Prognostics Approach for Power MOSFET under Thermal-Stress Aging José R. Celaya, PhD, SGT Inc., Prognostics Center of Excellence, NASA Ames Research Center Abhinav Saxena, PhD, SGT Inc., Prognostics Center of Excellence, NASA Ames Research Center Chetan S. Kulkarni, ISIS, Vanderbilt University Sankalita Saha, PhD, MCT., Prognostics Center of Excellence, NASA Ames Research Center Kai Goebel, PhD, Prognostics Center of Excellence, NASA Ames Research Center Key Words: Power MOSFET, Prognostics, PHM, Accelerated Life Test SUMMARY & CONCLUSIONS 1 INTRODUCTION The prognostic technique for a power MOSFET presented Prognostics is an engineering discipline focused on in this paper is based on accelerated aging of MOSFET predicting the time at which an in-service component will fail. IRF520Npbf in a TO-220 package. The methodology utilizes The science of prognostics is based on the analysis of failure thermal and power cycling to accelerate the life of the devices. modes, detection of early signs of wear and aging, and fault The major failure mechanism for the stress conditions is die- conditions. These signs are then correlated with a damage attachment degradation, typical for discrete devices with lead- propagation model and suitable prediction algorithms to arrive free solder die attachment. It has been determined that die- at a “remaining useful life” (RUL) estimate. The discipline attach degradation results in an increase in ON-state resistance that links studies of failure mechanisms to system lifecycle due to its dependence on junction temperature. Increasing management is often referred to as prognostics and health resistance, thus, can be used as a precursor of failure for the management (PHM). Power semiconductor devices such as die-attach failure mechanism under thermal stress. A feature MOSFETs (Metal Oxide Field Effect Transistors) are essential based on normalized ON-resistance is computed from in-situ components of electronic and electrical subsystems in on- measurements of the electro-thermal response. An Extended board autonomous functions for vehicle controls, Kalman filter is used as a model-based prognostics techniques communications, navigation, and radar systems. In current based on the Bayesian tracking framework. practices, maintenance schedules are usually based on The proposed prognostics technique reports on reliability data available from the manufacturer. However, preliminary work that serves as a case study on the prediction while this approach works well in aggregate on a large number of remaining life of power MOSFETs and builds upon the of components, failures on individual components are not work presented in [1]. The algorithm considered in this study necessarily averted. For mission critical systems it is had been used as prognostics algorithm in different extremely important to avoid such failures. This calls for applications and is regarded as suitable candidate for condition based prognostic health management methods. component level prognostics. This work attempts to further the validation of such algorithm by presenting it with real 1.1 Related Work degradation data including measurements from real sensors, In [2] a model-based prognostics approach for discrete which include all the complications (noise, bias, etc.) that are IGBTs was presented. RUL predictions were accomplished regularly not captured on simulated degradation data. using a particle filter algorithm where the collector-emitter The algorithm is developed and tested on the accelerated leakage current was used as the primary precursor of failure. A aging test timescale. In real world operation, the timescale of prognostics approach for power MOSFETs was presented in the degradation process and therefore the RUL predictions [3], where, the threshold voltage was used as a precursor of will be considerable larger. It is hypothesized that even though failure; a particle filter was used in conjunction with an the timescale will be larger, it remains constant through the empirical degradation model. degradation process and the algorithm and model would still Identification of parameters that indicate precursors to apply under the slower degradation process. By using failure in discrete power MOSFETs and IGBTs have received accelerated aging data with actual device measurements and considerable attention in recent years. Several studies have real sensors (no simulated behavior), we are attempting to focused on precursor of failure parameters for discrete IGBTs assess how such algorithm behaves under realistic conditions. under thermal degradation due to power cycling overstress. In [4], collector-emitter voltage was identified as a health which is measured and recorded in-situ. Therefore, the indicator; in [5], the maximum peak of the collector-emitter measured R was normalized to eliminate the case DS(ON) ringing at turn OFF transient was identified as the degradation temperature effects and reflect only changes due to variable; in [6] the switching turn-OFF time was recognized as degradation. Due to manufacturing variability, the pristine failure precursor; and switching ringing was used in [7] to condition R varies from device to device. In order to take DS(ON) characterize degradation. For discrete power MOSFETs, ON- this into account, the normalized R time series is shifted DS(ON) resistance was identified as a precursor of failure for the die- by applying a bias factor representing the pristine condition solder degradation failure mechanism [8, 9]. A shift in value. The resulting trajectory (ΔR ) from pristine DS(ON) threshold voltage was identified as failure precursor due to condition to failure, represents the degradation process due to gate structure degradation fault mode [10]. die-attach failure and represents the increase in R DS(ON) There have been some efforts in the development of through the aging process. degradation models that are a function of the usage/aging time These measurements do not have a fixed sampling rate. based on accelerated life test. For example, empirical On average, there is a transient response measurement every degradation models for model-based prognostics are presented 400 ns. This consists of a snapshot of the transient response in [2] and [3] for discrete IGBTs and power MOSFET which includes one full square waveform cycle. Therefore a respectively. Gate structure degradation modeling of discrete resampling of the curve was carried out to have uniform power MOSFETs under ion impurities has been presented in sampling and a reduced sampling frequency on the failure [11]. precursor trajectory. The signals were filtered by computing the mean of every one minute long window. There are six 2 ACCELERATED LIFE EXPERIMENTS available aged MOSFETs under thermal overstress. Figure 1 The development of prognostics algorithms face similar presents the ΔR trajectories for the six cases. DS(ON) constrains as reliability engineering in that both need information about failure events of critical electronics systems. These data are is rarely ever available. In addition, prognostics requires information about the degradation process leading to an irreversible failure; therefore, it is necessary to record in-situ measurements of key output variables and observable parameters in the accelerated aging process in order to develop and learn failure progression models. Thermal cycling overstress leads to thermo-mechanical stresses in electronics due to mismatch of the coefficient of thermal expansion between different elements in the component’s packaged structure. The accelerated aging applied to the devices presented in this work consists of thermal overstress. Latch-up, thermal run-away, or failure to turn ON due to loss of gate control are considered as failure Figure 1. (cid:2)(cid:2)(cid:2)(cid:3) (cid:3)(cid:2) trajectories for all MOSFETs. conditions. Thermal cycles were induced by power cycling the devices without the use of an external heat sink. The device 3 DEGRADATION MODELING case temperature was measured and directly used as control variable for the thermal cycling application. For power An empirical degradation model is suggested based on the cycling, the applied gate voltage was a square wave signal degradation process observed on ΔR for the six aged with an amplitude of ~15V, a frequency of 1KHz and a duty DS(ON) devices. It can be seen that this process grows exponentially as cycle of 40%. The drain-source was biased at 4Vdc and a a function of time and that the exponential behavior starts at resistive load of 0.2Ω was used on the collector side output of different points in time for different devices. An empirical the device. The aging system used for these experiments is degradation model can be used to model the degradation described in [5], and the accelerated aging methodology is process when a physics-based degradation model is not presented in [8]. available. This methodology has been used for prognostics of In-situ measurements of the drain current (I ) and the D electrolytic capacitors using a Kalman filter [12]. There, the drain to source voltage (V ) are recorded as the device is DS exponential degradation model was posed as a linear first- under aging regime. The ON-state resistance R in this DS(ON) order discrete dynamic system in the form of a state-space application was computed as the ratio of V and I on the DS D model representing the dynamics of the degradation process. ON-state of the square waveform. In the accelerated aging The proposed degradation model for the power MOSFET system, it is not possible to measure junction temperature application is defined as follows. Let (cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:3) (cid:3)(cid:2) be the directly, as a result, the increase in junction temperature is increase in ON-resistance due to aging. observed by monitoring the increase in R . Furthermore, DS(ON) junction temperature is also a function of the case temperature, (cid:2) (cid:2)(cid:2)(cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2) (cid:2) (1) In this model, (cid:2) and (cid:2) are also state variables that change through time. Therefore, the model is a non-linear dynamic where (cid:2) is time and (cid:2) and (cid:2) are model parameters that could system and Bayesian tracking algorithms like the extended be static or estimated on-line as part of the Bayesian tracking Kalman or particle filters are needed for on-line state framework. This model structure is capable of representing the estimation. The forward difference method is used to exponential behavior of the degradation process for the approximate the time derivatives in order to discretize the different devices. Table 1 presents parameter estimation model in equation (3). The first step in the process is results for model (1) based on non-linear least-squares (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2) estimation. The estimate for both parameters is presented (cid:2) (cid:2) (cid:2) (cid:2)(cid:2)(cid:2) (cid:2) (cid:2) (cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2) along with their corresponding sample variance. It is clearly (cid:2) observed that the parameters of the model will be different for Solving for (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2) and applying the method to (cid:2) and (cid:2) different devices. Therefore, the parameters (cid:2) and (cid:2) need to we get: be estimated online in order to ensure accuracy. Figure 2 (cid:2) (cid:2)(cid:2)(cid:2) (cid:2)(cid:2) (cid:2) (cid:2)(cid:2)(cid:2)(cid:3) (cid:2) (cid:2) (cid:2) (cid:2)(cid:2) (cid:2) (cid:2) presents the estimation results for device #36. (cid:2) (cid:2)(cid:2)(cid:2) (cid:2)(cid:2) (cid:2) (cid:2) (5) Device (cid:2) (cid:2) (cid:2)(cid:2)(cid:2) (cid:2)(cid:2)(cid:2) (cid:2) (cid:2)(cid:2)(cid:2) (cid:2)(cid:2) (cid:2) (cid:2) #08 3.70x10-4 3.24x10-2 1.01x10-8 4.27x10-6 4 PROGNOSTICS ALGORITHM DEVELOPMENT #09 7.92x10-4 1.79x10-2 3.50x10-9 1.18x10-7 #11 6.00x10-6 3.56 x10-2 1.10 x10-11 6.20 x10-6 A prognostics algorithm in this application predicts the #12 9.75 x10-7 4.70 x10-2 9.75 x10-7 4.70 x10-2 remaining useful life of a particular power MOSFET device at #14 2.60x10-4 3.60x10-2 1.64x10-9 -4.71x10-9 different points in time through the accelerated life of the #36 2.67x10-3 1.31x10-2 1.02x10-8 2.99x10-8 device. As indicated earlier, ΔR is used in this study as a DS(ON) Table 1. Static parameter estimation results for degradation health indicator feature and as a precursor of failure. The model in equation (1) applied to degradation data in Figure 1. prognostics problem is posed in the following way. • A single feature is used to assess the health state of the device (ΔR ). 0.06 DS(ON) #36 • It is assumed that the die-attached failure mechanism is Exponential fit the only active degradation during the accelerated aging 0.05 experiment. 0.04 • Furthermore, ΔRDS(ON) accounts for the degradation progression from nominal condition through failure. n) 0.03 • Periodic measurements with fixed sampling rate are o Δ Rds( 0.02 • Aav acirliasbpl fea fiolurr Δe RthDrSe(sOhNo).l d of 0.045 in ΔR is used. DS(ON) • The prognostics algorithm will make a prediction of the 0.01 remaining useful life at time t , using all the p measurements up to this point either to estimate the health 0 state at time t in a Bayesian tracking framework. p −0.01 4.1 Extended Kalman filter implementation 0 50 100 150 200 250 Aging time (hours) Figure 2. Non-linear least squares for device #36. Extended Kalman filter allows for the implementation of the Kalman filter algorithm for on-line estimation on non- 3.1 Dynamic degradation model for Bayesian tracking linear dynamic systems [13, 14]. This algorithm has been used in other applications for health state estimation and The degradation model presented in equation (1) is prognostics. The general form of extended Kalman filter is converted into a dynamic model in order to obtain the state- given as; space representation needed for Bayesian tracking. Defining (cid:2) (cid:2)(cid:2)(cid:2) (cid:2)(cid:2) (cid:2) (cid:2) (cid:2)(cid:2) (cid:2) (cid:2)(cid:2) (cid:2) (cid:2) the parameters (cid:2) and (cid:2) be time dependent parameters, then (6) the derivative of (1) is given by, (cid:2) (cid:2) (cid:2)(cid:2) (cid:2) (cid:2) (cid:2)(cid:2) (cid:2) (cid:2) (cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:3)(cid:2)(cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2). (2) where f and h are non-linear equations, (cid:2)(cid:2)(cid:2)(cid:2) is the model noise and (cid:2)(cid:2)(cid:2)(cid:2) is the measurement noise. Noise is considered Defining (cid:2) (cid:2)(cid:2) and (cid:2) = 0, the dynamic model to be normally distributed, with zero mean and known representation is given by, variance (cid:2) and (cid:2) for (cid:2) (cid:2) and (cid:2)(cid:2)(cid:2)(cid:2) respectively. For the prognostics implementation using the discrete (cid:2) (cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2) (cid:2) (cid:2) dynamic degradation model in equation (5), the state variable (cid:2) (cid:2)(cid:2)(cid:2) (3) is defined as (cid:2) (cid:2)(cid:2)(cid:2) (cid:2) (cid:2) (cid:2) (cid:2) (cid:2) (cid:2)(cid:2) (cid:2) (cid:2)(cid:2)(cid:2) (cid:2) (cid:2) (cid:2) (cid:2) (7) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2) (cid:2) Measured (8). TThhee rOefNor-er,e sf iisst aan cvee citso trh vea oluneldy fmuneacstiuorne dg ivvaelnu eb;y t heeqrueaftoiroen, R (ohms)00..0024 FPirletedriecdted tp=150 the measurement equation (cid:2) is given by equation (9). 0 120 140 160 180 200 220 240 (cid:2) (cid:2) (cid:2)(cid:2)(cid:2) (cid:2) (cid:2) (cid:2) (cid:2)(cid:2)(cid:2) (cid:2) (cid:2) (cid:2) (cid:2) (cid:2) (cid:2) (cid:2) (cid:2)(cid:2) (cid:2) (8) ms)0.04 (cid:2)(cid:2)(cid:2) (cid:2) R (oh0.02 tp=170 (cid:2) (cid:2)(cid:2)(cid:2) (cid:2) (9) 0 120 140 160 180 200 220 240 ms)0.04 5 RUL ESTIMATION RESULTS R (oh0.02 tp=190 This section presents the results of the algorithm 0 120 140 160 180 200 220 240 implemented. Four test cases are defined as follows following the leave one out validation concept: ms)0.04 • (cid:2)w(cid:2)i:t hP rtehdei crte RstU oLf othne d deveviciec e#s3 a6n, de sctiommapteu tien itRiaUl Lc oantd ittiimoness R (oh0.02 tp=200 (cid:2)(cid:2) (cid:2) (cid:3)(cid:4)(cid:2)(cid:2)(cid:3)(cid:4)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:2)(cid:3)(cid:4)(cid:2)(cid:4)(cid:2)(cid:2)(cid:2)(cid:4)(cid:2)(cid:4)(cid:2)(cid:3)(cid:3)(cid:2) 0 120 140 160 180 200 220 240 • (cid:2)(cid:2): Predict RUL on device #09, estimate initial conditions Figure 3. Health state (AΔgiRng time (hr)) tracking and RUL with the rest of the devices and compute RUL at times DS(ON) forecasting for test case T . (cid:2)(cid:2) (cid:2) (cid:3)(cid:4)(cid:2)(cid:2)(cid:3)(cid:4)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:2)(cid:3)(cid:4)(cid:2)(cid:4)(cid:2)(cid:2)(cid:2)(cid:4)(cid:2)(cid:4)(cid:2)(cid:3)(cid:3)(cid:2) 1 • (cid:2)(cid:2): Predict RUL on device #08, estimate initial conditions with the rest of the devices and compute RUL at times 0.04 Measured • (cid:2)(cid:2)(cid:2)(cid:2):(cid:2) Pre(cid:4)d(cid:2)ic(cid:2)(cid:4)t (cid:2)R(cid:2)U(cid:3)(cid:2)L(cid:2) o(cid:2)(cid:3)n(cid:3) d(cid:2)e(cid:2)(cid:3)vi(cid:3)c(cid:2)e(cid:2) (cid:2)#(cid:4)14(cid:4),(cid:2) (cid:3)e(cid:3)st(cid:2)im(cid:2)(cid:2)a(cid:4)t(cid:4)e(cid:2) (cid:3)in(cid:4)i(cid:2)tia l conditions R (ohms)0.020 FPirletedriecdted tp=90 with the rest of the devices and compute RUL at times (cid:2)(cid:2) (cid:2) (cid:3)(cid:2)(cid:2)(cid:4)(cid:2)(cid:2)(cid:3)(cid:2)(cid:2)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:3)(cid:4)(cid:2)(cid:2)(cid:2)(cid:4)(cid:4)(cid:2)(cid:3)(cid:3)(cid:2)(cid:2)(cid:2)(cid:4)(cid:4)(cid:2)(cid:3)(cid:4)(cid:2) 50 60 70 80 90 100 110 120 130 140 150 160 RUL estimates are computed by subtracting the time 0.04 cwrhoesnse tsh teh pe rfeadiilcutrieo nth wreassh moladd. eA fsr ommo rteh ed atitma eb ewchoemne ps raevdaiciltaebdl eR, R (ohms)0.020 tp=110 the predictions are expected to become more accurate and 50 60 70 80 90 100 110 120 130 140 150 160 more precise. Table 2 presents the initial conditions for all the test cases. The initial conditions for the parameters and their 0.04 cmoerarens paonndd insga mvaprliea nscteasn daarerd c odmevpiuatteido nb yo ft aktirnagin itnhge sdaemvpiclee R (ohms)0.020 tp=125 parameters in Table 1. The initial value for (cid:2) and its standard 50 60 70 80 90 100 110 120 130 140 150 160 deviation, are computed by using the first ten data points in 0.04 the tr Taeinsti ng d(cid:2)e(cid:2)v ices. (cid:2)(cid:2) (cid:2)(cid:2) (cid:2)(cid:2) (cid:2)(cid:2) (cid:2)(cid:2) R (ohms)0.020 tp=135 (cid:2)(cid:2) 2.6x10-4 3.2x10-4 3.5x10-2 1x10-2 3.8x10-5 2.4x10-3 50 60 70 80 90 100 110 120 130 140 150 160 (cid:2)(cid:2) 2.6x10-4 1.1x10-3 3.5x10-2 1x10-2 3.9x10-5 2.2x10-3 Aging time (hr) (cid:2)(cid:2) 2.6x10-4 1.1x10-3 3.5x10-2 1x10-2 3.8x10-5 2.3x10-3 Figure 4. Health state (ΔRDS(ON)) tracking and RUL (cid:2)(cid:2) 3.7x10-4 1.1x10-3 3.2x10-2 1x10-2 3.8x10-5 2.6x10-3 forecasting for test case T3. Table 2. Initial conditions for the state vector and its Table 3 and Table 4 present the state estimation results for corresponding variance for all the test cases. ΔR and the forecasting of ΔR after measurements DS(ON) DS(ON) are no longer available. Measurements are available up to time Figure 3 and Figure 4 present the RUL estimation results t , these are use by the algorithm to adjust the estate p for test cases T and T respectively. Analysis of the subplots estimation. The prediction portion starts after t . An estimate 1 3 p from top to bottom shows how the prediction progresses as of the expected value of RUL is presented along with the more data becomes available and the device gets closer to end sample standard deviation. These values are computed by of life. It also describes how prognostics consists of periodic Monte Carlo simulation using the last available state estimate RUL predictions through the life of the device. and the state transition equation in (8). The estimation error and relative accuracy (RA) are presented as performance metrics. RA is defined as (cid:3)(cid:2)(cid:2)(cid:3)(cid:2)(cid:2) (cid:2)(cid:2) (cid:3)(cid:4)(cid:2)(cid:2)(cid:2)(cid:3)(cid:4)(cid:2)(cid:2) . (10) The performance of the algorithm depends on the (cid:3)(cid:4)(cid:2)(cid:2) selection of the covariance matrix (cid:2) for the model noise (cid:2)(cid:2)(cid:2)(cid:2) and the variance (cid:2) for the measurement noise (cid:2)(cid:2)(cid:2)(cid:2). Their respective values have been used as tuning parameters for the (cid:2)(cid:2) (cid:2)(cid:2) algorithm. The covariance values were constant for all the (cid:2)(cid:2) RUL RUL tests cases. Error RA Error RA ((cid:2)(cid:3)(cid:4)(cid:2)) ((cid:2)(cid:3)(cid:4)(cid:2)) (cid:2)(cid:2)(cid:2)(cid:2)(cid:3)(cid:2)(cid:2)(cid:2) (cid:2) (cid:2) (cid:5)(cid:8)(cid:14) (cid:8)(cid:12)(cid:4)(cid:2)(cid:8)(cid:2)(cid:6) (cid:10)(cid:14)(cid:4)(cid:2)(cid:13)(cid:10) (cid:3)(cid:5)(cid:4)(cid:13)(cid:8)(cid:14) (cid:3)(cid:5)(cid:4)(cid:11)(cid:14)(cid:11) (cid:15)(cid:3)(cid:4)(cid:15)(cid:5)(cid:15) (cid:11)(cid:14)(cid:4)(cid:5)(cid:13)(cid:5) (cid:2) (cid:2) (cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:3)(cid:2)(cid:2)(cid:2) (cid:2) (cid:7)(cid:5)(cid:4)(cid:5)(cid:13)(cid:14)(cid:9) (cid:7)(cid:5)(cid:4)(cid:2)(cid:12)(cid:13)(cid:9) (cid:2) (cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:3)(cid:2)(cid:2)(cid:2) (cid:5)(cid:2)(cid:14) (cid:2)(cid:15)(cid:4)(cid:15)(cid:3)(cid:14) (cid:5)(cid:11)(cid:4)(cid:13)(cid:5)(cid:11) (cid:11)(cid:8)(cid:4)(cid:3)(cid:14)(cid:14) (cid:2)(cid:2)(cid:4)(cid:15)(cid:10)(cid:11) (cid:15)(cid:15)(cid:4)(cid:3)(cid:8)(cid:12) (cid:11)(cid:14)(cid:4)(cid:12)(cid:15)(cid:13) (cid:2) (cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)(cid:3) (cid:7)(cid:5)(cid:4)(cid:15)(cid:12)(cid:11)(cid:9) (cid:7)(cid:5)(cid:4)(cid:10)(cid:14)(cid:3)(cid:9) (cid:5)(cid:3)(cid:14) (cid:2)(cid:14)(cid:4)(cid:15)(cid:11)(cid:2) (cid:12)(cid:4)(cid:13)(cid:3)(cid:15) (cid:13)(cid:10)(cid:4)(cid:3)(cid:14)(cid:15) (cid:8)(cid:3)(cid:4)(cid:5)(cid:13)(cid:8) (cid:15)(cid:5)(cid:4)(cid:11)(cid:8)(cid:8) (cid:3)(cid:11)(cid:4)(cid:12)(cid:11)(cid:13) REFERENCES (cid:7)(cid:5)(cid:4)(cid:15)(cid:8)(cid:12)(cid:9) (cid:7)(cid:5)(cid:4)(cid:14)(cid:5)(cid:5)(cid:9) (cid:5)(cid:11)(cid:14) (cid:8)(cid:2)(cid:4)(cid:2)(cid:8)(cid:13) (cid:8)(cid:4)(cid:2)(cid:12)(cid:14) (cid:12)(cid:14)(cid:4)(cid:13)(cid:8)(cid:11) (cid:10)(cid:2)(cid:4)(cid:3)(cid:5)(cid:12) (cid:15)(cid:15)(cid:4)(cid:15)(cid:3)(cid:3) (cid:3)(cid:5)(cid:4)(cid:2)(cid:8)(cid:3) 1. Celaya, J.R., A. Saxena, S. Saha, and K. Goebel, (cid:7)(cid:5)(cid:4)(cid:5)(cid:15)(cid:12)(cid:9) (cid:7)(cid:14)(cid:4)(cid:3)(cid:13)(cid:15)(cid:9) "Prognostics of Power MOSFETs under Thermal Stress (cid:5)(cid:13)(cid:14) (cid:10)(cid:11)(cid:4)(cid:10)(cid:8)(cid:15)(cid:6) (cid:15)(cid:4)(cid:13)(cid:14)(cid:15) (cid:12)(cid:10)(cid:4)(cid:14)(cid:15)(cid:14) (cid:15)(cid:11)(cid:4)(cid:14)(cid:8)(cid:14) (cid:15)(cid:14)(cid:4)(cid:13)(cid:2)(cid:11) (cid:2)(cid:3)(cid:4)(cid:8)(cid:3)(cid:5) Accelerated Aging using Data-Driven and Model-Based (cid:7)(cid:5)(cid:4)(cid:14)(cid:14)(cid:5)(cid:9) (cid:7)(cid:14)(cid:4)(cid:2)(cid:8)(cid:13)(cid:9) Methodologies", in Annual Conference of the Prognostics (cid:15)(cid:12)(cid:4)(cid:15)(cid:3)(cid:2) (cid:15)(cid:5)(cid:4)(cid:10)(cid:5)(cid:12) (cid:5)(cid:12)(cid:14) (cid:14)(cid:4)(cid:13)(cid:13)(cid:12) (cid:12)(cid:11)(cid:4)(cid:14)(cid:2)(cid:15) (cid:5)(cid:3)(cid:4)(cid:2)(cid:11)(cid:13) (cid:2)(cid:3)(cid:4)(cid:15)(cid:3)(cid:10) and Health Management Society2011. (cid:7)(cid:14)(cid:4)(cid:13)(cid:5)(cid:3)(cid:9) (cid:7)(cid:14)(cid:4)(cid:8)(cid:11)(cid:12)(cid:9) 2. Saha, B., J.R. Celaya, P.F. Wysocki, and K.F. Goebel. (cid:15)(cid:5)(cid:4)(cid:10)(cid:10)(cid:3) (cid:5)(cid:12)(cid:4)(cid:2)(cid:8)(cid:13) (cid:5)(cid:12)(cid:2) (cid:10)(cid:4)(cid:11)(cid:2)(cid:11) (cid:13)(cid:2)(cid:4)(cid:14)(cid:2)(cid:13) (cid:5)(cid:10)(cid:4)(cid:10)(cid:10)(cid:12) (cid:2)(cid:12)(cid:4)(cid:8)(cid:3)(cid:15) "Towards prognostics for electronics components". in (cid:7)(cid:14)(cid:4)(cid:3)(cid:8)(cid:3)(cid:9) (cid:7)(cid:14)(cid:4)(cid:8)(cid:3)(cid:3)(cid:9) Aerospace conference, 2009 IEEE. 2009. (cid:5)(cid:2)(cid:4)(cid:3)(cid:11)(cid:10) (cid:5)(cid:3)(cid:4)(cid:12)(cid:13)(cid:10) (cid:15)(cid:14)(cid:14) (cid:8)(cid:4)(cid:8)(cid:11)(cid:3) (cid:11)(cid:11)(cid:4)(cid:11)(cid:13)(cid:5) (cid:5)(cid:14)(cid:4)(cid:12)(cid:14)(cid:14) (cid:3)(cid:14)(cid:4)(cid:12)(cid:10)(cid:3) 3. Saha, S., J.R. Celaya, V. Vashchenko, S. Mahiuddin, and (cid:7)(cid:14)(cid:4)(cid:2)(cid:5)(cid:11)(cid:9) (cid:7)(cid:14)(cid:4)(cid:8)(cid:10)(cid:14)(cid:9) K.F. Goebel. "Accelerated Aging with Electrical (cid:5)(cid:14)(cid:4)(cid:12)(cid:2)(cid:10)(cid:6) (cid:5)(cid:8)(cid:4)(cid:5)(cid:3)(cid:12) (cid:15)(cid:14)(cid:2) (cid:8)(cid:4)(cid:5)(cid:12)(cid:13) (cid:11)(cid:15)(cid:4)(cid:15)(cid:13)(cid:14) (cid:13)(cid:4)(cid:11)(cid:10)(cid:14) (cid:3)(cid:5)(cid:4)(cid:13)(cid:13)(cid:3) Overstress and Prognostics for Power MOSFETs". in (cid:7)(cid:14)(cid:4)(cid:8)(cid:15)(cid:11)(cid:9) (cid:7)(cid:14)(cid:4)(cid:10)(cid:13)(cid:12)(cid:9) IEEE EnergyTech 2011. 2011. (cid:3)(cid:4)(cid:3)(cid:14)(cid:13)(cid:6) (cid:5)(cid:5)(cid:4)(cid:11)(cid:10)(cid:11) (cid:15)(cid:5)(cid:14) (cid:10)(cid:4)(cid:2)(cid:15)(cid:10) (cid:3)(cid:2)(cid:4)(cid:15)(cid:11)(cid:15) (cid:3)(cid:4)(cid:5)(cid:3)(cid:5) (cid:3)(cid:2)(cid:4)(cid:2)(cid:13)(cid:11) 4. Patil, N., J. Celaya, D. Das, K. Goebel, and M. Pecht, (cid:7)(cid:14)(cid:4)(cid:10)(cid:5)(cid:8)(cid:9) (cid:7)(cid:14)(cid:4)(cid:10)(cid:2)(cid:12)(cid:9) "Precursor Parameter Identification for Insulated Gate Table 3. RUL estimation results for test cases(cid:2)(cid:2)(cid:2) and (cid:2)(cid:2). The Bipolar Transistor (IGBT) Prognostics". IEEE mean RUL and the standard deviation are presented in the Transactions on Reliability, 2009. 58(2): p. 276. first column; error and relative accuracy are presented in the 5. Sonnenfeld, G., K. Goebel, and J.R. Celaya. "An agile second and third column of each test case. accelerated aging, characterization and scenario simulation system for gate controlled power transistors". (cid:2)(cid:2) (cid:2)(cid:2) in IEEE AUTOTESTCON 2008. 2008. (cid:2)(cid:2) RUL Error RA RUL Error RA 6. Brown, D., M. Abbas, A. Ginart, I. Ali, P. Kalgren, and ((cid:2)(cid:3)(cid:4)(cid:2)) ((cid:2)(cid:3)(cid:4)(cid:2)) G. Vachtsevanos. "Turn-off Time as a Precursor for Gate (cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:3)(cid:6) (cid:2)(cid:3)(cid:4)(cid:5)(cid:13)(cid:14)(cid:6) 80 (cid:3)(cid:4)(cid:10)(cid:11)(cid:12) (cid:13)(cid:12)(cid:4)(cid:13)(cid:5)(cid:3) (cid:11)(cid:4)(cid:5)(cid:3)(cid:15) (cid:13)(cid:13)(cid:4)(cid:11)(cid:14)(cid:8) Bipolar Transistor Latch-up Faults in Electric Motor (cid:7)(cid:5)(cid:4)(cid:8)(cid:3)(cid:5)(cid:9)(cid:6) (cid:7)(cid:5)(cid:4)(cid:10)(cid:12)(cid:5)(cid:9)(cid:6) Drives". in Annual Conference of the Prognostics and (cid:8)(cid:11)(cid:4)(cid:2)(cid:15)(cid:5)(cid:6) (cid:2)(cid:10)(cid:4)(cid:3)(cid:13)(cid:14)(cid:6) 90 (cid:2)(cid:4)(cid:5)(cid:5)(cid:11) (cid:12)(cid:14)(cid:4)(cid:15)(cid:11)(cid:11) (cid:16)(cid:14)(cid:4)(cid:15)(cid:3)(cid:15) (cid:12)(cid:12)(cid:4)(cid:2)(cid:5)(cid:14) Health Management Society 2010. 2010. (cid:7)(cid:5)(cid:4)(cid:14)(cid:5)(cid:13)(cid:9)(cid:6) (cid:7)(cid:5)(cid:4)(cid:15)(cid:8)(cid:11)(cid:9)(cid:6) 7. Ginart, A., M. Roemer, P. Kalgren, and K. Goebel. (cid:8)(cid:3)(cid:4)(cid:3)(cid:14)(cid:2)(cid:6) (cid:3)(cid:5)(cid:4)(cid:3)(cid:14)(cid:12)(cid:6) 100 (cid:16)(cid:8)(cid:4)(cid:14)(cid:14)(cid:5) (cid:12)(cid:14)(cid:4)(cid:3)(cid:5)(cid:3) (cid:16)(cid:5)(cid:13)(cid:4)(cid:15)(cid:5)(cid:13) (cid:2)(cid:13)(cid:4)(cid:14)(cid:15)(cid:3) "Modeling Aging Effects of IGBTs in Power Drives by (cid:7)(cid:14)(cid:4)(cid:12)(cid:13)(cid:8)(cid:9)(cid:6) (cid:7)(cid:5)(cid:4)(cid:2)(cid:15)(cid:14)(cid:9)(cid:6) Ringing Characterization". in IEEE International (cid:2)(cid:8)(cid:4)(cid:13)(cid:13)(cid:5)(cid:6) (cid:3)(cid:2)(cid:4)(cid:14)(cid:11)(cid:2)(cid:6) 110 (cid:7)(cid:5)(cid:4)(cid:15)(cid:10)(cid:11)(cid:9)(cid:6) (cid:16)(cid:15)(cid:15)(cid:4)(cid:10)(cid:15)(cid:12) (cid:10)(cid:5)(cid:4)(cid:2)(cid:11)(cid:10) (cid:7)(cid:5)(cid:4)(cid:11)(cid:2)(cid:11)(cid:9)(cid:6) (cid:16)(cid:10)(cid:5)(cid:4)(cid:11)(cid:12)(cid:13) (cid:8)(cid:4)(cid:13)(cid:14)(cid:11) Conference on Prognostics and Health Management. 2008. (cid:8)(cid:14)(cid:4)(cid:15)(cid:13)(cid:11)(cid:6) (cid:8)(cid:11)(cid:4)(cid:11)(cid:14)(cid:8)(cid:6) 120 (cid:7)(cid:14)(cid:4)(cid:13)(cid:2)(cid:13)(cid:9)(cid:6) (cid:16)(cid:5)(cid:11)(cid:4)(cid:3)(cid:10)(cid:5) (cid:15)(cid:15)(cid:4)(cid:5)(cid:14)(cid:14) (cid:7)(cid:5)(cid:4)(cid:14)(cid:8)(cid:2)(cid:9)(cid:6) (cid:16)(cid:15)(cid:8)(cid:4)(cid:10)(cid:14)(cid:8) (cid:16)(cid:10)(cid:4)(cid:13)(cid:8)(cid:11) 8. Celaya, J., A. Saxena, P. Wysocki, S. Saha, and K. Goebel. "Towards Prognostics of Power MOSFETs: (cid:10)(cid:14)(cid:4)(cid:10)(cid:5)(cid:14)(cid:6) (cid:10)(cid:3)(cid:4)(cid:15)(cid:2)(cid:13)(cid:6) 125 (cid:7)(cid:14)(cid:4)(cid:2)(cid:13)(cid:3)(cid:9)(cid:6) (cid:16)(cid:5)(cid:15)(cid:4)(cid:3)(cid:3)(cid:3) (cid:15)(cid:13)(cid:4)(cid:5)(cid:3)(cid:13) (cid:7)(cid:14)(cid:4)(cid:11)(cid:13)(cid:5)(cid:9)(cid:6) (cid:16)(cid:5)(cid:11)(cid:4)(cid:13)(cid:13)(cid:2) (cid:15)(cid:4)(cid:13)(cid:5)(cid:13) Accelerated Aging and Precursors of Failure". in Annual (cid:15)(cid:5)(cid:4)(cid:3)(cid:5)(cid:8)(cid:6) (cid:15)(cid:8)(cid:4)(cid:2)(cid:2)(cid:14)(cid:6) Conference of the Prognostics and Health Management 130 (cid:7)(cid:14)(cid:4)(cid:8)(cid:14)(cid:15)(cid:9)(cid:6) (cid:16)(cid:13)(cid:4)(cid:12)(cid:12)(cid:12) (cid:15)(cid:13)(cid:4)(cid:11)(cid:3)(cid:3) (cid:7)(cid:14)(cid:4)(cid:8)(cid:12)(cid:10)(cid:9)(cid:6) (cid:16)(cid:5)(cid:5)(cid:4)(cid:5)(cid:3)(cid:14) (cid:5)(cid:3)(cid:4)(cid:11)(cid:10)(cid:11) Society 2010. 2010. (cid:5)(cid:15)(cid:4)(cid:3)(cid:5)(cid:2)(cid:6) (cid:5)(cid:10)(cid:4)(cid:8)(cid:12)(cid:12)(cid:6) 9. Celaya, J.R., N. Patil, S. Saha, P. Wysocki, and K. 135 (cid:7)(cid:14)(cid:4)(cid:15)(cid:10)(cid:13)(cid:9)(cid:6) (cid:16)(cid:8)(cid:4)(cid:12)(cid:12)(cid:3) (cid:10)(cid:8)(cid:4)(cid:2)(cid:10)(cid:12) (cid:7)(cid:14)(cid:4)(cid:15)(cid:11)(cid:15)(cid:9)(cid:6) (cid:16)(cid:2)(cid:4)(cid:14)(cid:11)(cid:13) (cid:10)(cid:12)(cid:4)(cid:2)(cid:11)(cid:11) Goebel. "Towards Accelerated Aging Methodologies and (cid:2)(cid:4)(cid:11)(cid:12)(cid:2)(cid:6) (cid:3)(cid:4)(cid:2)(cid:12)(cid:12)(cid:6) Health Management of Power MOSFETs (Technical 140 (cid:16)(cid:10)(cid:4)(cid:5)(cid:3)(cid:15) (cid:16)(cid:15)(cid:14)(cid:4)(cid:5)(cid:15)(cid:2) (cid:16)(cid:10)(cid:4)(cid:5)(cid:12)(cid:11) (cid:3)(cid:4)(cid:14)(cid:13)(cid:5) (cid:7)(cid:14)(cid:4)(cid:5)(cid:8)(cid:5)(cid:9)(cid:6) (cid:7)(cid:14)(cid:4)(cid:5)(cid:3)(cid:13)(cid:9)(cid:6) Brief)". in Annual Conference of the Prognostics and Table 4. RUL estimation results for test cases(cid:2)(cid:2)(cid:2) and T4. The Health Management Society 2009. 2009. mean RUL and the standard deviation are presented in the 10. Celaya, J.R., P. Wysocki, V. Vashchenko, S. Saha, and K. first column; error and relative accuracy are presented in the Goebel. "Accelerated aging system for prognostics of second and third column of each test case. power semiconductor devices". in 2010 IEEE AUTOTESTCON. 2010. AAAI and ASME. 11. Ginart, A.E., I.N. Ali, J.R. Celaya, P.W. Kalgren, S.D. Poll, and M.J. Roemer. "Modeling SiO2 Ion Impurities Chetan Kulkarni Aging in Insulated Gate Power Devices Under ISIS, Vanderbilt University Temperature and Voltage Stress". in Annual Conference 1025 16th Avenue of the Prognostics and Health Management Society 2010. Nashville, TN, 37212, USA 2010. email: [email protected] 12. Celaya, J., C. Kulkarni, G. Biswas, and K. Goebel. "Towards Prognostics of Electrolytic Capacitors". in Chetan S Kulkarni is a Research Assistant at ISIS, Vanderbilt AIAA Infotech@Aerospace. 2011. St. Louis, MO. University. He received the M.S. degree in EECS from Vanderbilt 13. Rasmussen, C.E. and C.K.I. Williams, "Gaussian University, Nashville, TN, in 2009, where he is currently a PhD Processes for Machine Learning"2006. MIT Press. student. 14. Meinhold, R.J. and N.D. Singpurwalla, "Understanding the Kalman Filter". The American Statistician, 1983. Sankalita Saha, PhD 37(2): p. 123-127. NASA Ames Research Center (MCT) Prognostics Center of Excellence MS 269-4 Moffett Field, CA 94035, USA BIOGRAPHIES email: [email protected] José R. Celaya, PhD NASA Ames Research Center (SGT Inc.) Sankalita Saha is a research scientist with Mission Critical Technologies at the Prognostics Center of Excellence, NASA Ames Prognostics Center of Excellence Research Center. She received the M.S. and PhD degrees in MS 269-4 Electrical Engineering from University of Maryland, College Park in Moffett Field, CA 94035, USA 2007. Prior to that she obtained her B.Tech (Bachelor of Technology) degree in Electronics and Electrical Communications Engineering email: [email protected] from the Indian Institute of Technology, Kharagpur in 2002. José R. Celaya is a research scientist with SGT Inc. at the Prognostics Kai Goebel, PhD Center of Excellence, NASA Ames Research Center. He received a NASA Ames Research Center PhD degree in Decision Sciences and Engineering Systems in 2008, a Prognostics Center of Excellence M. E. degree in Operations Research and Statistics in 2008, a M. S. degree in Electrical Engineering in 2003, all from Rensselaer MS 269-1 Polytechnic Institute, Troy New York; and a B. S. in Cybernetics Moffett Field, CA 94035, USA Engineering in 2001 from CETYS University, México. email: [email protected] Abhinav Saxena, PhD NASA Ames Research Center (SGT Inc.) Kai Goebel received the degree of Diplom-Ingenieur from the Prognostics Center of Excellence Technische Universitt Mnchen, Germany in 1990. He received the MS 269-4 M.S. and PhD from the University of California at Berkeley in 1993 and 1996, respectively. Dr. Goebel is a senior scientist at NASA Moffett Field, CA 94035, USA Ames Research Center where he leads the Diagnostics and Prognostics groups in the Intelligent Systems division. In addition, he email: [email protected] directs the Prognostics Center of Excellence and he is the technical lead for Prognostics and Decision Making of NASA’s System-wide Abhinav Saxena is a Research Scientist with SGT Inc. at the Safety and Assurance Technologies Program. He worked at General Prognostics Center of Excellence NASA Ames Research Center, Electric’s Corporate Research Center in Niskayuna, NY from 1997 to Moffett Field CA. His research focus lies in developing and 2006 as a senior research scientist. He has carried out applied evaluating prognostic algorithms for engineering systems using soft research in the areas of artificial intelligence, soft computing, and computing techniques. He is a PhD in Electrical and Computer information fusion. His research interest lies in advancing these Engineering from Georgia Institute of Technology, Atlanta. He techniques for real time monitoring, diagnostics, and prognostics. He earned his B.Tech in 2001 from Indian Institute of Technology (IIT) holds 15 patents and has published more than 200 papers in the area Delhi, and Masters Degree in 2003 from Georgia Tech. Abhinav has of systems health management. been a GM manufacturing scholar and is also a member of IEEE,