Informacion tomada de

Liu, L., “Robust fault detection and diagnosis for permanent magnet synchronous motors,” Ph.D. dissertation, Dept. Mech. Eng., Florida State University, Tallahassee, FL, 2006.


Fault Detection and Diagnosis

Faults in engineering systems are difficult to avoid. In complex systems, any fault possesses the potential to impact the entire system’s behavior. In a manufacturing process, a simple fault may result in off specification products, higher operation costs, shutdown of production lines, and environmental damage, etc. In a continuously operated system, ignoring a small fault can lead to disastrous consequences. For instance, a position sensor fault in an automated guided vehicle (AGV) may bring the vehicle towards hazardous directions or cause a collision. Without taking these fault conditions into consideration, fully automated systems are not reliable. To improve reliability of automated systems Blanke, et al proposed the concept of fault tolerable and reconfigurable systems in [1][2]. Figure 1.1 illustrates the three layer architecture for a fault tolerant control (FTC) system. The bottom level comprises traditional control loop with sensors, actuators, signal processing, and controllers. The middle level consists of fault detection, diagnosis, and accommodation. Fault accommodation is performed through reconfiguration or other remedial actions initiated by the autonomous supervisor, which is at the top level. Though this architecture separates these functions into three independent layers, they are highly coupled. Located in the middle, FDD plays an important role by linking the other parts into an integrated system. Thus FDD becomes an essential component in a robust FTC system. Routinely, particularly more than five years ago, fault detection was implemented using hardware redundancy, i.e. installing multiple sensors, actuators, and other components into the original system. Many measurements from an identical component are then compared for consistency. Any discrepancy among these measurements indicates a fault in that component. According to Frank (1990) [3], Gertler (1998) [4], Patton, Frank and Clark (1989) [5], hardware redundancy methods have significant disadvantages, though being theoretically simple, and also due to the requirement of many safety and reliability standards. The redundant hardware results in a more expensive and complex system. Also, the increased size and weight of the extra hardware is not always acceptable in some special environments, such as marine and aerospace applications. Due to these drawbacks, hardware redundancy is most used to evaluate sensor conditions to improve sensing and control robustness. Frank, Patton, Clark and Angeli claimed in [3][5][6] that hardware redundancy is not capable of detecting anomalies inside monitored process. However, detection and diagnosis of anomalies, such as a system internal fault, is very important to obtain a robust FTC system.

i. As asserted by Gertler and Patton.Figure 1. As claimed by Frank (1990) [3] and Kinnaert (2003) [7].e. model based methods with analytical redundancy are becoming an attractive approach to FDD for FTC. While analytical redundancy is more capable and advantageous to hardware redundancy in many ways. Model based approaches rely on analytical redundancy. The acquisition of accurate nonlinear system models is challenging. et al in [4] and [5]. this method brings its own issues. If the nonlinear system’s operating points deviate from the linearization points. Model based fault detection and diagnosis have been extensively explored for systems represented by linear models. One of the most important problems is robustness to modeling errors.1 A three layer fault tolerant control architecture To overcome the shortcomings of hardware redundancy. etc. Though desired. The main idea is to check the consistency between system measurements and the computed values based on an analytical model. Fault detection and diagnosis based on an inaccurate model may lead to wrong detection results. especially when taking the fault conditions into account. nonlinear models with high accuracy are preferable to represent systems with large dynamic range. weight. the model can only be used about some particular linearization point. However. With the development of advanced modeling techniques. explicit mathematical models of the monitored system. model based methods are advantageous in many aspects. size. model based approaches can be used to detect not only sensor faults but also problems inside of the system. various noise effects. most practical engineering systems are nonlinear in nature. recent FDD development has focused mainly on model based techniques. the linear model can no longer represent the system accurately. such as cost. it is unrealistic to expect perfect modeling for complex systems. FDD based on nonlinear models is an area requiring further explorations. as opposed to physical redundancy. and uncertain or time varying system parameters can cause significant deviation in component output from that of its model. . For instance. As analytical modeling provides insights to system dynamics. by replacing the extra hardware with mathematical models. Early detection of a system internal fault can greatly improve system reliability by avoiding disastrous failures. Hence. disturbances of an unknown structure. robustness has been the most important requirement for successful fault detection and diagnosis since the initial development of model based approaches. The causes for errors are diverse. When a nonlinear system is described by a linear model.

However. fault identification determines the size and general characteristics of that fault [4]. In this dissertation. location. Components within this system are susceptible to various faults. Later on. Systematic development of automatic FDD did not come until the widespread applications of computers. and noises. the tasks of fault detection and fault diagnosis are interconnected to each other.1. Gertler (1998) defined the task of fault diagnosis as fault isolation and fault identification. In simple cases. there is still a large gap between the established theories and practical applications. measuring devices were introduced that provided more accurate information about important variables. These terms are consistent with Isermann et al defined in [12].1. Numerous methods for FDD have been published. Just “What is fault diagnosis?” has not been standardized. As defined by Iserman and Ulieru (1993). and second to further diagnose the fault based on the observed analytical symptoms. such as shutdown for on demand maintenance. Although definitions are clearly given for different functions. 2. A typical FDD framework in applications is shown in Figure 2.2 FDD Terminologies Before giving more details on FDD theories. But these devices were also prone to malfunction. First. humans have been concerned about these machines’ conditions. especially for the FDD of nonlinear systems. the operating status of the system is monitored by sensors.1 Introduction Since the emergence of machines. In this diagram. A FDD module takes measured data to detect and diagnose the faults. Fault diagnosis usually consists of identifying the fault’s type. fault isolation specifies the sources associated with the detected fault.1. and size. a good design of symptoms (residuals) generation for fault detection purpose can greatly facilitate fault diagnosis in later stages. etc. fault isolation and fault identification are integrated together as fault diagnosis. to detect whether a fault appears in the system by analytical symptom generation. the main task of FDD can be subdivided into two parts [12]. continuous operation with automatic reconfiguration. Such a control system with effective FDD design can greatly improve a system’s reliability and availability. For instance. The primitive FDD were based on observations of machines’ operating conditions. it is helpful to introduce other important definitions for the technical terms used in this field. disturbances. These terms are defined by the FDD . such as availability and reliability.2. Therein. The diagnosis results are then fed back to the control system for corrective actions.

the consequences of a failure are usually more serious.. see [4][5][13][14][15][16] for references. etc. 2. These influences are undesirable because they may result in false alarms. malfunctions. Besides faults. Isermann.initiators. the proposed FDD algorithms were applied to chemical plants and airplanes. symptom means a sign or an indication of disorder or disease. residual is a fault indicator constructed based on deviation between a system’s actual and its normal behaviors. Meanwhile. Residuals caused by different faults need to be distinguishable from each other. Considering the relatively small magnitude and slowly time varying properties. Since the research on FDD was initiated. a fault can be classified into two types. Based on a fault’s influence on system models. According to Figure 2. soft faults are more difficult to detect than hard faults. Faults in a system may lead to degraded performance. The most typical additive faults are sensor offsets. usually refers to slow changes of system’s parameters or drift of measured data caused by sensors problems. In medicine. A hard fault. In these papers. Gertler. such as system operation changes. additive faults and multiplicative faults. hard faults are usually simulated with a step change in inputs or system parameters. an additive fault influences a variable Y by a simple addition of the fault f. whereas multiplicative faults are parameter changes within a system. Gertler summarized the main features of model based FDD and presented the robustness and sensitivity considerations in [21]. Basseville presented both online and offline FDD algorithms in his survey paper with . in their published books and some well recognized papers. Based on its severity and timeliness. as well as Lainiotis and Park (1973) [17][18]. False alarms can be avoided by robust FDD design. In FDD. During simulation. The same word is utilized in FDD to represent the characteristics resulting from a fault in the system.g.2 Model Based Fault Detection and Diagnosis Study of model based FDD began in 1970’s. By definition.2. Isermann reviewed FDD design based on modeling and estimation method in 1984 [20]. or failures. sometimes called incipient fault. sometimes called abrupt fault. sensor noises. In 1988. such as partial or complete system breakdown. may also exert influence to residuals. especially on linear time invariant (LTI) systems. A fault is defined as “an unpermitted deviation of at least one characteristic property of a variable from an acceptable behavior” according to Isermann [14]. Willsky wrote the first major survey paper on model based FDD [19]. faults can be categorized into two types. many other factors. Different from fault. etc. Frank. In 1976. Symptoms of a fault are usually reflected as deviations of a signal from its normal value. A multiplicative fault has more complicated impact upon a system by the product of the variable U with f. FDD is generally considered to have entered its modern era with the publications of Mehra and Peschon (1971). Patton. Clark. e. A soft fault. After that. disturbances. fault symptoms are usually evaluated by residuals. means an abrupt complete loss in sensing signals or abrupt changes of other type functions.. Incipient faults may indicate potential malfunctions. robustness has been an important goal.

This 1997 paper focused on fuzzy logic and introduced the concept of knowledge observer. a well structured residual set can be designed as (2.. In general.a particular emphasis on statistical methods for detection [22]. A review of these methods is given in the following subsections. some nonlinear model based FDD methods were proposed. an ideal residual should follow (2.1) indicates that r(t) should be zero when the fault f(t) does not occur.2). residuals should be designed insensitive to disturbances. This topic was surveyed by Frank in 1990 [3]. and other uncertainties. Then. a set of residuals must be designed with each residual responding only to a specific subset of faults. as shown in Figure 2. many algorithms have been proposed for residual generation. fi(t) denotes the ith fault. With the development of Artificial Intelligence (AI). a survey on which was made by García and Frank [24]. and be nonzero otherwise. the residual generator uses system measurements as input to generate residuals as fault symptoms. and r(t) represents the residual designed for the detection of that fault. This design means that residual ri(t) only reponds to fi(t). In addition. when a number of n residuals are designed to detect n faults in a system.2). According to selected FDD references. to distinguish different faults.1) for fault detection purpose: where f(t) represents a fault. Gertler introduced the design of structured residuals for LTI systems in [4]. diagnostic observers. Frank and Clark (1989) [5]. Equation (2. . Residual generator plays an important role in model based FDD. In (2. these residuals are sent to the residual evaluator to make diagnosis decisions. in an effective FDD system. and remain zero as long as this fault does not occur. Frank and Köppen-Seliger did a followup survey on AI based FDD in [23]. Gertler (1998) [4] and Patton. For example. This design is usually called structured residual. a model based FDD system consists of two modules: residual generator and residual evaluator. In the mid 1990s.3. Therefore. a lot of AI based FDD designs were proposed. These algorithms can usually be classified into three categories. e. and ri(t) denotes the ith residual. In the past thirty years. In this diagram. which are parity relations.g. noise. In 1997. while remaining sensitive to faults. etc. and parameter estimations.

In [16]. In this figure. 2.5. by Chow and Willsky [13]. While being the most computationally efficient and simplest in implementation. diagnostic observers also require accurate knowledge of system model . Since its computational effort is relatively small. Parity relations directly check the consistency between the model and measured system outputs. Diagnostic observers utilize the errors between estimated and measured outputs as residuals.1 Fault detection with parity relations Parity relation sometimes is also called consistency relation. The estimated states can then be used to observe the system outputs.2. Isermann claimed that parity relation is the most straightforward way to detect additive faults. parity relation is the predominant FDD algorithm for additive faults. y denotes the measured system output. Lou. The schematic diagram of this approach is shown in Figure 2. This method is especially useful when dealing with the cases where states or outputs are not measurable.3) with u denoting the control signal input. Some representative work can be found in [26][27][28]. GP represents the monitored system. the method has difficulty identifying parameter changes caused by multiplicative faults. To overcome the problem of imprecise models. robust parity relations were introduced in the mid 1980s.2. Similar to parity relation method. which can be calculated according to (2. This approach was first introduced by Frank [3] and Geltler [4].4).2 Fault detection with diagnostic observers If a system is observable. and GM is the analytical model of that system. ye denotes the estimated system output.2. The main idea of this method is shown in Figure 2. However. residuals based on this method are susceptible to model inaccuracies due to system uncertainties.4. and r denotes the residual. thus is the most intuitive fault detection method. Residual r based on parity relations can then be mathematically expressed in (2. the parity relation based approach assumes that the model parameters are known and constant. observers can be used to estimate states of a system using measured data. This method is very sensitive to abrupt additive faults. Diagnostic observer design takes advantages of widely used observer designs. Willsky and Verghese [25]. Thus. or analytical redundancy relation.

6. and Isermann [33][34].2. For some problems. parameter estimation based fault detection is used for FDD. Parameter estimation methods only require the structure of the system model to be known. p represents pe’s nominal value. The parameters of the model are estimated and compared to their values during normal operating conditions. changes in model parameters are directly related to system internal faults. Hofling and Isermann (1995) proved the mathematical equivalence between the parameter estimation and the parity relation with an accurate model. A disadvantage of this method is that a Persistent Excitation (PE) condition is usually required. the parameters of a system model may normally vary continuously in certain ranges. and r represents the residual as calculated in (2. Also. Isermann (1991) states that parameter estimation provides deeper insights into system dynamics and thus makes fault diagnosis easier. In many cases.parameters. Parameter estimation is proved to be advantageous in this regard according to Gertler. Gertler (1995). pe represents the estimated parameter.3 Fault detection with parameter estimation Both the parity relations and the diagnostic observers require an accurate model of the system. parameter estimation allows the detection of very small changes of the system conditions. the parameter estimation approach is especially suitable for the detection of the multiplicative faults according to Gertler (1998) and Isermann (1984. In [33][34]. Some related work can be found in [17][30][31][32]. Since accurate system models are often unavailable for practical systems.5): In [35]. 2. 2005) [4][14][20]. Since the above two methods cannot handle this problem. In this figure. Hofling. Thus. The relationship between diagnostic observers and parity relation based designs was discussed by Mgni and Mouyon for linear systems [29]. robust designs are important for diagnostic observers. including slowly developing faults. . This parameter estimation based fault detection method is illustrated in Figure 2.

expert systems. Representative papers in this aspect are found. raising the threshold hinders the detection of small incipient faults. This is a major concern for high performance applications. Active approaches are applied to the residual generator. It is a difficult task especially for complex nonlinear systems. e. Passive approaches are applied to the residual evaluator. The lack of accuracy makes the detection of small incipient faults difficult. instead of quantitative models in their study on nonlinear model based FDD. Development of quantitative models usually requires explicit description of systems’ dynamic behaviors. The objective is to minimize false alarms by using proper decision making logic from the noised residual [41]. 2. If the system is operating away from the linearization point. i. Techniques used in qualitative models based FDD include fuzzy logic. The objective is to minimize the influence of the model impreciseness on residuals [4]. In recent FDD papers. But linearization of nonlinear models will lead to modeling errors since the linearized model is only valid around some linearization point [4]. et al (1992) [36]. These methods can be broadly classified into two categories.4 Application Issues of Model Based FDD The above methods are relatively mature for linear systems but not for nonlinear systems.2. To avoid this problem. FDD techniques based on quantitative models usually provide more accurate diagnosis but are very sensitive to modeling errors. . several approaches have been proposed. qualitative models are used to describe system variables with a finite number of modalities and sets of rules. quantitative model based FDD techniques are still necessary. the performance of the linear system based FDD algorithm cannot be guaranteed. researchers tend to utilize qualitative. Since practical systems are usually nonlinear in nature. statistical approaches.Since these three residual generation approaches have their pros and cons. Modeling errors may lead to false alarms.g. Qualitative models are generally an incomplete or fuzzy description of the actual system. together with expert knowledge or experience. Linearization may be a possible way to apply linear system based FDD algorithms to nonlinear systems. et al (1995) [38]. the FDD techniques developed for linear systems cannot be extended to nonlinear systems directly. et al (1995) [37].e. To improve robustness of FDD designs. Goode. Vinson. Glass. active and passive approaches. Qualitative models can be built upon relatively simple understanding about a system. a combination of two or more may provide a better solution. as summarized by Staroswiecki (2000) [39]. To improve FDD system performance especially in high performance applications. etc. However. often fuzzy ones. A simple way to decrease the false alarm rate is to increase the threshold for fault decision making. and some other AI algorithms. Venkatasubramanian wrote a good review on this subject in 2003 [40].

vol. pp. Proceedings SafeProcess. New York. and M. Issues of Fault Diagnosis for Dynamic Systems. 1. UK. S. 25-27 June 2001. Chow.” IEEE Control Systems Magazine. 1. MARCEL DEKKER Inc. “Concepts and methods in faulttolerant control. 2005. no. 37. pp. Clark. 1984. vol. Angeli.”International Journal of Computer Science and Applications . pp. vol. Fault Diagnosis in Dynamics Systems.” IEEE proceedings of the 2001 American Control Conference. vol.” IEEE Trans. 1988. P. “Trends in the Application of Model based Fault Detection and Diagnosis of Technical Processes. IL. P.” Automatica. a good FDD design should address both of the above two problems. Bianchi. i. Estimation And System Identification: Discrete Data Case. 709-719. pp. “Process Fault Detection Based on Modeling and Estimation Methods – A Survey. pp. 3. “Model based fault-detection and diagnosis – status and applications. Theory and Applications. pp. Diagnosis and Prognostics:Software Agent Solutions. 29. S. UK. and A. McCoy. S. N. vol. 26. K. “Analytical redundancy and the design of robust failure detection systems. Prentice Hall. Man and Cybernetics. 20.. Blanke. pp. and D. Lainiotis. Automatic Control. 37-51. no. vol. and M. no. Electromechanical Systems. Willsky.A holistic view. 609-633. Wu . 1997. C.” Automatica. Kinnaert M. 1. (USA). Bolognani. M. R. 1990. Frank and R. pp. L. 2606 – 2620. A. 459-474. K. E. Isermann. “A Survey of Design Methods for Failure Detection in Dynamic Systems. no. 3-11. “Online Fault Detection Techniques for Technical Systems: A Survey. Frank. and C. “An Innovations Approach To Fault Detection And Diagnosis to Dynamic Systems. M. 1989. P. 637-640. “Joint Detection.” IEEE Power Engineering Society Summer Meeting. 5. vol. no. Fault Detection and Diagnosis in Engineering Systems . J. vol. A.” 1st IEEE Symposium on Electric Ship Technology (ESTS). no. 26-27 Jul. 2001. 2000. M. Lyshevski. R. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] M. Blanke. K. R. Peschon. Ulieru.e. 17. Springer-Verlag. Patton. Gertler. CRC Press LLC. Isermann. 601-611. vol. “Survey of Model Based Failure Detection and Isolation in Complex Plants. Cartes. N.” Automatica. 1997. pp. 2003. pp 693-702. P. 4. – Oct. vol. 5. E. M. Frank. chapter 7. Clark. Lunau.” Control Engineering Practice. 7. Isermann. London. vol. Ind. USA. 12. and N.743 – 748. M. 6. and R. pp. 1348-1355. “Fault Detection. 343-346. 7. pp. S. J.” International Journal Of Control. Isermann.” Control Engineering Practice. pp. 12-30. vol. Balle. Zigliotto. 5. “Fault Diagnosis Based On Analytical Models For Linear And Nonlinear Systems . no. Mehra. and J.. 1976. J. 2002. Willsky. “Integrated fault detection and diagnosis. pp. 1993. R. and S. J.” Annual Reviews in Control. Applicat. P. J. Bøgh. 1984. vol. . D. and Applied Mechatronics. E. “Trends in ship electric propulsion. T. Gertler. no. 2005. pp. 5. 8. 5. and P.Summarizing this section. J.1. R. vol. 1971. A. R. 1973.” IEEE Trans. Patton. 387-404. Electric Machines. 2004. Liu. 29. 1. Staroswiecki. “Fault-tolerant control systems . Izadi-Zamanabadi.” IEEE Proceedings on Systems. “Fault Diagnosis In Dynamic Systems Using Analytical and Knowledge Based Redundancy – A Survey and New Results.A Tutorial”. 2000. system nonlinearity and robustness. R. N. “High-performance PM synchronous motor drive for an electrical scooter. 17-20 Oct. J. Chicago. R. Park. vol. Washington. 71-85.” Automatica. G. Sept. Logan.

WA. Seattle. M. 3. Kavuri. “Diagnosing Parametric Faults – From Identification to Parity Relations. 1992. M. 23-30. C. 4. Massoumnia.” IEEE Transactions on Automatic Control. Chen. H.” Automatica. 1997. Venkatasubramanian. pp.” Proceedings of American Control Conference. 1615-1620. pp. 3-14. G.” System and Control Letters. Basseville. T. Hofling.131 – 138. 1986. “Observer Based Fault Detection and Isolation: Robustness and Applications.” IEEE Transactions on Automatic Control. A. “Parameter Estimation Triggered by Continuous Time Parity Equations. pp. Hofling. S. 10.[22] [23] [24] [25] [26] [27] [28] [29] space [30] [31] [32] [33] [34] [35] [36] [37] [38] M. 333-344. V. A. 42. 5. 1995. Willsky. “Estimation Of Physical Parameters For Dynamic Processes With Application to An Industrial Robot. Mechanical Systems and Signal Processing. M. Vinson. D. Clark. 3. vol. vol. 5. 3. vol.” IEEE Intelligent Systems and Their applications. P. J. and Mo-yuen Chow. 1991. 671-682. Mgni. Seattle. “Optimally robust redundancy relations for failure detection in uncertain systems. and P. “A Review Of Process Fault Detection And Diagnosis Part II: Qualitative Models And Search [39] [40] .-F. J. and A. A. Rengaswamy. 2. pp. April 1995. “Detecting Changes in Signals and Systems – A Survey. 1995. “Quantitative and Qualitative Models for Fault Detection and Isolation. Glass. no. Gertler. vol. no. “Failure Detection And Identification. pp. 5. García. Grantham. pp. no. 1997. 316–321.1995. 441 – 447. no. Aug. Feb. and L. 1145-1146. 2000.” Automatica. 663-670.” Control Engineering Practice. “Using a neural/fuzzy system to extract heuristic knowledge of incipient faults in induction motors. and P. 301 – 325. Chen.” Engineering Applications of Artificial Intelligence. J. Portoroz. R. C.” IFAC SafeProcess Conference. vol. vol. no.” Proceedings of American Control Conference.1989. Todtli. and J.1. Germany.” Proceedings of the 18th IEEE Conference on Decision and Control . P. “Qualitative model based fault detection in air-handling units. R.” IEEE Transactions on Industrial Electronics. 1990. 24. Goode. “Deterministic Nonlinear Observer Based Approaches to Fault Diagnosis: A Survey. 1994. no. no. V. 1. Staroswiecki. 14.” Control Engineering Practice. 1988. 12-17. Isermann. and G. Ungar.M. R. and R. pp. “Dedicated observer approach to instrument failure detection. “Automatic rebuilding of qualitative models for diagnosis. “A Survey of Parity Space Approaches to Fault Diagnosis. T. 309-326.” Proceedings of 6th Mediterranean Electrotechnical Conference. 1. Gruber. 237–241. R. and B. Patton.” IEEE Control Systems Magazine. J. no. 11 – 22. X. no. pp.” Mechanical Systems and Signal Processing. R. Frank. 1991. S. N. Part IMethodology. Xi-Cheng Lou. M. 22. “On residual generation by observer and parity approaches. 1994. Frank. 2. pp. Electrotechnical and Computer Conference. pp. “New Developments Using Artificial Intelligence in Fault Diagnosis. and J. Isermann. A. 14. Patton. and P. pp. 15. N. Baden-Baden. vol. J. vol. Willsky. no. and J. pp. Aug. 3. vol. 1979. S. WA. P. 34. vol. vol. no. vol. 7. J. Köppen-Seliger. Roos. Slovenia. and S. pp. pp. “Fault Detection of a DC Motor Using Grey Box Modeling and Continuous Time Parity Equations”. Verghese. S. 5. Frank. 1997. Ding. vol. ERK. pp. vol. pp. E. M. Verghese. 4. 431-436. “Fault Detection via Factorization Approach. M. Mouyon. 39.

. 27.” Proceedings of the 30th IEEE Conference on Decision and Control.” Computers & Chemical Engineering. vol. X. Frank. 1991.1. “An approach to robust residual generation and evaluation. M. 3. and P. Dec. vol.[41] Strategies. 11-13. no. 313-326. pp. 2003. Ding. pp. 656-661.