Anomaly detection for IGBTs using Mahalanobis distance

Microelectronics Reliability 55 (2015) 1054–1059

Contents lists available at ScienceDirect

Microelectronics Reliability journal homepage: www.elsevier.com/locate/microrel

Anomaly detection for IGBTs using Mahalanobis distance Nishad Patil, Diganta Das, Michael Pecht ⇑ Center for Advanced Life Cycle Engineering (CALCE), University of Maryland, College Park, MD 20742, United States

a r t i c l e

i n f o

Article history: Received 3 March 2014 Received in revised form 9 February 2015 Accepted 6 April 2015 Available online 29 April 2015 Keywords: Anomaly detection Fault detection Mahalanobis distance IGBT

a b s t r a c t In this study, a Mahalanobis distance (MD)-based anomaly detection approach has been evaluated for non-punch through (NPT) and trench field stop (FS) insulated gate bipolar transistors (IGBTs). The IGBTs were subjected to electrical–thermal stress under a resistive load until their failure. Monitored on-state collector–emitter voltage and collector–emitter currents were used as input parameters to calculate MD. The MD values obtained from the healthy data were transformed using a Box–Cox transform, and three standard deviation limits were obtained from the transformed data. The upper three standard deviation limits of the transformed MD healthy data were used as a threshold for anomaly detection. The anomaly detection times obtained by using the MD approach were compared to the detection times obtained by using a fixed percentage change threshold for the on-state collector–emitter voltage. Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction Insulated gate bipolar transistors (IGBTs) are the power semiconductor switch of choice for operating voltages above 200 V [1]. Due to their low on-state voltage drop and simple gate drive requirements, IGBTs have been widely used in medium- and high-power motor drives and power supplies. IGBTs have been reported to fail under excessive electrical and thermal stresses in variable speed drives [2] and are considered as reliability problems in wind turbines [3], inverters in hybrid electric vehicles [4], and railway traction motors [5]. To address these reliability issues, there is an increased focus on the development of diagnostic and prognostic techniques for IGBTs. Anomaly detection approaches have been reported in literature for identifying faults in IGBT power inverters which involve monitoring system level currents and voltages, collector–emitter voltage, gate emitter voltage and collector emitter current for fault diagnosis. These methods have been developed to identify system level faults and the location of the faulty IGBTs and are not meant to detect faults in the IGBTs themselves [2,6]. A prognostic approach was reported for IGBTs in power modules in automotive applications [7]. This approach involved measurement of the IGBT collector–emitter saturation voltage at vehicle start-up and comparison of the measured value to a look-up table that contained healthy values of on-state collector–emitter voltage VCE(ON). A 15% change for VCE(ON) was used as a threshold for anomaly detection. ⇑ Corresponding author. E-mail address: [email protected] (M. Pecht). http://dx.doi.org/10.1016/j.microrel.2015.04.001 0026-2714/Ó 2015 Elsevier Ltd. All rights reserved.

In another study [8], a 7% change in VCE(ON) was proposed as a threshold for anomaly detection for power module failures as a result of wire-bond degradation. In this study, we present an alternate approach to fault detection in IGBTs which involves the use of Mahalanobis distance (MD) for anomaly detection. One of the advantages of the MD approach is that it can be derived from multiple parameters which could potentially provide earlier anomaly detection times in comparison to using a single parameter. This approach was implemented on data obtained by electrical–thermal stress experiments performed on non-punch through (NPT) and trench field stop (FS) IGBT devices. The VCE(ON), on-state collector–emitter current ICE(ON), and package temperature were monitored in-situ during the test. MD was calculated using VCE(ON) and ICE(ON) parameters, and a threshold was defined to detect anomalies. The detection times obtained by this approach were compared to the detection times obtained by the use of a fixed percentage change threshold for the on-state collector–emitter voltage. This paper is organized in five sections. Section 2 describes the procedure used to stress the IGBT devices to failure as well as the data obtained from the experiments, Section 3 describes the equations used to calculate MD and the methodology used for determining the thresholds for anomaly detection, Section 4 describes the results, and Section 5 summarizes this work.

2. Experimental procedure Ten NPT (IRGB15B60KD) IGBTs labeled N1–N10 and FS (IRGB4056) IGBTs labeled F1–F10 from International Rectifier were

1055

N. Patil et al. / Microelectronics Reliability 55 (2015) 1054–1059

evaluated in this study. A schematic of the NPT and FS IGBT is shown in Fig. 1. The IGBT devices were packaged in a TO-220A package along with a soft recovery diode. The devices were rated for a collector–emitter voltage of 600 V and gate–emitter voltage of 20 V. The maximum junction temperature rating was 150 °C for NPT and 175 °C for FS IGBTs. The detailed description of the experimental setup used for power cycling of IGBTs is provided in [10]. To perform the IGBT electrical–thermal stress tests, the IGBTs were switched on and off with a gate voltage of 15 V, duty cycle of 50%, switching frequency of 1 kHz, and collector–emitter voltage of 5 V. The switching was performed until the device reached a pre-defined maximum temperature as illustrated in Fig. 2. Once the maximum temperature Tmax was achieved, switching was stopped until the device cooled to Tmin after which switching was resumed again. The temperature increase was caused by self-heating of the IGBTs. No external heat source was used. The main contributor to the self-heating was due to conduction losses in comparison to the switching losses as the switching frequency was low (1 kHz). The temperature at the start of the tests was room temperature. Tmean in the experiments was set to 300 °C for NPT and 250 °C for FS IGBTs, and the Tmin and Tmax temperatures were set to a range of ±15 °C from the mean temperature. In this stress condition, failures observed were either due to latch-up (loss of gate control leading to increase in collector– emitter current) or failure to turn on. The failure mode and failure times for the devices tested are given in Tables 1 and 2 . In-situ measurement of the collector–emitter voltage, collector–emitter current, and package temperature was performed every 400 ms until failure of the IGBT under test and recorded using a National Instruments data acquisition system (NI-DAQ). Temperature monitoring was performed using an infrared sensor focused on the front surface of the package, and current measurements were performed by a Hall-effect current sensor. Fig. 3 illustrates the latch-up failure mode observed for an FS IGBT. The current peaks in Fig. 3 represent the current as the IGBT heats up from Tmin to Tmax. The current reduces to zero as the IGBT cools down to Tmin before switching begins again. At latch-up the current exceeded 18 A upon which the collector–emitter voltage supply was cut off to prevent device burn-out. Periodically during the test, a square gate pulse of magnitude 1.5 times greater than the stress voltage with a 1 ms duration and 50% duty cycle was applied to the gate. The collector–emitter current and collector–emitter voltage responses to this gate pulse were recorded using an oscilloscope as shown in Figs. 4 and 5.

Fig. 2. Schematic of temperature and switching stress profile.

Table 1 Failure time and failure mode for NPT IGBTs. Device ID

Failure time (min)

Failure mode

N1 N2 N3 N4 N5 N6 N7 N8 N9 N10

47.1 55.8 48.7 60.0 39.4 56.2 42.6 41.0 54.8 34.8

Latch-up Failure to turn on Latch-up Failure to turn on Latch-up Failure to turn on Latch-up Latch-up Latch-up Latch-up

Table 2 Failure time and failure mode for FS IGBTs. Device ID

Failure time (min)

Failure mode

F1 F2 F3 F4 F5 F6 F7 F8 F9 F10

36.8 51.0 39.5 24.6 56.2 58.2 61.0 42.7 56.7 52.6

Latch-up Latch-up Latch-up Latch-up Latch-up Latch-up Failure to turn on Latch-up Latch-up Latch-up

16 14

10 8

I

CE

(A)

12

6 4

Trench field stop Non Punch through

Fig. 1. Schematic of the non-punch through and trench field stop IGBTs [9].

2 0

37:37

37:54

38:12

38:29

38:46

39:03

39:21

39:38

39:55

Time (Minutes) Fig. 3. Latch-up of FS IGBT (F3) as recorded using the NI-DAQ.

To determine the effects of the electrical–thermal stress without the influence of temperature changes, the collector–emitter voltages and currents at the mean test temperature were extracted

N. Patil et al. / Microelectronics Reliability 55 (2015) 1054–1059

3.1. Fixed threshold approach Fixed thresholds based on percentage change in the collector– emitter voltage (VCE) have been reported in literature for early warning [7,8] and as a failure criteria for IGBTs [16–18]. However, there are no specific guidelines to make the choice of threshold. These thresholds are typically based on application conditions or expert judgment. In this study, a 5% increase in VCE(ON) was used as a fixed threshold. The anomaly detection times with the fixed threshold were taken to be the first occurrence of the VCE(ON) going above 5% of its initial value at the mean temperature.

15

ICE (A)

10

5

0

0

500

1000

1500

2000

2500

3000

3500

4000

4500

Pulse Duration (1 division=0.2µs) Fig. 4. Collector–emitter current response to input gate pulse.

5000

4 3 2 1 0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

Pulse Duration (1 division=0.2µs) Fig. 5. Collector–emitter voltage response to input gate pulse.

I CE (ON) V CE (ON)

2.6

13.0

2.5

12.5

2.4

12.0

2.3

11.5

2.2

11.0

2.1

10.5

VCE (ON) (V)

13.5

2.0 0

10

20

30

40

50

Time (Minutes) Fig. 6. ICE(ON) and VCE(ON) vs. time for NPT IGBT ‘‘N3’’ at mean aging temperature.

13.2

2.4

I CE (ON) V CE (ON)

13.0

2.3

12.8

2.3

12.6

2.2

12.4

2.2

12.2

2.1

12.0

2.1

VCE (ON) (V)

Approaches to implementing anomaly detection depend on the type of data available from the system under consideration. When healthy data from a system is available, anomaly detection can be implemented by determining a detection threshold based on the healthy data in order to identify outliers. Threshold detection therefore is an important step in diagnostics in order to have advanced warning of failure. Threshold values are typically defined based on expert knowledge of known fault conditions and economic factors such as the need to reduce the number of false alarms. These diagnostic approaches may not be able to detect anomalies when a priori knowledge of faults is not available. It is therefore useful to implement a generalized probabilistic approach to determine thresholds for anomaly detection. We evaluate a fixed threshold approach and a probabilistic MD approach in the following sections.

5

I CE (ON) (A)

3. Anomaly detection

6

ICE (ON) (A)

for this study. For every collector–emitter current and voltage waveform obtained at the mean test temperature, a portion of the waveform was sampled in the transistor on-state. A total number of 500 points were sampled as depicted by the window shown in Figs. 4 and 5. The mean of these 500 points was then plotted against test time, an example for NPT IGBT is shown in Fig. 6 and FS IGBT is shown in Fig. 7. The on-state collector–emitter current ICE(ON) was observed to reduce with time, and the on-state collector–emitter voltage VCE(ON) increased. Finally, failure occurred by two modes. The first failure mode was loss of gate control due to latch-up as a result of the activation of the parasitic thyristor inherent in the device structure. The second failure mode was an increase in device resistance to an extent wherein the device failed to turn on. X-ray analysis was performed before and after the stress tests, and die attach degradation was observed in all the devices. Threshold voltage shifts were also observed indicating damage to the gate oxide [11]. The ICE(ON) and VCE(ON) parameters extracted at the mean test temperature were used for anomaly detection by the MD approach which is described in the next section.

VCE (V)

1056

2.0

11.8 0

10

20

30

40

Time (Minutes) Fig. 7. ICE(ON) and VCE(ON) vs. time for FS IGBT ‘‘F3’’ at mean aging temperature.

3.2. MD-based approach The MD approach implemented in this study uses a probabilistic method to define the threshold for anomaly detection based on the healthy behavior of the device. The MD approach uses multiple parameters for assessing the health of the system. MD is a distance measure that has been used in applications such as anomaly detection, pattern recognition, and process control [12]. In electronics, MD has been used for detecting anomalies in notebook computers [13], multilayer ceramic capacitors, and embedded planar capacitors [14,15]. The MD approach reduces multivariate data to univariate data. It is sensitive to changes between various parameters monitored as it takes the correlation between the different parameters into account. Additionally, it is not sensitive to the differing scales of the parameters monitored, as MD values are calculated using normalized parameters [13].

1057

N. Patil et al. / Microelectronics Reliability 55 (2015) 1054–1059

Fig. 8 shows data for two parameters X1 and X2. The MD is the squared distance between any observed vector ‘x’ and the mean vector ‘l’ of the healthy data. In Fig. 8, the test data is within the healthy space if one considers a single parameter X1. However, if one considers the healthy data in terms of the parameters X1 and X2, it is seen that the test data is not part of the healthy data and is therefore labeled as an anomaly. One of the advantages of the MD approach for anomaly detection of IGBTs is that a health assessment of the device can be performed by including both the current and voltage parameters. As the MD approach determines anomalies by comparing test data to the healthy data obtained from the same device, it provides a better estimate of the health of the device under consideration when compared to the use of a fixed threshold. To implement the MD approach, ICE(ON) and VCE(ON) data at the mean test temperature as illustrated in Figs. 6 and 7 were partitioned into healthy data and test data. The first 50 observations were classified as healthy data which corresponded to approximately the first 5–8 min of the electrical–thermal stress test. The subsequent observations were used as test data. The parameters that form the input for MD computation are denoted by i, where i = 1, 2, . . ., p. In our study ICE(ON) and VCE(ON) were used as input parameters, hence p = 2. The number of observations recorded for each parameter is denoted by j, where j = 1,2, . . .n. Xij denotes the value of parameter i at time instance j. Each individual observation of a given parameter in the data vector was normalized as given by Eq. (1). The sample mean (X i ) and sample standard deviation (Si) for the input parameters was computed by using Eq. (2).

Z ij ¼

Xi ¼

iÞ ðX ij  X Si n 1X

n

j¼1

X ij ;

ð1Þ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi uPn   u  2 t j¼1 X ij  X i Si ¼ ðn  1Þ

obtained were found to not follow a normal distribution. The Box– Cox power transformation was used to transform the healthy MD values into a normal distribution. This approach has been demonstrated to be effective in threshold estimation for anomaly detection of electronic products [13]. The healthy MD values before and after the Box–Cox transformation are shown in Fig. 9. The mean (l) and standard deviation (r) of the transformed healthy MD values were used to obtain 3r bounds about the mean. The upper bound (l + 3r) was used as a threshold for anomaly detection as increasing MD values indicate degradation in the IGBT. When the transformed test MD values crossed the threshold, an anomaly was said to have occurred.

4. Results and discussion In this section, the anomaly detection times obtained by using the MD approach are compared to the detection times obtained by using a fixed percentage change threshold for VCE(ON). The transformed MD test and VCE(ON) data along with their respective thresholds for an NPT IGBT ‘‘N3’’ is shown in Fig. 10. The part failed due to latch-up at 49 min, and anomalies in the device were detected at 12 min using the MD approach and 17 min using the 5% increase in VCE(ON) threshold. The transformed MD test and VCE(ON) data for an FS IGBT ‘‘F3’’ is shown in Fig. 11. The part failed due to latchup at 40 min, and the anomaly was detected at 12 min using the MD approach and at 36 min using the 5% increase in VCE(ON) threshold. A 3r threshold above the mean was used for the transformed MD for anomaly detection, but this can be set to any value in principle. Another approach to anomaly detection besides using MD is to use a l + 3r threshold derived from the healthy data for a single parameter, either the ICE(ON) or VCE(ON). The l + 3r threshold based

ð2Þ 20

The healthy MD values were computed by using Eq. (3) with the normalized parameters obtained from Eq. (1), where Zj was the normalized ICE(ON) and VCE(ON) at time j

1 T 1 Z C Zj p j

ð3Þ

where C is the correlation matrix. Eq. (4) was used to calculate the correlation matrix. n 1 X C¼ Z j Z Tj ðn  1Þ j¼1

Frequency

MDj ¼

15

10

5

ð4Þ 0 0

To compute the test MD values, the mean and standard deviation obtained from the healthy MD computation were used to normalize the test ICE(ON) and VCE(ON). These normalized test parameters were then used to compute the test MD by Eq. (3) where the correlation matrix was obtained from the healthy data. The healthy MD values

μ

x

Healthy data

Fig. 8. Anomaly detection using MD.

3

4

5

8

6

4

2

0 -3

x2

2

MD

Frequency

Test data

x1

1

-2

-1

0

1

2

Transformed MD Fig. 9. Histograms of the healthy MD for one NPT device before and after the Box– Cox transformation.

1058

N. Patil et al. / Microelectronics Reliability 55 (2015) 1054–1059

15

2.7

11

2.6

9

2.5

Transformed MD crosses threshold

7 5

2.4 Transformed MD threshold

3

2.3

VCE(ON) crosses threshold

1 -1

VCE(ON) (V)

Transformed MD

2.8

Transformed MD VCE(ON)

13

2.2 VCE(ON) threshold

2.1

-3 -5

2.0 0

10

20

30

40

50

Fig. 12. Covariance ellipsoid of the healthy space of FS IGBT ‘‘F3’’ and detected anomaly using MD approach.

Time (Minutes) Fig. 10. Transformed MD test data and VCE(ON) for NPT IGBT ‘‘N3’’.

on the healthy data is an improvement over using a fixed threshold on a single parameter. However, there are situations in which measurements from a system may be within the thresholds of the individual parameters but are not part of the healthy space. In such situations, the MD approach can detect the anomaly while the single parameter threshold (even one based on the healthy data) would not be able to do so. In Fig. 12, the covariance ellipsoid for the healthy data of FS IGBT ‘‘F3’’ is shown along with the anomaly. From this figure, one can observe that although the test data point had a lower VCE(ON) than the previous measurements, it was labeled anomalous as it did not fall into the healthy space. In Tables 3 and 4, the anomaly detection times obtained for the NPT and FS IGBTs evaluated are summarized for the two approaches. A percent time to failure after detection metric is computed for each device. This metric refers to the ratio of the difference between the detection time and the failure time. Higher values of this metric indicate earlier anomaly detection times. For both the NPT and FS IGBTs, the MD-based approach was observed to provide earlier anomaly detection times in comparison to the 5% change in VCE(ON) threshold for seven of the ten devices, two devices had the same detection time while for one device the 5% change in VCE(ON) threshold yielded an earlier detection time. The differences in the anomaly detection times between the two approaches were significantly higher for the FS IGBTs. The VCE(ON) increase is much slower with degradation in the FS IGBT due to the field stop layer. In two of the FS IGBTs tested, the VCE(ON) did not increase beyond 5% resulting in no anomalies being detected by the use of the VCE(ON) threshold.

Transformed MD VCE(ON)

40

Transformed MD

35

VCE(ON) crosses threshold

2.2

VCE(ON) threshold

30 25

2.1

20 15

2

10

Transformed MD crosses threshold

5

1.9

Transformed MD threshold

0 -5

0

10

Device ID

Detection time using transformed MD (min)

Time for 5% increase in VCE(ON) (min)

Failure time (min)

% Time to failure after detection using transformed MD

% Time to failure after 5% increase in VCE(ON)

N1 N2 N3 N4 N5 N6 N7 N8 N9 N10

13.8 12.8 11.8 18.7 11.4 12.0 16.2 17.2 14.6 15.0

17.8 16.7 16.5 18.7 17.1 16.7 17.4 17.2 8.0 18.0

47.1 55.8 48.7 60.0 39.4 56.2 42.6 41.0 54.8 34.8

70.8 77.1 75.8 68.9 71.0 78.7 62.0 58.0 73.3 57.0

62.3 70.1 66.2 68.9 56.5 70.3 59.2 58.0 85.3 48.4

Table 4 Detection time obtained for FS IGBTs (minutes). Device ID

Detection time using transformed MD (min)

Time for 5% increase in VCE(ON) (min)

Failure time (min)

% Time to failure after detection using transformed MD

% Time to failure after 5% increase in VCE(ON)

F1 F2 F3 F4 F5 F6 F7 F8 F9 F10

14.3 15.8 12.4 9.1 12.0 14.7 56.3 16.8 9.0 51.0

– 50.7 35.7 – 44.9 51.2 56.3 16.8 10.7 50.5

36.8 51.0 39.5 24.6 56.2 58.2 61.0 42.7 56.7 52.6

61.1 69.1 68.7 62.8 78.6 74.7 7.7 60.6 84.1 3.0

– 0.7 9.6 – 20.2 12.1 7.7 60.6 81.1 4.0

2.3

VCE (ON) (V)

45

Table 3 Detection time obtained for NPT IGBTs (minutes).

20

30

40

1.8

Time (Minutes) Fig. 11. Transformed MD test data and VCE(ON) for FS IGBT ‘‘F3’’.

In Table 5, the percentage change in VCE(ON) required to obtain a comparable detection time as the MD approach is shown for NPT and FS IGBTs. As observed from Table 5, a reduction in the fixed percentage threshold to 1% or 2% from 5% for VCE(ON) would allow for equivalent or earlier detection times between the MD approach and the fixed threshold approach for the NPT IGBTs. However, for the FS IGBTs even a 1% threshold for VCE(ON) does not provide equivalent or earlier detection times compared to the MD approach for four of the devices. This result highlights the limitations of using a single parameter for anomaly detection. A single parameter does not provide the complete picture of the health of a system. The MD approach by taking into account more than one parameter enables the comparison of test data to the ‘‘healthy

N. Patil et al. / Microelectronics Reliability 55 (2015) 1054–1059 Table 5 Equivalent percentage change in VCE(ON) at time of anomaly detection by MD approach. Device ID

N1 N2 N3 N4 N5 N6 N7 N8 N9 N10

Detection time using transformed MD (min)

% Increase in VCE(ON) at detection

Device ID

13.8 12.8 11.8 18.7 11.4 12.0 16.2 17.2 14.6 15.0

2.0 1.7 1.8 5.0 1.6 1.4 4.8 5.0 9.6 2.8

F1 F2 F3 F4 F5 F6 F7 F8 F9 F10

Detection time using transformed MD (min)

% increase in VCE(ON) at MD detection

14.3 15.8 12.4 9.1 12.0 14.7 56.3 16.8 9.0 51.0

0.1 0.5 0.2 1.1 0.5 1.2 5.0 5.0 3.5 6.1

space’’. This allows the MD approach to detect anomalies that a single parameter fixed threshold cannot as illustrated in Fig. 12. 5. Summary and conclusions NPT and FS IGBTs were subjected to electrical–thermal stresses under a resistive load until failure. The ICE(ON) and VCE(ON) parameters extracted at the mean test temperature were used to compute the Mahalanobis distance (MD). The l + 3r upper bound obtained from healthy MD values was then used as a threshold for anomaly detection. This approach was compared to anomaly detection times obtained by using a single parameter threshold based on a 5% change in VCE(ON). The MD-based anomaly detection approach was demonstrated to be able to detect anomalies before failure for all the devices evaluated. The advantage of using the MD-based approach was especially evident in the analysis of detection times for the FS devices, where even a threshold of 1% change in VCE(ON) did not provide equivalent or earlier detection times compared to the MD approach for four of the devices. The anomaly detection times observed in this study indicate that the MD approach that incorporates the degradation in ICE(ON) in addition to VCE(ON) is a better approach for anomaly detection in NPT and FS IGBTs compared to using a fixed percentage change in VCE(ON). The performance of the MD-based approach to detect anomalies in IGBTs was demonstrated using accelerated stress test data. Implementing this approach for real time diagnostics will require additional experiments performed at different temperatures to obtain an acceleration factor which will aid in determining anomaly detection times in actual operating conditions. This work illustrates how using multiple parameters rather than a single parameter, and a probabilistic method to define the threshold based on the healthy behavior provides improved anomaly detection results in IGBTs. Multiple parameter-based detection metrics have the potential for integration into higher-level systems

1059

where interaction measures between parameters relate to degradation. Acknowledgments The authors would like to thank the Center for Advanced Life Cycle Engineering (CALCE) at the University of Maryland and the more than 100 companies and organizations that support its research annually. The authors would also like to thank the members of the Prognostics and Health Management Consortium at CALCE for their support of this work. The IGBT experiments were performed at the Prognostics Center of Excellence, NASA Ames research center. The authors thank Dr. Jose Celaya and Dr. Kai Goebel of NASA Ames for their guidance. References [1] Baliga B. Power semiconductor devices. New York: Springer; 2008. [2] Lu B, Sharma S. A literature review of IGBT fault diagnostic and protection methods for power inverters. IEEE Trans Ind Appl 2009;45(5):1770–7. [3] Arifujjaman M, Iqbal M, Quaicoe J. Reliability analysis of grid connected small wind turbine power electronics. Appl Energy 2009;86:1617–23. [4] Hirschmann D, Tissen D, Schroder S, De Doncker RW. Reliability prediction for inverters in hybrid electrical vehicles. IEEE Trans Power Electron 2007;22:2511–7. [5] Perpina X, Serviere J, Jorda X, Fauquet A, Hidalgo S, Urresti-Ibanez J, et al. IGBT module failure analysis in railway applications. Microelectron Reliab 2008;48:1427–31. [6] Rothenhagen K, Fuchs F. Performance of diagnosis methods for IGBT open circuit faults in voltage source active rectifiers. In: Proceedings of the IEEE power electronics specialists conference; 2004. p. 4348–54. [7] Xiong Y, Cheng X, Shen Z, Mi C, Wu H, Garg V. Prognostic and warning system for power-electronic modules in electric, hybrid electric, and fuel-cell vehicles. IEEE Trans Industr Electron 2008;55:2268–76. [8] Bing Ji, Pickert V, Wenping Cao, Zahawi B. In situ diagnostics and prognostics of wire bonding faults in IGBT modules for electric vehicle drives. IEEE Trans Power Electron 2013;28(12):5568–77. [9] Nakano H, Onozawa Y, Ikawa O. New IGBT-PIM with 6th generation chip and package technologies. Fuji Electr Rev 2008;54(2):52–6. [10] Patil N, Das D, Pecht M. A prognostic approach for non-punch through and field stop IGBTs. Microelectron Reliab 2012;52:482–8. [11] Patil N, Celaya J, Das D, Goebel K, Pecht M. Precursor parameter identification for IGBT prognostics. IEEE Trans Reliab 2009;58(2):271–6. [12] De Maesschalck R, Jouan-Rimbaud D, Massart D. The Mahalanobis distance. Chemom Intell Lab Syst 2000;50(1):1–18. [13] Kumar S, Chow TSW, Pecht M. Approach to fault identification for electronic products using Mahalanobis distance. IEEE Trans Instrum Meas 2010;59(8):2055–64. [14] Nie L, Azarian M, Keimasi M, Pecht M. Prognostics of ceramic capacitor temperature-humidity-bias reliability using Mahalanobis distance analysis. Circuit World 2007;33(3):21–8. [15] Alam M, Azarian M, Osterman M, Pecht M. Prognostics of failures in embedded planar capacitors using model-based and data-driven approaches. J Intell Mater Syst Struct 2011:1293–304. [16] Held M, Jacob P, Nicoletti G, Scacco P, Poech MH. Fast power cycling test for insulated gate bipolar transistor modules in traction application. Int J Electron 1999;86:1193–204. [17] Coquery G, Lefranc G, Licht T, Lallemand R, Seliger N, Berg H. High temperature reliability on automotive power modules verified by power cycling tests up to 150 °C. Microelectron Reliab 2003;43:1871–6. [18] Forest F, Rashed A, Huselstein J, Martiré T, Enrici P. Fast power cycling protocols implemented in an automated test bench dedicated to IGBT module ageing. Microelectron Reliab 2015;55:81–92.