Abstract

This paper examines the design of minimal-order residual generators for the purpose of detecting and isolating actuator and/or component faults in dynamical systems. We first derive existence conditions and design residual generators using only first-order observers to detect and identify the faults. When the first-order functional observers do not exist, then based on a parametric approach to the solution of a generalized Sylvester matrix equation, we develop systematic procedures for designing residual generators utilizing minimal-order functional observers. Our design approach gives lower-order residual generators than existing results in the literature. The advantages for having such lower-order residual generators are obvious from the economical and practical points of view as cost saving and simplicity in implementation can be achieved, particularly when dealing with high-order complex systems. Numerical examples are given to illustrate the proposed fault detection and isolation schemes. In all of the numerical examples, we design minimum-order residual generators to effectively detect and isolate actuator and/or component faults in the system.

1. Introduction

Industrial systems have become more and more complicated and expensive, and thus the requirement of safety and reliability in system operations is paramount [1]. In fact, high complex modern engineering systems are vulnerable to unavoidable faults. Faults can happen due to an internal event, change in environmental conditions, error in the design of the system, sensor failure, actuator malfunctioning, or even human mistakes during operation. The unexpected faults disrupt the system operation, break down part of or the whole system, and can lead to fatal consequences. Fault detection and isolation (FDI) providing failure signals, therefore, have been considered as a vital aspect of system control research. This research issue has attracted extensive studies to safeguard equipment as well as amend the safety and reliability of modern system performance. Since the early 1970s, there have been a large number of fruitful results in FDI which can be found in, for example, [1–30] and the references therein.

In fault diagnosis area, there are two well-known model based fault diagnosis approaches. The first is the fault estimation based approach (see, e.g., [3–8]). By this approach, the systems are decoupled into fault-free and fault dependent parts. The faults can then be estimated via the observers of the fault-free part. The other is the residual-based approach which makes use of state observers to generate diagnostic signals or, in other words, residual generators. In [15], the authors provided a useful comparison between the two approaches. This study is interesting as it provides insightful knowledge and some shortcomings of each approach. The philosophy behind the residual-based approach is to estimate the system state vector based on the control inputs and the measured outputs. The residual generators are then constructed by a properly weighted output estimation error. The residuals are able to detect and identify the fault happening in the systems [9–14]. Obviously, the residual signals are expected to be close to zero in a fault-free condition and deviate from zero in the presence of a fault. For FDI purpose, a decision rule which is normally based on a threshold set is engaged to test for the likelihood of faults occurring in the systems.

Regarding the residual-based approach, the residual signals in most studies are generated via primarily employing full-order state observers or filter schemes which can be found in, for instance, [16–25]. However, the design and implementation of the residual generators using full-order state observers normally require high degrees of complexity and computational work, especially when dealing with high-order complex systems. As stated in [30], in certain applications such as fault diagnostic or control system design, an estimation of the entire state vector is not necessary. Meanwhile, there have been some FDI schemes based on reduced-order state observers (see, e.g., [26–29]). In these schemes, the residual generators are constructed based on the observers of the partial state vector. By that a significantly lower order of the residual generators, in comparison with the ones employing full-order state observers, can be achieved. With an attempt to simplify the design of the residual generators, this trend is identified as very important from the practical point of view (see [31–35]).

In this paper, by using minimal-order functional observers, we present a new FDI scheme to detect and identify unpredictable actuator and/or component fault signals entering from the inputs of dynamical systems. The systems considered in this paper have state variables, measured outputs, and control inputs. Our approach to FDI utilizes the advantages of functional observer schemes which can be found in [36–41]. Due to the fact that functional observers are employed instead of full-order state observers (or, for that matter, reduced-order Luenberger state observers) to generate the residual signals, substantial reduction in the complexity of the overall designed FDI schemes can now be consequently achieved. Particularly, we first present a simple solution approach where residual generators are designed using only first-order functional observers to detect and isolate faults in the systems. When first-order functional observers are not possible to design, then the residual generators are designed based on minimal-order functional observers. To minimize the observer order, we employ a parametric technique to the solutions of a generalized Sylvester matrix equation appearing in the existence conditions. Regarding the timely fault detection (FD), designing the residuals only involves the functional observers with the order as low as , where denotes the smallest integer larger than . This is a new finding and has not yet been reported in the literature. The result is significant since it is clear that, for , the order of the residual generator is much lower than any of the existing results in the literature. Furthermore, for timely isolation of independent actuator and/or component faults in the system, we propose to construct a bank of residual generators, each with the order as low as . Thus, it is clear that our proposed FD and FI schemes, taking advantage of functional observers, are identified as most beneficial for complex large-scale systems where, by default, and are large values.

The organization of this paper is as follows. In the next section (Section 2), we present the system description and preliminaries where we consider linear dynamical systems with unpredictable actuator and/or component fault signals entering from the system inputs. This section also introduces a novel residual generator that is constructed from reduced-order functional observers. Section 3 presents a detailed analysis on the design of minimal-order residual generators to trigger, in a timely manner, any fault that enters into the system. This follows by Section 4 where a bank of residual generators is designed to isolate likely faults. To achieve this, each residual generator is designed, using first-order or minimal-order functional observers, to be insensitive to one specific fault but sensitive to the remaining faults. Subsequently, a logic table can be drawn up from the residual generator outputs to detect and isolate the faults. Section 5 presents extensive numerical examples to highlight the attractive features of our proposed FDI schemes. Finally, Section 6 concludes the paper.

2. System Description and Preliminaries

In this paper, we consider dynamical systems with outputs and inputs and with unpredictable actuator fault signals entering from the system inputs. In state space models, the systems are governed by the following equations:where , , and are the state, input, and output vectors, respectively. , , and are known constant system matrices. is the fault identity matrix. We assume that , without loss of generality, the fault identity matrix is also assumed to be a known full-column rank matrix, and the faults , , are linearly independent. This assumption avoids vagueness which may appear when some faults occur simultaneously; as a result, the residual generators may not detect these faults due to the zero overall effect of these faults [26].

Let us consider the following reduced-order functional observer:where , , , , and are observer parameters to be determined such that is an asymptotic estimate of a linear function when no fault appears in the system, that is, , and is a matrix to be determined for the purpose of FDI.

We now define a residual generator, , which is used to trigger the faults in the system:where and are residual parameters.

Figure 1 shows block-diagram implementation of the residual generator as defined in (2)-(3). By that, only known information of the inputs, , and the outputs, , of the system is utilized to generate the residual generator .

Figure 1 

Block-diagram implementation of a residual generator.

Obviously, regarding the FD purpose, parameters , , , , , and should be determined to meet the required functions of the residual generator such thatwhere , implies a faultless condition, and implies a faulty condition.

Error dynamics , which is defined to be the difference between the estimation and the linear function , is expressed as follows:

By taking the derivative of the error vector (5) and then substituting from (2) and from (1), the error dynamics can be written as follows:

It is clear from (6) that, under the faultless condition, the error asymptotically converges to zero if the following conditions hold.

Proposition 1. Under faultless conditions, that is, , is an asymptotic estimate of (i.e., asymptotically) for any initial conditions and if

Proof. When no fault occurs in the system, if conditions (8) and (9) are satisfied then (6) is reduced to . Thus, if is Hurwitz then asymptotically. Note also that if all the eigenvalues of can be arbitrarily assigned, then with any prescribed rate. This completes the proof of Proposition 1.

Definition 2. The functional observer (2) is an Asymptotic Estimator if is an asymptotic estimate of when .

Since Proposition 1 is satisfied and based on Definition 2, the error dynamics of an asymptotic observer takes the following form:

Substituting (5) into (3), the residual generator can be obtained as follows:

According to (4) and (11), the residual generator can satisfy its proposed functions if the following conditions hold.

Proposition 3. An Asymptotic Estimator of form (2) can satisfy according to (4) for any and if

Proof. If and also (12) and (13) are satisfied then according to (10) we have asymptotically and according to (11) we have asymptotically. When a fault appeared in the system, that is, , and condition (13) holds, based on (10) we have . Accordingly, the residual signals (11) deviate from zero, which indicates the fault happened in the system. Thus, the required functions of the residual generator as stated in (4) are satisfied. This completes the proof of Proposition 3.

Now, the design of the functional observer (2) and the residual generator (3) reduces to determining the unknown matrices , , , , , and such that conditions (7)–(9), (12), and (13) are satisfied.

3. Fault Detection Scheme

In Section 2, we showed that the functional observer (2) and the residual generator (3) can detect faults in the system if all the unknown parameters satisfying the conditions stated in Propositions 1 and 3 are found. Observe that if is known, then from (9), , and condition (13), that is, , can be easily verified. In this section, we propose a systematic procedure for solving coupled matrix equations (8) and (12) with the requirement that matrix is Hurwitz. Note that, here, not only do we require that matrix be Hurwitz, but also we desire its eigenvalues to be placed at some prespecified locations in the -plane in order to achieve timely detection of faults.

Let us first simplify matrix equations (8) and (12) by a partition technique defined in [38]. Regarding this, an invertible matrix is introduced:where denotes the Moore-Penrose inverse of , that is, , and denotes an orthogonal basis for the null-space of , that is, . Now the following partitions are defined:where submatrices , , , , , and .

Postmultiplying both sides of (8) and (12) by the matrix , we obtain

By substituting (15)–(17) into (18), we achieve the following equations:

It is clear from (19) and (20) that matrices and can be computed once , , , and are found to satisfy (21) and (22).

In the following, we discuss the solutions to coupled matrix equations (21) and (22). Observe that the solution to (21) depends on the characteristic of matrix . Hence, we classify into two cases, namely, Case 1 where is a column matrix and Case 2 where is a row matrix.

Case 1 (first-order residual generators). In this case, we consider that ; accordingly is a column matrix. For this, we only need to design a first-order residual generator to detect faults in the system. Regarding this, the design procedure is detailed as follows.

Theorem 4. A first-order functional observer (2) always exists for system (1) when is a column matrix; that is, . Furthermore, a residual generator (3) can be constructed to detect faults in the system if condition (13) is satisfied.

Proof. To design a first-order functional observer, that is, , can be chosen to be any negative scalar. Thus, letBy letting , (21) is reduced to the following:Since is a column matrix, a solution to (24) where always exists. Let be a matrix of row basis vectors for the row-null-space of ; that is, . Therefore, the solution for in (24) can be achieved by first computing according toand then can be selected as any row of .
Matrix in (22) can be arbitrarily chosen to be any nonzero scalar, say, , since . Finally, if condition (13) is satisfied, a residual generator using a first-order functional observer exists to detect faults in the system. Matrices , , and can then be easily obtained from (9), (19), and (20), respectively, to complete the design of the residual generator. This completes the proof of Theorem 4.

It is worth pointing out that, in the derivation of a first-order residual generator, we do not impose the requirement that the matrix pair is observable.

Case 2 (minimal-order residual generators). In this case, we consider that ; accordingly is a row matrix. Matrix is considered to be full rank; that is, . For this, we design a minimal-order residual generator for FD by presenting a solution to coupled matrix equations (21) and (22) via a parametric technique. Here, has preassigned distinct eigenvalues satisfying condition (7). For completeness, let us first present a parametric solution [42] to generalized Sylvester matrix equation (21). For the solution of (21), we require that the pair be observable. This implies that the pair is observable [38].
Let with distinct eigenvalues be defined as follows:where is any freely chosen invertible matrix and , for and for all . With as defined in (26), and satisfying (21) are given in the following parametric forms [42]:where are free vectors satisfying if . denotes the complex conjugate of . and are coprime polynomial matrices satisfying the following coprime factorization:The reader can refer to [42] for a numerically reliable algorithm to compute and . Also, as suggested in [38], and can be conveniently computed according to the following equations:where and denote the determinant and the adjugate matrix of matrix , respectively. For any given , the characteristic polynomial can be obtained:where the coefficients , , are real constants. The adjugate matrix is then obtained as follows:where , , are computed by using the coefficients of and matrix , whereNote that if the preassigned eigenvalues of are complex then the resulting matrices , , and are also complex. As reported in [14], we can obtain real matrices , , and by applying the following simple computation. Here, without loss of generality, we assume that and , , are a pair of the eigenvalues containing complex values, and all other eigenvalues of are real and distinct. The real matrices , , and can be obtained as follows:where and denote the real and the imaginary parts of , respectively.

Now, by incorporating as defined in (28) into (22), we can solve for . Let us now present the result for Case 2 by the following theorem.

Theorem 5. With and , a residual generator (3) with an order as low as can be constructed to detect faults in the system if condition (13) is satisfied.

Proof. First, let us express as follows:where , , are arbitrarily real numbers.
Once are scalars, by substituting (28) and (36) into the transpose of (22) and after some rearrangement the following equation is obtained:Now, (37) can be expressed as follows:where and andSimilar to the solution to (24), let be defined such that ; then a solution to (38) always exists for if . Accordingly, if is a row matrix, that is, the number of columns is greater than the number of rows, the solution to always exists. That implies the following result:and then exists and can be taken to be any column of .
As a result, with prescribed eigenvalues for (i.e., , , are given) and for any arbitrary invertible matrix , matrices , , and are computed based on (26), (27), and (28), respectively. Matrix is then calculated by (16) and condition (13) can be verified. If (13) is satisfied, the residual generator can detect faults in the system. Finally, , , , and are obtained from (9), (19), (20), and (36), respectively, to complete the design of the observer and the residual . This completes the proof of Theorem 5.

Remark 6. The observer order can be assigned to satisfy condition (40), where a solution to (38) always exists for . Thus, for the purpose of designing minimal-order observers, we only need to choose .

Remark 7. External disturbances or the uncertainties are unavoidable during the operation of the systems. If we take disturbances in the system where the disturbance matrix is defined as and the disturbance vector is , system (1) now can be governed by the following equations:

Based on the disturbance decoupling approach [7], the influence of the unknown disturbance can be eliminated. Using this approach, the conditions in Proposition 1 of this paper are as follows:

Since the condition is added, the solutions to the observer parameters will be more strict. Nevertheless, based on the solution approach in this paper, it is possible to design minimal-order observer-based residual generators to detect faults in the systems with unknown disturbances. In this case, the residual functions are also defined as (4) which means the threshold can be set as zero. In addition, it is also possible to extend the result to include nonlinearities in the systems. This topic deserves further research and will be a subject for future research. Another issue that deserves further research is related to the robust threshold selection. How to solve the robust threshold selection problem under uncertainties is an important research question and the reader is referred to the most recent and interesting research [43].

Remark 8. We conclude this section by considering the case where and . Note that, for this case, a first-order residual generator can be designed since a solution to (24) where exists due to the fact that . Indeed, there exists such that . Accordingly, is obtained and the rest of the observer and residual generator parameters can be derived in the same way as reported in Case 1. A numerical example can be easily constructed to demonstrate this case. However, due to space limiting reason, a numerical example is not given in this paper.

4. Fault Isolation Scheme

In Section 3, we discussed the design of minimal-order residual generators for FD in the systems. However, identifying the faults is also considered as an important task in system control since it provides the fault positions [15]. The information about the fault positions is then sent to the decision-maker for further actions dealing with the faults. In this regard, we present a systematic procedure to construct a bank of minimal-order residual generators which is used to isolate the faults. Hence, our proposed FI scheme significantly reduces the complexity in the implementation. The scheme is thus more suited to deal with complex high-order systems. Before showing the procedure, let us mention a logic for the fault isolation from the following [44].

We consider, in this paper, that there may be likely faults occurring in the system. Accordingly, for the purpose of FI, we propose to design a bank of residual generators, in which each residual generator is expected to be insensitive to one specific fault but sensitive to the rest of the remaining faults. Thus, a th residual generator, , is designed to be insensitive to the th fault, that is, , and sensitive to the rest of the remaining faults. By that, when there is no fault in the system, the outputs of the residual generators remain as zero. However, when a fault, say, , has occurred, the output of the th residual generator remains as zero; meanwhile, the outputs of the other residual generators are not zero but turn to , where is as defined in (52). Consequently, by measuring the output values of the residual generators or using a set of residual thresholds, when a fault appears, we can identify the position of the fault.

For simpler explaining of the logic, let us take an example where there are four possible faults, that is, , , in the system. Accordingly, a bank of four minimal-order residual generators, , , is designed to isolate the faults. Let us use a logic “1” to indicate that the output of a residual generator is bigger than a threshold value, that is, , and a logic “0” to indicate that the output of a residual generator remains as zero (or less than a threshold value). Table 1 shows the logic for the fault isolation. In the table, for instance, when a fault has occurred, residual is insensitive to ; hence it appears as “0.” Meanwhile, other residuals can trigger so they are expressed as “1.” Therefore, we can identify that fault has occurred. However, it is noted that this logic is only correct when the faults happen independently and also not simultaneously. In Section 2, we have made an assumption that avoids such a situation and hence the logic expressed in Table 1 can thus be used to isolate the faults in the system.

Now, by utilizing the above FI logic and for the rest of this section, we present systematic procedures to design each residual generator in the residual bank, , with minimal order. For this, let us rearrange and partition the fault vector, , and the fault identity matrix, , as follows:where is the th fault, contains the remaining faults, , and .

Thus, system (1) can now be rewritten as

Similar to Section 2, let us now consider a reduced-order functional observer for the purpose of fault isolation:where , , , , and are observer parameters to be determined such that is an asymptotic estimate of the linear function when and is a matrix to be determined for the purpose of fault isolation.

Let an error vector be defined as follows:By taking the derivative of (46) and using (44) and (45), the following error dynamics is obtained:

Proposition 9. Under no-fault conditions, is an asymptotic estimate of the function (i.e., asymptotically) for any and if

Proof. The proof of Proposition 9 is the same as the proof of Proposition 1; thus it is omitted here.

By using the estimate and the system outputs, , the residual generator , which is insensitive to fault and sensitive to the other faults, , is now proposed:where , , and .

Since faults occur independently, matrices and need to be found to satisfy the following functions of :where , means that, except fault which may or may not happen, other faults do not happen, and means that any of the faults can happen in the system.

If all the conditions in Proposition 9 are satisfied, from (45), (46), and (51), the residual can be governed in the following equations:

Proposition 10. The residual generator is insensitive to fault and sensitive to faults for any and if

Proof. The proof can be constructed by following similar lines as in the proof of Proposition 3 and thus it is omitted here.

From Propositions 9 and 10, we can now determine the unknown parameters , , , , , and to complete the design of the residual generator . Clearly, matrix can be calculated from (50) and condition (56) can be easily verified if is found. Therefore, the residual generator design now reduces to find the solutions to unknown matrices , , , , and which satisfy three conditions (49), (54), and (55). Here, we again assign stable eigenvalues to and then the fault isolation scheme can be achieved in the following sections.

As in Section 3, let us use matrix defined in (14) and the partitions in (15) and (17) and the following partitions:where , , , and .

Accordingly, matrix equations (49), (54), and (55) can be partitioned to the following: