffic network can be broken down into percentages of vehicles moving along different OD trips. Thus, linear relationships can be established between OD-pair and link flows. Let and be OD-pair and link flow vectors, respectively; their linear relationships can be described by a matrix as follows:
Figure 5 shows a benchmark case, known as the Nguyen-Dupuis network  in the literature, consisting of 13 plausible origin/destination places that are interconnected by 19 bidirectional links. In that scenario, vehicles can move from one place to another through suitable routes. Figure 5 shows indices assigned to links along with their directions. Therefore, for an OD-pair, any possible path is defined as a series of oriented link indices; for example, the sequence denotes an alternative for a displacement from 1 (origin) to 2 (destination).
In what follows, it is assumed that matrix and the OD-pairs are known. Below are all the OD pairs considered in this example and their potential paths as well as matrix . They are the same as those tested in :
Note that characterizes the physics of the whole traffic network because it relates the defined OD-pair flows with all the 38 possible oriented traffic link flows: A question arises when we want to know if a given sensor network allows to estimate the state of the traffic system or where sensors should be placed in order to complete an observable sensed system. Two cases are going to be taken into account concerning these issues.
7.1.1. Case 1: Observable Configuration
Consider a sensor network consisting of 8 traffic flow meters that result in a measured variable vector whose magnitudes might be estimated by means of a submatrix of and the system state variables , that is, OD-pair traffic flows, as follows:
The question arises as to whether OD-pair traffic flows can be estimated from this sensor set among the aforementioned oriented link flows.
In Figure 6, the elementary networks derived from the coefficient matrix of (7.4) are shown. Note how OD-pairs play the role of network nodes, while OD-pair traffic flows are the network node potential levels. In the figures, branch admittance values are indicated; indices were assigned to the branches and are shown in the figures by smaller numbers next to the arrows.
(a) E 1← v 2
(b) E 2←v 4
(c) E 3← v 5
(d) E 4← v 7
(e) E 5← v 12
(f) E 6← v 15
(g) E 7← v 27
(h) E 8← v 38
Figure 7 shows the entire associated network and how a measured spanning tree, highlighted using thick line, was found among other possibilities. Note that each elementary network is related to one and only one branch in the resulting measured spanning tree. This tree is not unique but the existence of, at least one, guarantees the topological observability of the system for the sensor set defined in (7.4).
7.1.2. Case 2: Not Observable Configuration
In a second case, a total of 6 traffic flow meters are considered. The question arises as to whether system observability can be achieved by incorporating additional sensors. And if it is not possible, which is the maximum observable subsystem?
Let be the initial sensor set. This is a subset of the observable configuration discussed earlier. Therefore, the linear equations that characterize this case are given by:
Figure 8 shows the resulting associated network, , and one of the possible maximum measured graphs, (thick lines). Note that OD-pairs 2-4 and 4-3 are clearly not observable, that is, their traffic flows cannot be estimated by means of the available measurements. A more detailed analysis leads to the conclusion that measurements , , and are not wholly contained in and, therefore, their associated elementary networks , , and , respectively, should be removed from the network in order to search for the maximum observable subnetwork. This argument should be repeated until the resulting subnetwork is made up exclusively of elementary networks associated to wholly contained measurements. That is the case after removing , the elementary network associated to measure . From there, the maximum observable traffic subsystem is immediate and is given by OD-pairs 3-1 and 3-4 and oriented traffic link flow sensed values .
To achieve a totally observable system, it is necessary to add two new traffic flow meters that allow to join the maximum measured graph in Figure 8 and the isolated nodes given by OD-pairs 2-4 and 4-3. Each row in matrix of (7.2) corresponds to an oriented traffic link flow and, in particular, those rows with nonzero coefficients in columns related to isolated OD-pairs are plausible candidates to improve the system observability. Thus, the inclusion of one of the sensed values from:
that allow joining the OD-pair 2-4 node, in conjunction with one of the following:
that allow joining OD-pair 4-3 node, would permit observing the whole traffic system.
7.2. Electric Power System Example
As it was mentioned in the introduction, observability analysis in electric power systems has been an important research topic for decades. In particular, this issue can be addressed by means of topological methods, when the set of measured variables are made up exclusively of bus voltages and active and reactive powers that are injected into or flow through the system . In those cases the system can be considered as a decoupled system , that is, a pair of two independent subsystems: one of which can be analyzed by means of active power measurements, and is known as - subsystem; the other one, the - subsystem, can be studied exclusively from bus voltages and reactive powers measured in the system. Only the - subsystem is going to be analyzed in this example. Such a subsystem is observable  when a sufficient number of well-placed reactive powers are measured, and, at least, one node voltage is known at any node.
An electric power system is commonly represented as a mesh where the edges denotes the lines in charge of transporting the electric energy, and where the nodes are the places where the lines are incident, that is, the places where electricity is generated, consumed, or transformed. Figure 9(a) shows the topology of an example of an electric power system with 8 nodes. The places where reactive powers are acquired in the system are shown in Figure 9(b), where two kinds of measurements may be distinguished:(i)node measurements, numbered as 1, 2, and 3 in Figure 9(b), corresponding to reactive powers injected into the system through a node. These derive in equations of the form: where denotes the -th node reactive power, represents the voltage at node , is a coefficient related to measurement and node , and denotes a constant term related to measurement ;(ii)branch measurements, numbered as 4, 5, 6, and 7 in Figure 9(b), corresponding to reactive powers that flow through the lines. These derive in equations of the form: where denotes a branch reactive power that is acquired in a line that joins nodes and .
(a) system topology
(b) measurement configuration
(c) associated network
(d) simplified associated network
(e) measured spanning tree
Finally, a voltage measure at node 1 is also considered, resulting in an equation as follows: Summarizing, the linear equations that characterize the - subsystem and the given measurement configuration are as follows: Figure 9(c) shows the associated network derived from (7.11), where branch admittances were suppressed in order to clarify the drawing. The numbers close to the oriented edges of the graph denote the order of the measurement from which the edge is derived, that is, the order of the elementary network in which it is defined. Note that the only branch associated to measurement and, in general, to any node voltage measure, is the one that joins the node where the voltage is acquired and the reference node. As a result, the reference node is implicitly connected to the rest of the nodes due to the inclusion of just one node voltage measurement, and a simplified associated network may be taken into account as shown in Figure 9(d), where thicker lines represent the edges that are present in the entire individuals of the search space of measured graphs. Note that those edges are the ones due to the node voltage and branch reactive power measurements.
Finally, one of the possible measured spanning trees is shown in Figure 9(e), after the assignment of each of the eight measurements considered to one of the edges in the associated network. The existence of such a tree permits concluding that the electric power - subsystem is topologically observable for the given sensing system.
In this paper, a new topological approach to the determination of the observability of a physical system where a sensor network is defined has been presented. The techniques developed in this paper were inspired by the contributions of researchers in the scope of electric power systems and generalized to other physical sparse linear systems. The terms parametric unobservability and topological observability have been introduced and justified in a formal way, which allows characterizing those parameter dependent cases where an algebraic approach to the observability issue led to different results than the topological one. A sensing system has been considered for any linear physical system or, at least, linearized after a first order derivative. From there, an associated network has been defined, and it has been demonstrated that the existence of certain constrained graphs, known as measured graphs, in the scope of the associated network permits characterizing the topological observability of the system. From this graph approach, the determination of the maximum observable subsystem can be carried out in case of unobservability. The technique has been illustrated with the help of two examples in the scope of traffic sensing structures and electric power systems.
This work was partially funded by the Xunta de Galicia, the MICINN of Spain and European Regional Development Funds through projects 09DPI012166PR, 10DPI005CT, and TIN2011-28753-C02-01.
F. C. Schweppe, “Power system static-state estimation, parts I,II and III,” IEEE Transactions on Power Apparatus and Systems, vol. 89, no. 1, pp. 120–135, 1970.View at: Google Scholar
A. Monticelli and F. F. Wu, “Network observability: theory,” IEEE Transactions on Power Apparatus and Systems, vol. 104, no. 5, pp. 1042–1048, 1985.View at: Google Scholar
G. R. Krumpholz, K. A. Clements, and P. W. David, “Power system observability: a practical algorithm using network topology,” IEEE Transactions on Power Apparatus and Systems, vol. 99, no. 4, pp. 1534–1542, 1980.View at: Google Scholar
V. Veverka and F. Madron, Material and Energy Balancing in the Process Industries: from Microscopic Balances to Large Plants, Computeraided Chemical Engineering, Elsevier Science, 1997.
V. H. Quintana, A. Simoes-Costa, and A. Mandel, “Power system topological observability using a direct graph-theoretic approach,” IEEE Transactions on Power Apparatus and Systems, vol. 101, no. 3, pp. 617–626, 1982.View at: Google Scholar
S. Vazquez-Rodriguez, A. Faina, and B. Neira-Duenas, “An evolutionary technique with fast convergence for power system topological observability analysis,” in Proceedings of the IEEE Congress on Evolutionary Computation (CEC '06), pp. 3086–3090, 2006.View at: Google Scholar
J. R. Rice, Matrix Computations and Mathematical Software, McGraw-Hill Computer Science Series, McGraw-Hill, New York, NY, USA, 1983.
H. J. Kim, “A new algorithm for solving ill-conditioned linear systems,” IEEE Transactions on Magnetics, vol. 32, no. 3, pp. 1373–1376, 1996.View at: Google Scholar
S. K. Kurtzl, “A direct algorithm for solving ill-conditioned linear algebraic systems,” Advances, vol. 42, pp. 629–633, 2000.View at: Google Scholar
M. Cosnard and L. Grigori, “Using postordering and static symbolic factorization for parallel sparse lu,” in Proceedings of the 14th International Parallel and Distributed Processing Symposium (IPDPS '00), pp. 807–812, 2000.View at: Google Scholar
W. F. Tinney, V. Brandwajn, and S. M. Chan, “Sparse vector methods,” IEEE Transactions on Power Apparatus and Systems, vol. 104, no. 2, pp. 295–301, 1985.View at: Google Scholar
K. A. Clements, G. R. Krumpholz, and P. W. Davis, “Power system state estimation with measurement deficiency: an algorithm that determines the maximal observable subnetwork,” IEEE Transactions on Power Apparatus and Systems, vol. 101, no. 9, pp. 3044–3052, 1982.View at: Google Scholar
S. Nguyen and C. Dupuis, “An efficient method for computing traffic equilibria in networks with asymmetric transportation costs,” Transportation Science, vol. 18, no. 2, pp. 185–202, 1984.View at: Google Scholar