You are on page 1of 6


11, NOVEMBER 2013


Fault Detection for Modular Multilevel Converters Based on Sliding
Mode Observer
Shuai Shao, Patrick W. Wheeler, Jon C. Clare, and Alan J. Watson

Abstract—This letter presents a fault detection method for modular multilevel converters which is capable of locating a faulty
semiconductor switching device in the circuit. The proposed fault
detection method is based on a sliding mode observer (SMO) and
a switching model of a half-bridge, the approach taken is to conjecture the location of fault, modify the SMO accordingly and then
compare the observed and measured states to verify, or otherwise,
the assumption. This technique requires no additional measurement elements and can easily be implemented in a DSP or microcontroller. The operation and robustness of the fault detection
technique are confirmed by simulation results for the fault condition of a semiconductor switching device appearing as an open
Index Terms—Fault detection, modular multilevel converter
(MMC), sliding mode observer (SMO), switching model.

HE modular multilevel converter (MMC) has drawn considerable interest, as it offers very attractive features
1) modular construction with scalable, manufacturable, standardised cells (half-bridge);
2) submodules are fed by floating dc capacitors, no multipulse transformer is required;
3) high power and high voltage capability, extendable by
adding additional cells;
4) flexible control of the voltage level and simple realization
of redundancy if required.
Fault detection is an important issue for an MMC. When an
open-circuit fault occurs, the output voltage and current are distorted, moreover, the voltages of the dc floating capacitors will
keep increasing, leading to further, vast destruction. Therefore,
it is vital to locate the fault after its occurrence and take measures
such as bypassing the faulty cell to reconfigure the MMC.


Manuscript received June 27, 2012; revised November 6, 2012 and
December 24, 2012; accepted January 15, 2013. Date of current version May 3,
2013. Recommended for publication by Associate Editor J. Jatskevich.
The authors are with the Power Electronics, Machines and Control Group,
School of Electrical and Electronic Engineering, University of Nottingham,
Nottingham NG7 2RD, U.K. (e-mail:; Pat.;; Alan.watson@
Color versions of one or more of the figures in this paper are available online
Digital Object Identifier 10.1109/TPEL.2013.2242093

Given the large numbers of identical cells (half-bridge) and
the symmetrical structure of the converter, the process of fault
location in an MMC is challenging if significant extra cost is
to be avoided. An effective but inefficient way to detect faults
is to add additional sensors to each semiconductor switching
device [5], to each cell [6], or to use a gate drive module capable of detecting faults and providing feedback [7]. These
additional sensors and signals increase not only the cost but
also the implementation complexity. Some conventional fault
detection methods for voltage source converters (VSCs), such
as the calculation of the output current trajectory using Park’s
Vector [8], [9], or comparison of the actual ac voltage and the
reference quantity [10] are, however, not suitable for an MMC,
as there is not enough information present to locate the fault.
Fault diagnosis methods for a cascaded H-bridge (CHB) are
applicable to an MMC because of the similar structure. The
authors in [11] presented a detection approach for a CHB multilevel converter, which analyzes magnitude of the switching
frequency component (vs ) of the output phase voltage: vs becomes significantly larger after the occurrence of a fault due
to the imbalanced cancellation of the switching frequency harmonics. The faulty cell can be located according to the angle
of vs [11]. However, the faulty switching device cannot be located, and it is complex to implement and easy to get the wrong
diagnosis in transient operation. In [12], the authors proposed
an artificial intelligence (AI)-based algorithm to detect the fault
of a CHB, the major drawbacks are the accuracy (only 76% in
some cases) and the long training time required for the circuit
and all the fault scenarios.
This letter proposes a sliding mode observer (SMO)-based
fault detection method for an MMC. The method uses the converter arm currents and the cell capacitor voltages as the inputs,
which are already available as measurement inputs to the control
system, no additional sensors are required. Using this method
not only the faulty cell, but also the faulty switching device can
be located. Moreover, inherited from the easy implementation
and robustness of the SMO [13]–[15], this method has good
immunity to conditions such as transient operation, degradation
of capacitance over time, and measurement inaccuracies.
A half-bridge (see Fig. 1) is the basic cell of an MMC. In
order to diagnose an open-circuit fault, it is essential to identify
the characteristics of a half-bridge as observed from its dc-side
and ac-side both in normal and fault conditions.

0885-8993/$31.00 © 2013 IEEE

The SMO offers desirable features such as robustness to parameter uncertainty and insensitivity to measurement noise [13]–[15]. This is verified in the all of the simulation results where generous values of 5 V and 1 μs are included for the device voltage drop and dead-time delay respectively. Therefore. C is the dccapacitance. but the switching states S have to be modified. (a) Eight cell MMC circuit (each cell represents a half-bridge). The equations which characterize the upper arm can be expressed as   1 E diP = − S1 vc1 + S2 vc2 + S3 vc3 +S4 vc4 + vo − dt l 2 dvci 1 = Si iP (i = 1. 1(b). Thus. 28. idc = iac . NOVEMBER 2013 TABLE II SWITCHING STATE S IN FAULT CONDITION Fig. NO. . SLIDING MODE OBSERVER An observer is a contrivance designed from a real system. then the . for example the saturation function of an observed-measured error. For all other conditions. (a) Normal condition.4868 IEEE TRANSACTIONS ON POWER ELECTRONICS. as shown in Fig. 1. we consider vc1 in this case) are selected to be observed. 1 are the gate signals for the switches. Vac = 0. Consider the half-bridge with an open circuit fault at T1 . 2(a). as (3) and (4) present) to force the observed output to converge to the actual output [13]. 1) Normal (Fault-Free) Condition: As shown in Fig. [16]. the switching state can be modified in a similar way. 2. T1 is ON and T2 is OFF. iac < 0. III. iac is forced to go through D2 instead of T1 as the result of the open-circuit fault. . When the gate signal is 1. Based on [16] and [18]. Switching model of half-bridge. . The analysis assumes ideal devices and instantaneous commutation. [17]. . 1(a). the SMO can be implemented in the field-programmable gate array (FPGA) [16]. vc4 are the capacitor voltages of cell 1 to cell 4. when g1 = 1. [15]. (b) Fault condition (an open-circuit fault at T 1 ). . g2 = 0. the SMO can be developed for a single-phase eight cell MMC as shown in Fig. g2 = 1. . uses high-gain feedback in the observer vector (normally in the form of a high-frequency switching function. 11. 2(a). 2. the switching models can still be described as (1). . however. respectively. alternatively. when g1 = 0. vc4 . the switching state S should be changed from 1 to 0. When the open-circuit fault occurs on T2 . . as shown in Fig. idc = 0. when it is 0. S1 . TABLE I SWITCHING STATE S IN NORMAL CONDITION g1 and g2 in Fig. the corresponding switch turns OFF. VOL. S4 are the corresponding switching states of the half-bridges given by Table I. so as to estimate its internal state [13]. If iP and one of the capacitor voltages (vc1 . (b) SMO simulated results in normal conditions. [17]. . 2) Fault Condition: In the fault condition (one open-circuit fault of the switch). With simple realization. the relationship between the ac-side and dc-side voltages and currents can be calculated as  Vac = S · Vdc (1) idc = S · iac where S is the switching state given by Table I. robust against nonideal device characteristics. . Table II demonstrates the modifications of switching states of a faulty half-bridge. 4) dt C (2) where l is the arm inductance. thus Vac = Vdc . 3. The fault detection method is. . The SMO Fig. and are complementary. . the corresponding switch turns ON. When g1 = 1. the half-bridge operates just as normal. vc1 . generally in the same mathematical form as the original system.

VOL. and observe iP and vci . B. steady conditions. 2). Note that the decision of fault occurrence is made only if the observed-simulated error (|ˆiP − iP |) lasts for a period of time (700 time-steps in this letter. Monitor Mode The aim of this mode is to determine whether the MMC is operating normally. sat(x) is the saturation function. then a fault has occurred. 2(a). 3. NO. set ˆiP (t0 ) = iP (t0 ). L1 sat(ˆiP − iP ) and L2 sat(−lS1 L1 sat(ˆiP − iP )) comprise the observer vector. and it lasts for 700 time steps (2 μs per step). Flowchart of the fault detection system. then iP and vc4 are observed. where Vthreshold is the threshold value of voltage . D1 (Assumption and Modification): Set the assumed faulty switch as Cell i(i = 1. 2. as the measurement noise may also lead to |ˆiP − iP | ≥ IThreshold . 3. modify Si (replace S with SF ) of SMO in (3) based on the Table II. NOVEMBER 2013 4869 SMO equations are  1 dˆiP = − S1 vˆc1 + S2 vc2 + S3 vc3 + S4 vc4 dt l  E + vo − − L1 sat(ˆiP − iP ) 2 1 dˆ vc1 = S1ˆiP − L2 sat(−lS1 L1 sat(ˆiP − iP )) dt C (3) where ˆiP and vˆc1 denote the observed states associated with the actual states iP and vc1 . and judgement to locate the open-circuit fault begins. if the assumed faulty switch is Cell 4. Tj (j = 1. where Ithreshold is the threshold value of current error. The process of each mode is detailed as follows. If |ˆiP − iP | ≥ IThreshold . L1 and L2 are the observer gains (large constant. We denote this moment as t0 and go to the detection mode. and a procedure including assumption. Consider the upper arm of the MMC circuit in Fig. and suppose that it initially operates in normal. The observed and actual simulated states are shown in Fig. (4) sat(x) = x ⎪ ⎩ −1 x ≤ −1 According to the analysis of [16]. The MMC circuit is simulated using Simulink/PLECS. The lower arm current and capacitor voltage can be observed in the same way. [18]. 11. modification. empirical value).IEEE TRANSACTIONS ON POWER ELECTRONICS. The fault detection systems consists of two modes. FAULT DETECTION ON AN MMC USING SLIDING MODE OBSERVER The basic idea of this detection method is to compare the observed state and the actual simulated state of the MMC. Fig. the system described in (3) is observable when S1 = 1. and the observer equations are  dˆiP 1 = − S4(F ) vˆc4 + S1 vc1 + S2 vc2 + S3 vc3 dt l  E + vo − − L1 sat(ˆiP − iP ) 2 1 dˆ vc4 = S4(F )ˆiP − L2 sat(−lS4(F ) L1 sat(ˆiP − iP )). Detection Mode The aim of this mode is to locate the open-circuit after its occurrence. else. T1 . then an open-circuit fault has occurred. for example 20000) to guarantee the sliding mode. and if they are different for a predefined period of time. 2(b). from which one can see that ˆiP and vˆc1 accurately match iP and vc1 . IV. respectively. whose derivations are detailed in [16]. (5) dt C vci − vci | < D2 (Judgement): If |ˆiP − iP | < IThreshold and |ˆ VThreshold . A. repeat the monitor mode. this is useful to prevent “false positives”. iP and one of the any capacitor voltages (for example vc1 ) are observed. 3. vˆci (t0 ) = vci (t0 ). as specified in Fig. 28. For example. which is defined as ⎧ x≥1 ⎪ ⎨1 −1 < x < 1 . 4).

Fault detection process: assume the faulty switch is C ell 4.1 s. One can see that ˆiP and vˆc1 cannot track iP and vc1 after t = 0. measurement inaccuracy. and it lasts for 100 ms (empirical value). IThreshold is the peak value of the arm current. go to D1 to try another switch. 5. as shown in Fig. In Fig. when the assumed faulty switch is the actual one–Cell 4. T 1 . Fig. 7. Fault detection at the 5% load condition. NOVEMBER 2013 Fig. 7(b). In the monitor mode (0 to 0. 4). VOL. Fault detection process: assume the faulty switch is cell 1. T 1 . 7(a). and it can be judged that the faulty switch is not vc1 − vc1 | > VThreshold within 100 ms. Fig. After 0. and imbalanced voltage. The Detection mode can be executed in parallel for each cell. Second. Fig. At the 5% load. 5 and Fig. SIMULATED RESULTS Consider an open-circuit fault occurring at Cell 4. 2. The observed states ˆiP and vˆc1 cannot track the simulated states iP and vc1 after the modification. T1 . A simulation has been undertaken . First. 5 and 6. 4(a). a significant difference between observed and simulated current (|ˆiP − iP | > IThreshold ) appears soon after the algorithm enters into the detection mode. 4. and the algorithm enters into the detection mode. 3. Tj (j = 1. Cell 1. the light load condition is considered. 2). Note that the threshold value (IThreshold and VThreshold ) are chosen in such a way that the observed-simulated error vci − vci |) are significant enough to allow a de(|ˆiP − iP | and |ˆ cision to be made with low risk of “false positive”. The open-circuit fault can still be located at the light load condition. 6 illustrate the process of the fault detection.4870 IEEE TRANSACTIONS ON POWER ELECTRONICS. the assumed faulty switch is the actual one–Cell 4. 6. In this letter. Fig. 6. 11. the occurrence of the open-circuit fault is detected with the condition |ˆiP − iP | ≥ IThreshold . T 1 . Fig. and VThreshold is 1/10 of the dc capacitor voltage. V. T1 at t = 0. In Fig. ROBUSTNESS The proposed SMO-based fault detection method has immunity to the parameter uncertainty. when the assumed faulty switch is not the actual one. as illustrated in Fig. T1 . else alter i or j. as shown in both Figs. (a) Open-circuit fault occurs at C ell 4. such that the open-circuit fault can be located within a very short time. as shown in Fig. then the assumption is correct—the faulty switch is Cell i(i = 1. the assumed faulty switch is Cell 1. T1 . 5. T1 as |ˆ Finally. It can be seen that the observed states ˆiP and vˆc4 can track iP and vc4 again. the observed states can follow the actual states.1 s. the observed current ˆiP can follow the simulated current iP . error. NO.1 s. the observed states cannot follow the actual simulated states after the occurrence of the open-circuit fault. 4(b) shows that if no fault detection algorithm is applied. 28. VI.1 s). (b) Observed and actual simulated states without using the fault detection algorithm.

“Design and control of a modular multilevel HVdc converter with redundant power modules for noninterruptible energy transfer. VII. Telford. 8. Park.-H. 2009. Rothenhagen and F.2C. Rodriguez. 7. 5. Peuget. N. vol. “Fault detection and isolation on a pwm inverter by knowledge-based model. O. [6] O. Electron. Conf.” IEEE Trans. 32nd Annu. Ind. Yu. pp.. Electron. which include device voltage drops and commutation delays. Lima. 2005. as Fig. 8. pp. Robustness test: assume the faulty switch is cell 1. 32nd IEEE IAS Annu. 3. “A novel fault detection scheme for voltage fed PWM inverter. Anaya-Lara. have shown that it is effective in locating the faulty switch and very robust against both parameter uncertainty and measurement error. NO. 28.IEEE TRANSACTIONS ON POWER ELECTRONICS. 587–593. pp. 1) Parameter uncertainty: Let the arm inductance used in the calculation be 110% of the actual value. [2] M. pp. NOVEMBER 2013 4871 cannot match iP and vc1 . Construction of experiments is currently underway. Power Electron. 27. With parallel execution. Dec. C. “Fault detection of open-switch damage in voltage-fed pwm motor drive systems. [10] R. Khomfoi and L. In regard to locating the fault (after 0. IET.” in Proc. 9. [3] G.. T. Park.. Song and A.or full-bridge converters. Jun. B. pp. IEEE Conf. pp. while the control system and the fault detection algorithm will be carried out in the TMS320C6713 DSP. REFERENCES Fig. Son. and J. Lezana. Spec. Electron. Ind. and A. 6. and J. to demonstrate these properties. and J.. Flett. Power Del. 2012. Hur. Jun. 3. Fuchs. lcal = 1. D. no. “An innovative modular multilevel converter topology suitable for a wide power range. no. Adam. pp. the observed states ˆiP and vˆc4 can track iP and vc4 . S. 2700–2708. Sep. 9 is 1. Huang and F.. 2654–2659. Record. IEEE Bologna Power Tech Conf. Williams. 2010. 2.25 times the value of vc4 in Fig.” in Proc. McDonald.-S. Ind. vol. Lesnicar and R.. Jun. The accuracy of the measurement may limit the application of this method–it is not able to locate the fault when the systematic measurement error is larger than 7%. “Modular multilevel inverter: Pulse width modulation and capacitor balancing technique.” IEEE Trans. pp. K. Nov. vol. 6. vol.. In the monitor mode (0 to 0. R. Appl. S. S. Aguilera. 10. vol. vol..-H. 2003. This method is also independent of operation frequency of an MMC. the open-circuit fault can be located within a very short time. T 1 . Marquardt. IEEE Power Electron. 2006. This method can also be used on any other multilevel half. Rognon. In the presence of parameter uncertainty and measurement noise and imbalanced capacitor voltage. E. VOL. and part of the simulation parameters have been amended as follows to imitate the parameters and measurement errors and the imbalanced capacitor voltage. whereas the assumed switch is not the actual one.25Vc4 before the occurrence of the open-circuit fault. 54. Power Electron. 1737–1746. Electron. when the assumed switch is the actual faulty one. 1471–1478. and J.” Power Electron.-W. Hyun. and the dc-capacitance 120% of the actual value.” in Proc. Aug.” IEEE Trans. Jul. Baek. 1997. Burt. Ind. the SMO is going to be implemented in the ACTEL A3P400 FPGA. 2003. no. 1611–1619. Chung. Akagi. “Fault detection on multicell converter based on output voltage frequency analysis.. the faulty switch can still be located. p. Lee. “IGBT fault protection based on di/dt feedback control. Eur. ˆiP and vˆc1 [1] A. The results are illustrated in Figs. 57. Ccal = 1.” IEEE Trans.. [12] S. no. Courtine. H. 1478– 1484. This letter proposes an SMO-based fault detection method for an MMC. Simulations. Y. “Fault diagnosis and reconfiguration for multilevel inverter drive using AI-based techniques. Ind. 2.. 8 presents. vol. 702–715. T. Huang. 2010. 9 illustrates.1s). . Oct. T 1 . Power Electron. One can also notice that the capacitor voltage is imbalanced during this period: vc1 in Fig. Hagiwara and H. Robustness test: assume the faulty switch is C ell 4. vol. [7] W. 24. 2007. [11] P. Jul. this is similar to to the approaches taken in [17] and [18]. vci cal = (1 + ΔS + ΔR ) · vci .” IEEE Trans. [4] G. [9] K.-T. G. 2009. Mar. 18.” IEEE Trans. 56. vol.-J. and D. 8 and 9. 2007. the observed current can still track the simulated current. Nam. no. Jacobina.-J. Conf. 8. CONCLUSION Fig. p. pp. 3) Imbalanced capacitor voltage: Let Vc1 = 1. 3. de Araujo Ribeiro. 2275–2283. 6. U. “Performance of diagnosis methods for IGBT open circuit faults in three phase voltage source inverters for ac variable speed drives. “Control and experiment of pulsewidthmodulated modular multilevel converters. Meet. as Fig.1s). [8] R. “Fault-tolerant design and control strategy for cascaded h-bridge multilevel converter-based statcom. no. namely. 11. 2) Measurement inaccuracy: Add 2% of the systematic error (ΔS ) and 10% of the random error (ΔR ) to the measured capacitor voltage.” in Proc. Appl. 2954–2968. pp. Other conditions are the same as the ones in Section V. Lee. [5] F.-S.” in Proc. no. Conf. namely.1l. da Silva. Tolbert.

Clare.. Amer. and T. dissertation. Wheeler. Dept. and J. [16] M. 2009. Eng. FL.” in Proc. Slotine. [14] J. 11. “On sliding mode observers via equivalent control approach. Electron. 2007. U. P. NO. no. pp. “Digital sliding-mode observer implementation using FPGA. [15] U. 1051–1067. Lienhardt. Jun. 54. Apr. Vadim and I. Power Electron. Guldner. Gateau. E. Conf.. O. . [18] M. J. [17] A. no.D. Drives. 4. “On sliding observers for nonlinear systems. Sliding Mode Control in ElectroMechanical System (Automation and Control Engineering Series). pp.4872 IEEE TRANSACTIONS ON POWER ELECTRONICS. vol. K. Haskara. 28. Boca Raton. 2010. 2010. J. vol.. and E. Electron.-J.” IEEE Trans. Shi. 1–6. A. Aug. Mach. 5th IET Int. Nottingham.” Ph. 1865–1875. the Unversity of Nottingham. pp. Ind. Elec. “Sliding mode observation of capacitor voltage in multilevel power converters. 71. 6. Almaleki. VOL. pp. G. J. 1986.” in Proc. 1998.. Almaleki. 1794– 1800. Misawa.” Int. and J. Meynard. Control. Control Conf. “Sliding mode observation of capacitor voltage in multilevel power converters...K. Utkin. NOVEMBER 2013 [13] V. USA: CRC Press.-M. Hedrick.