978-1-4577-0557-1/12/$26.00 ©2012 IEEE 1

Data-driven Fault Diagnosis in a Hybrid Electric Vehicle

Regenerative Braking System Chaitanya Sankavaram

1, Bharath Pattipati

1, Krishna Pattipati

1, Yilu Zhang

2, Mark Howell

2, and Mutasim Salman

2

1Department of Electrical and Computer Engineering,

University of Connecticut, 371 Fairfield Road, U-2157, Storrs, CT 06269, USA 2GM R&D, General Motors Company, 30500 Mound Rd, Warren, MI 48090, USA

{chaitanya, krishna}@engr.uconn.edu

Abstract— Regenerative braking is one of the most

promising and environmentally friendly technologies used in

electric and hybrid electric vehicles to improve energy

efficiency and vehicle stability. In this paper, we discuss a

systematic data-driven process for detecting and diagnosing

faults in the regenerative braking system of hybrid electric

vehicles. The process involves data reduction techniques,

exemplified by multi-way partial least squares, multi-way

principal component analysis, for implementation in memory-

constrained electronic control units and well-known fault

classification techniques based on reduced data, such as

support vector machines, k-nearest neighbor, partial least

squares, principal component analysis and probabilistic neural

network, to isolate faults in the braking system. The results

demonstrate that highly accurate fault diagnosis is possible

with the pattern recognition-based techniques. The process can

be employed for fault analysis in a wide variety of systems,

ranging from automobiles to buildings to aerospace systems.

TABLE OF CONTENTS

1. INTRODUCTION .................................................................1 2. MODELING OF REGENERATIVE BRAKING SYSTEM ........2 3. FAULT UNIVERSE AND MONITORED SIGNALS ................3 4. FAULT DETECTION AND DIAGNOSIS PROCESS ................4 5. EXPERIMENTAL RESULTS ................................................8 6. CONCLUSIONS ..................................................................9 ACKNOWLEDGMENTS ..........................................................9 REFERENCES ........................................................................9 BIOGRAPHIES .....................................................................10

1. INTRODUCTION

Hybrid electric vehicles (HEVs) employ regenerative

braking to improve fuel economy, enable energy

regeneration and provide environmental protection [1]. The

primary function of a regenerative braking system (RBS) is

to convert kinetic energy into electrical energy and store it in

batteries during braking mode for later use in propelling the

vehicle (see Fig 1) [2]. Failures in a regenerative braking

system may significantly degrade the performance and

efficiency of vehicles. Hence, an intelligent diagnostic

process is crucial to quickly detect and isolate faults in order

to aid in vehicle health monitoring and, consequently,

enhance the reliability of vehicular systems.

Diagnostic methods have mainly evolved upon three

major paradigms, viz., physics-based modeling, data-driven

and knowledge-based approaches. The physics-based

modeling approach employs consistency checks between the

sensed measurements and the outputs of a mathematical

model. The expectation is that inconsistencies are large in

the presence of malfunctions and small in the presence of

normal disturbances, noise and modeling errors. Two main

methods of generating the consistency checks are based on

observers (e.g., Kalman filters, reduced-order unknown

input observers, interacting multiple models, particle filters)

and parity relations (dynamic consistency checks among

measured variables stemming from hardware or information

redundancy relations). A data-driven approach is preferred

when the system monitoring data for nominal and degraded

conditions is available. Neural network and statistical

classification methods are illustrative of data-driven

techniques. The knowledge-based approach uses graphical

models such as dependency graphs (digraphs), Petri nets,

multi-signal (multi-functional) flow graphs, and Bayesian

networks for diagnostic knowledge representation and

inference [3][4][5].

In this paper, we develop a systematic data-driven fault

detection and diagnosis (FDD) process for diagnosing faults

in a regenerative braking system. For FDD analysis,

Powertrain System Analysis Toolkit (PSAT) [6], a vehicle

simulation software tool, was used to create a

Matlab/Simulink® model of RBS with series-parallel

drivetrain configuration. The fault diagnosis process

involves data reduction techniques, such as multi-way partial

least squares, multi-way principal component analysis [7],

Fig. 1: Regenerative Braking System - Energy Flow

Diagram in HEVs

2

Fig. 2: Regenerative Braking System Matlab/Simulink® model from PSAT

and well-known fault classification techniques based on

reduced data, exemplified by support vector machines

(SVM), k-nearest neighbor (KNN), partial least squares

(PLS), principal component analysis (PCA) and

probabilistic neural network (PNN) to isolate faults in the

braking system [8][9]. The current electronic control units

(ECUs) for control and diagnosis have 1-2 MB of memory

and 24-50 MHz of processor speed (due to cost constraints);

therefore, intelligent data reduction techniques are crucial

for effective fault diagnosis in automotive systems. Although

the fault detection and diagnosis approach presented in this

paper is targeted at automotive regenerative braking

systems, it has direct applications to aerospace, building and

other cyber-physical systems.

Organization of the paper

The paper is organized as follows. In Section 2, we briefly

present the nominal model of Regenerative Braking System

(RBS). Section 3 describes the fault universe and the

monitored variables in the RBS model. Section 4 presents

the data-driven FDD process to detect and isolate faults in

the RBS. Section 5 describes the experimental results and

finally, Section 6 concludes the paper with a summary.

2. MODELING OF REGENERATIVE BRAKING

SYSTEM

We modeled the regenerative braking system with a

series-parallel drivetrain [10] configuration using PSAT

software (see Fig 2). The details of the vehicle drivetrain

components are given in Table I.

The RBS model consists of a driver model, powertrain

controller, component controller, and the powertrain model.

The powertrain controller is the supervisory controller

making the high-level decisions that affect the general state

of the powertrain (e.g. engine on/off), the operating mode of

the vehicle (e.g. propelling, regenerative braking etc.), and

accordingly deliver the torque requests to the component

controller. Subsequently, the torque requests are converted

into component commands at the component controller.

These commands are treated as the actuator commands by

the individual components in the powertrain model to

achieve the requested torque and consequently, report the

system status (e.g., engine speed, battery state of charge) to

the supervisory controller. The models are described briefly

in the following subsections [11]. Mathematical details will

be presented in [12].

3

Fig. 3: Vehicle Powertrain Model with Series-Parallel Configuration

Driver Model:

The driver model simulates the drive cycles (e.g., Urban

Dynamometer Driving Schedule (UDDS) [13]), by setting

accelerator and brake pedal positions to achieve the desired

vehicle speed. The output from this block is the driver’s

torque demand at the wheels, which constitutes the input to

the powertrain controller (PTC).

Powertrain Controller:

The key functionality of the PTC is to send commands to

the component controller so that the desired torque demands

are met. The PTC has the control strategies for (a)

propelling, when torque at the wheels is positive, (b)

shifting, i.e., gear selection strategy, and (c) braking, when

torque at the wheels is negative. In addition, PTC also

computes the physical limits for each of the components, for

instance, the maximum available torque for the engine and

the maximum propelling torque for the generator.

Component Controller:

In the component controller, the desired torque demands

from the PTC are converted into component commands,

e.g., the desired torque from the engine is converted into

percentage throttle command. These commands are

processed by the component blocks in the powertrain model.

Powertrain Model:

This model comprises all the components that mimic the

behavior of hardware components, such as the engine, the

battery and the motor. Fig 3 shows the individual

component blocks in the powertrain model (with a series-

parallel drivetrain configuration).

Table I. Series-Parallel Drivetrain Configuration

COMPONENT SPECIFICATIONS

Motor1 Permanent Magnet (PM)

Electric Motor with

Continuous Power = 25KW

Peak Power = 50 KW

Motor2 UQM Power-phase PM

motor

Continuous Power = 55KW

Peak Power = 100 KW

Engine Gasoline Engine

1.497 liter

57 KW

Energy Storage SAFT Li-ion Battery

Capacity = 6 Ah

Number of cells = 75

Gearbox 2 Gear Manual Transmission

Gear ratios = 1.86 and 1

Wheel Axle 2 Wheel drive

Vehicle Vehicle body mass = 800 Kg

Frontal area = 1.8 m2

Drag coefficient = 0.38

3. FAULT UNIVERSE AND MONITORED SIGNALS

The RBS model is subjected to 12 sensor and parametric

faults listed in Table II. There are 25 signals that are

monitored in the braking system including (a) sensor

signals, such as temperature, speed, and current

measurements from the hardware components in the

powertrain model; (b) motor, wheel, and engine torque

demands sent from the powertrain controller to the

component controller; and (c) component commands sent

from the component controller to the hardware components

in the powertrain model. Table III shows the list of

4

monitored signals in the Regenerative Braking System.

The sensor and parametric faults are injected via

simulation-based fault injection experiments. The most

common faults in the sensors are additive faults; these are

the discrepancies between the measured and the true values

of the monitored signals. Therefore, the sensor faults (F1,

F2, and F4-F9) are simulated as additive biases on the

measured signals. For instance, a 10% increase in the engine

speed is used to model the engine speed sensor fault.

Similarly, the parametric faults F3 and F10 are simulated as

10% increase in their nominal values, whereas 10% decrease

in the wheel radius is considered as wheel radius fault (F12).

Mathematically, the fault scenarios are simulated using the

following equation,

(1 ).faulty nominalX X (1)

In (1), Xfaulty is the parameter value under faulty condition,

Xnominal is the nominal parameter value and Δ is the percent

change in the parameter value (fault severity). Our approach

is generic and does not preclude more sophisticated fault

models (e.g., coupling faults, multiplicative faults, stuck

with constant bias fault, abrupt/incipient faults,

persistent/intermittent faults and so on [14]).

The simulations are conducted for 1400 seconds and the

data is collected every 0.1 seconds. Hence, there are a total

of 14001 data points for each nominal and faulty case. The

data is arranged in the form of a tensor as I J KX R where I, J and K are the nominal and fault

cases, measured (monitored) signals, and time samples,

respectively (see Fig 4).

Fig. 4: Three Dimensional Data Arrangement

Table II. Fault List for Regenerative Braking System

FAULT FAULT DESCRIPTION

F1 Battery Current Sensor Fault

F2 Battery Temperature Sensor Fault

F3 Clutch Inertia Fault

F4 Engine Speed Sensor Fault

F5 Motor1 Current Sensor Fault

F6 Motor1 Speed Sensor Fault

F7 Motor2 Current Sensor Fault

F8 Motor2 Speed Sensor Fault

F9 Vehicle Speed Sensor Fault

F10 Wheel Inertia Fault

F11 Battery SOC Fault

F12 Wheel Radius Fault

Table III. List of Monitored Signals in RBS

Monitored

Signal Signal Description

S1 Battery SOC

S2 Motor 2 Torque Demand

S3 Wheel Torque Demand

S4 Motor 1 Torque Demand

S5 Engine Torque Demand

S6 Battery Temperature

S7 Battery Current

S8 Driver Torque Demand

S9 Motor1 Command

S10 Gearbox Speed

S11 Wheel Input Speed

S12 Wheel Output Speed

S13 Wheel Torque

S14 Vehicle Linear Speed

S15 Motor1 Speed

S16 Motor1 Current

S17 Clutch Speed

S18 Engine Command

S19 Motor2 Command

S20 Motor2 Speed

S21 Motor2 Current

S22 Engine Speed

S23 Clutch Speed

S24 Mechanical Accessory Torque

S25 Wheel Command

4. FAULT DETECTION AND DIAGNOSIS PROCESS

The data-driven FDD process for the RBS system consists

of an offline training phase and an online testing phase, as

shown in Fig. 5. The process primarily involves three major

steps: data reduction, fault detection and fault

classification.

In the off-line training phase, data reduction techniques

are employed on the residuals (deviation of actual

measurements from the expected ones) from different fault

scenarios; and the reduced data is then used to train the fault

5

Fig. 5: Fault Detection and Diagnosis Process for Regenerative Braking System

classifiers (SVM, KNN, PLS, PCA, and PNN). In the online

testing phase, the trained classifiers are used to classify the

faults based on reduced data. The following subsections

describe briefly the steps involved in the FDD process.

Data Reduction Techniques

Data reduction techniques are of significant importance in

real world applications to overcome the issue with high-

dimensional datasets (due to multiple modes of system

operation and sensor data collected over time) [15][16].

Often, all the measurements are not salient for understanding

the essential phenomena of interest, and features extracted

using the data reduction techniques enable real-time

implementation of data-driven diagnostic algorithms via

compact memory footprint, improved computational

efficiency and generally enhanced diagnostic accuracy [17].

Statistical data reduction techniques, such as multi-way

principal component analysis (MPCA) and multi-way partial

least squares (MPLS), are among the widely investigated

data reduction techniques [17][18]. These techniques reduce

the dimensionality of datasets by transforming the data to a

lower-dimensional feature space, and this reduced space

often gives information about the salient structure of the

high-dimensional data space. In other words, a number of

correlated variables in the original data are transformed into

a smaller number of uncorrelated new variables in the

reduced space, called principal components or the score

vectors. These reduced variables capture good amount of

variation in the data. The MPCA is used to reduce the

dimensionality of data, and produces a representation that

preserves the correlation structures among the monitored

variables. The PCA is optimal in terms of capturing the

variation in data. The MPLS is another dimensionality

reduction technique that considers both pattern (independent

data) and class (response) information. MPLS technique is

widely used for its ability to enhance classification accuracy

on high-dimensional datasets, and its computational

efficiency. The reduced data can be processed with

classifiers for categorizing the various fault classes.

Furthermore, both MPLS and MPCA techniques can be used

for fault detection. For example, in the MPLS technique, if

the distance of the score vectors is close to the origin of the

new reduced feature space, then the system is judged to be

in the nominal state. The distance to these score vectors

from the origin is computed via Hotelling statistic [7]. In this

paper, we employed the MPLS technique on the residuals

primarily for data reduction (from 559 MB to 12 KB).

Fault detection Techniques

Fault detection is performed by monitoring the

amount/rate of deviation of a parameter from its nominal

value, also known as residuals. We employ cumulative sum

(CUSUM) test [19], and simple thresholds for the detection

of faults. Here, the residuals of monitored variables are

computed at each measurement time, and the cumulative

sum of squared residuals is calculated. Let Sik be the

cumulative sum of squared residual of monitored signal i at

time k, mathematically written as,

6

2

1

k

ik ij

j

S r

(2)

where rij is the residual of the signal i at jth

time instant, and,

is the measurement difference between the actual sensor

output and nominal no-fault output. In the no-fault case, the

expected value of rij is zero, whereas in the presence of a

fault, the rij does not have zero mean. Hence, a simple

threshold test on the cumulative sum of the squared residuals

can determine whether the system is in a faulty state or

nominal state i.e., if Sik exceeds a threshold T then a fault is

detected (see equation (3)).

fault detected

no fault detected

ik

ik

S T

S T

(3)

Fault Classification

Statistical and pattern classification techniques such as

support vector machines, k-nearest neighbors, principal

component analysis, partial least squares, and probabilistic

neural networks are employed for isolating faults in the RBS

model. A brief explanation of these techniques is given in

the following subsections.

Support Vector Machine—Support vector machine is one of

the most widely used supervised learning algorithms for

classification. Given a set of training samples belonging to

different classes, support vector machines find an optimal

decision boundary (also called hyperplane) that maximizes

the margin between the classes [20]. Therefore, SVM is also

known as the maximum-margin classifier. The vectors that

define the optimal separating hyperplane are known as the

support vectors. These support vectors lie closest to the

decision surface and can be used to estimate the fault class

of a test feature vector. In the case of non-linear

classification, where a linear boundary is not appropriate, a

kernel function is used for mapping the data onto a higher

dimensional space and the optimal hyperplane is then

constructed. Some of the kernel functions that are commonly

used are polynomial, Gaussian, radial basis functions, and

hyperbolic tangent.

k - Nearest Neighbor—The KNN classifier is a simple non-

parametric method that classifies test vectors based on the

samples from the training data [8]. The classifier finds the k-

nearest points to a test vector from the training data, and the

class with the maximum a posteriori probability within those

k points is declared as the most-likely class. Normally, k is

chosen as an odd number to avoid ties. Mathematically, the

posterior class probabilities P(ci|xnew) are given by,

| ii inew

kP c x p c

k (4)

where ki is the number of vectors belonging to class ci within

the k-nearest points. P(ci) is the prior probability of class ci.

A new test sample xnew is assigned to the class ci with the

highest a posteriori class probability P(ci|xnew).

Principal Component Analysis—PCA is a multivariate

statistical procedure that transforms the training data into a

lower-dimensional space by transforming a number of

correlated variables into a smaller number of uncorrelated

new variables called principal components. These

components represent the selection of a new coordinate

system obtained by rotating the original variables and

projecting them onto the reduced space defined by the first

few principal components. Here, the first principal

component describes the largest amount of variation in the

data, the second one the second largest amount of variation

in the data and so on. Each principal component is

represented as a linear combination of the columns (J), and

has a specific numerical value for each of the rows (I). In

matrix form, the PCA model can be written as,

1

, L

T

s f ff

X i j t p E

(5)

where L is the number of principal components. The loading

vectors (pf) are orthonormal and provide the directions with

maximum variability. The score vectors (tf) from the

different principal components are the coordinates for the

objects in the reduced space. A classification of a new test

pattern is done by obtaining its predicted scores and

residuals. If the test pattern is similar to a specific class in

the training network, the scores will be located near the

origin of the reduced space, and the residual should be

small. The distance of test data from the origin of the

reduced space can be measured by Hotelling statistic [7].

Partial Least Squares—PLS is a regression technique used

to find a set of components that explain the covariance

between the independent training data matrix (sensor data)

X, and the dependent matrix (fault classes) Y. The goal of

the PLS algorithm is to determine highly correlated latent

variables (or the score vectors) that not only capture the

variations in the input data X, but also the variations that are

most predictive of the output data Y. Once the latent

variables are extracted, a least squares regression is

performed to estimate the fault class. The score vectors are

determined using nonlinear iterative partial least squares

(NIPALS) algorithm [21].

Probabilistic Neural Network—PNN is a supervised method

that computes the likelihood of an input vector belonging to

a specific class based on the learned probability distributions

of each class. The learned patterns can also be weighted

with a priori probability (relative frequency) of each

category and misclassification costs to determine the most

likely class for a given input vector. If the relative frequency

of the categories is unknown, then all the categories can be

assumed to be equally likely and the determination of

category is solely based on the closeness of the input feature

vector to the distribution function of a class [8].

7

Fig. 6: Driver Profile (Profile 1) Fig. 7: Driver Profile (Profile 2)

Fig. 8: Battery SOC with and without Noise (Profile 1)

Fig. 9: Cumulative Sum of Squared Residuals from 25 Monitored Signals for Battery Current Sensor Fault (F1)

8

Fig. 10: Cumulative Sum of Squared Residuals from Battery SOC Signal (S1) for all the Faults

Table IV. Classification Accuracy of each Classifier for each Fault

Classifier

Fault

SVM KNN

(k=1)

KNN

(k=2)

KNN

(k=3)

PLS PCA PNN

F1 100 100 100 100 100 100 100

F2 100 100 100 100 100 100 100

F3 100 100 100 100 83.33 100 100

F4 100 100 100 100 100 100 100

F5 100 100 100 100 100 100 100

F6 100 100 100 100 100 100 83.33

F7 100 100 100 100 100 100 83.33

F8 100 100 100 100 100 100 100

F9 100 100 100 100 100 100 100

F10 100 100 100 100 100 100 100

F11 100 100 100 100 100 100 100

F12 100 100 100 100 50 100 100

5. EXPERIMENTAL RESULTS

The RBS model was simulated under nominal and 12

faulty scenarios with two driver profiles: Profile1 and

Profile2 shown in Figures 6 and 7. These driver profiles

(drive cycles) are collected from a Chevy 2007 model year

vehicle and acts as input to the driver block of the RBS

model. The duration of the driver profile is 1400 seconds

and the data is sampled every 0.1 seconds. We monitored 25

signals for each fault case; hence, the total data collected

from the fault simulations is 13x25x14001 for each profile.

A Gaussian measurement noise with random seed is added

to the signals with a variance of 0.6% of the squares of the

magnitude of the signals (this corresponds to a signal-to-

noise (SNR) ratio of 22.2dB). Fig 8 is a snapshot of a

sample signal with and without noise added to the signal.

Fault Detection

Here, CUSUM test is employed for fault detection. For

CUSUM, the cumulative sum of the squared residuals for

the 25 monitored variables is collected. A threshold is

defined and the cumulative sum of the squared residual is

compared against this threshold to determine whether the

system is in faulty state or in the nominal state. Fig. 9 shows

the cumulative sum of the squared residuals for battery

current sensor fault. Here, x-axis represents the sample

number and y-axis represents the cumulative sum of squared

residuals from 25 monitored signals. A suitable threshold on

the residuals detects the presence of fault(s), for instance, a

threshold of 500 on signal S22 indicates the presence of a

fault at t = 89.4 seconds. Similarly, a threshold of 5 on S1

detects F11, the battery SOC fault (see Fig 10). Once a fault

is detected, data reduction techniques and classification

techniques can be employed to isolate the fault.

Data Reduction and Fault Classification Results

MPLS technique is employed to reduce the data into

fewer components. Here, the data is reduced to 10 score

vectors (559 MB → 12 KB) and this reduced data is used

for fault isolation. As mentioned earlier, the classifiers

SVM, KNN, PLS, PCA and PNN are employed to isolate

9

the sensor and parametric faults. Here, the number of

patterns available is 156 (26 patterns from Profiles 1 and 2

and 130 patterns by adding noise to the measured signals).

This data is divided into two equal subsets, i.e., 50% for the

training of classification algorithms and the remaining 50%

of the data is used for validating the trained classification

algorithms. Table IV presents the individual classification

accuracies of each classifier for each fault. It is seen that

SVM, KNN and PCA performed best by achieving 100%

accuracy on each fault. Table V summarizes the overall

classification accuracy achieved by each of the individual

classifiers. Here, all the classifiers performed well with

greater than 97% accuracy, except for PLS. The sample of

faults and signals considered in this paper are to test the

viability of the presented approach. It is necessary to explore

more realistic scenarios, such as different driving conditions,

road scenarios, a variety of fault types, etc. and validate the

proposed approach. However, the preliminary experimental

results are very promising.

The systematic FDD process presented in this paper

featuring fault detection, multivariate statistical data

reduction and fault classification techniques is a generic

approach for fault analysis and can be easily applied to any

real-world engineering system. Indeed, variants of this

process have been applied to a number of engineering

systems [17][18] [22]-[27].

Table V. Overall Classification Accuracy

Classification

Algorithm

Classification

Accuracy (%)

SVM 100

KNN (k=1) 100

KNN (k=2) 100

KNN (k=3) 100

PLS 94.87

PCA 100

PNN 97.44

6. CONCLUSIONS

In this paper, a systematic data-driven fault detection and

diagnosis approach for an automotive regenerative braking

system is presented. The RBS system is modeled using

PSAT software and simulated under nominal and faulty

scenarios. We employed MPLS technique for data

reduction, and statistical and pattern recognition techniques

for fault isolation. The approach presented here

demonstrated high diagnostic accuracy (100% fault

classification accuracy with SVM, KNN and PCA

classifiers) and can be used for fault analysis in any vehicle

systems. It has the potential for real-time implementation in

automotive and aerospace systems due to the significant

reduction of the data size without compromising on the fault

classification accuracy at a reduced computational load/

time. In future, we plan to investigate the robustness of the

FDD scheme with varying driver profiles. In addition, we

plan to model the RBS with multiple ECUs and CAN bus

communication, and investigate the application of our FDD

scheme for isolating software and communication-related

faults. Finally, we will also predict the degradation of RBS

parameters and estimate its remaining useful life.

ACKNOWLEDGMENTS

The work reported in this paper was supported by

National science foundation (NSF) under grants ECCS-

0931956 (NSF CPS) and ECCS-1001445 (NSF GOALI).

We thank GM R&D and NSF for their support of this work.

Any opinions expressed in this paper are solely those of the

authors and do not represent those of the sponsor.

REFERENCES

[1] J. K. Ahn, K. H. Jung, D. H. Kim, H. B. Jin, H. S. Kim

and S. H. Hwang, ―Analysis of a regenerative braking

system for hybrid electric vehicles using an electro-

mechanical brake‖, International Journal of Automotive

Technology, vol. 10, no. 2, pp. 229-234, 2009.

[2] http://www.cvel.clemson.edu/auto/systems/regenerative

_braking.html, accessed on October 10th

, 2011.

[3] K. R. Pattipati, A. Kodali, K. Choi, S. Singh, C.

Sankavaram, S. Mandal W. Donat, S.M. Namburu, S.

Chigusa, L. Qiao and J. Luo, "An integrated diagnostic

process for automotive systems," in D. Prokhorov, (ed.)

Studies in Computational Intelligence (SCI), Vol. 132,

pp. 191-218, 2008.

[4] C. Sankavaram, A. Kodali, D. F. M. Ayala, K. Pattipati,

S. Singh, and P. Bandyopadhyay, ― Event-driven data

mining techniques for automotive fault diagnosis‖, 21st

International Workshop on Principles of Diagnosis,

Portland, OR, October 2010.

[5] C. Sankavaram, B. Pattipati, A. Kodali, K. Pattipati, M.

Azam, and S. Kumar, "Model-based and data-driven

prognosis of automotive and electronic systems", 5th

Annual IEEE Conference on Automation Science and

Engineering, Bangalore, India, August 22-25, 2009.

[6] http://www.transportation.anl.gov/modeling_simulation/

PSAT/index.html, accessed on October 10th

, 2011.

[7] P. Nomikos, ―Detection and diagnosis of abnormal

batch operations based on multi-way principal

component analysis,‖ ISA Transactions, Vol. 35, pp.

259–266, 1996.

[8] R. 0. Duda, P. E. Hart, and D. G. Stork, Pattern

Classification, second edition, New York: Wiley

lnterscience, 2001.

[9] C. M. Bishop, Pattern Recognition and Machine

Learning, Springer, 2006.

10

[10] M. Ehsani, Y. Gao, and J. M. Miller, ―Hybrid electric

vehicles: architecture and motor drives‖, Proceedings of

the IEEE , vol.95, no.4, pp.719-728, April 2007.

[11] http://www.transportation.anl.gov/pdfs/HV/412.pdf,

accessed on October 10th

, 2011.

[12] C. Sankavaram, K. Pattipati, Y. Zhang, M. Howell, and

M. Salman, ―Fault Diagnosis and Prognosis in Cyber-

Physical Systems with an Application to Hybrid Electric

Vehicle Regenerative Braking System‖, to be submitted

to IEEE transactions on SMC – Part C.

[13] http://www.epa.gov/nvfel/testing/dynamometer.htm,

accessed on October 11th

, 2011.

[14] T. Kurtoglu, S. Narasimhan, S. Poll, D. Garcia, L.

Kuhn, J. D. Kleer, A. Gemund, and A. Feldman,

―Towards a Framework for evaluating and comparing

diagnosis algorithms‖, in Proceedings of the Twentieth

International Workshop on Principles of Diagnosis

(DX’09), Stockholm Sweden, pp. 373-382, 2009.

[15] I. K. Fodor, "A survey of dimension reduction

techniques," [Online]. Available: http://www.llnl.

gov/CASC/sapphire/pubs/1 48494. pdf.

[16] R. E. Bellman, Adaptive Control Processes, Princeton

University Press, Princeton, NJ, 1961.

[17] K. Choi, J. Luo, K. Pattipati, S.M. Namburu, L. Qiao,

and S. Chigusa, ―Data Reduction Techniques for

Intelligent Fault Diagnosis in Automotive Systems‖,

Proceedings of the IEEE Autotestcon, Anaheim, CA,

September, 2006.

[18] K. Choi, S. M. Namburu, M. S. Azam, Jianhui Luo, K.

R. Pattipati and A. Patterson-Hine, "Fault diagnosis in

HVAC chillers," Instrumentation & Measurement

Magazine, IEEE, vol. 8, pp. 24-32, 2005.

[19] J. Chen and R. Patton, Robust Model-based Fault

Diagnosis for Dynamic Systems, Kulwer Academic,

Norwell, MA, 1999.

[20] C. J. C. Burges, ―A tutorial on support vector machines

for pattern recognition,‖ Data Mining and Knowledge

Discovery, vol. 2, pp. 121-167, 1998.

[21] P. Geladi, and B. R. Kowalski, ―Partial least-squares

regression: a tutorial‖, Analytica, Chemica, Acta, 1986.

[22] S. M. Namburu, M. Azam, J. Luo, K. Choi, and K. R.

Pattipati, ―Fault detection, diagnosis and data-

driven modeling in HVAC Chillers,‖ in Proc. SPIE

Conf., Orlando, FL, March 2005.

[23] W. Donat, K. Choi, W. An, S. Singh, and K. R.

Pattipati, ―Data Visualization, Data Reduction, and

Classifier Fusion for Intelligent Fault Detection and

Diagnosis in Gas Turbine Engines‖, ASME Journal of

Engineering for Gas Turbines and Power, June 2007.

[24] K. Choi, S. Singh, K. R. Pattipati, J. W. Sheppard, S.

M. Namburu, S. Chigusa, D. V. Prokhorov, and L.

Qiao, ―Novel classifier fusion approaches for fault

diagnosis in automotive systems,‖ IEEE Transactions on Instrumentation and Measurements, Vol. 57, No. 3,

pp. 602-611, March 2009.

[25] A. Kodali, W. Donat, S. Singh, K. Choi, and K. R.

Pattipati, ―Dynamic fusion and parameter optimization

of multiple classifier systems,‖ ASME Turbo Expo

2008: Power for Land, Sea and Air, Berlin, Germany,

June 2008.

[26] A. Kodali, S. Vemana, K. Choi, K. R. Pattipati,

―Diagnostic ambiguity and parameter optimization in

classifier fusion: Application to gas turbine engine

data,‖ IEEE Autotestcon Conference, Salt Lake City,

Utah, pp. 433-438, September 2008.

[27] R. Ghimire, C. Sankavaram, A. Ghahari, K. Pattipati, Y.

Ghoneim, M. Howell, and M. Salman, ―Integrated

Model-based and Data-driven Fault Detection and

Diagnosis Approach for an Automotive Electric Power

Steering System‖, IEEE Autotestcon Conference,

Baltimore, MD, September 2011.

BIOGRAPHIES

Chaitanya Sankavaram received the B.Tech. degree in

Electrical and Electronics Engineering from Sri

Venkateswara University, Tirupathi, India, in 2005. She

is currently working toward the Ph.D. degree in

Electrical and Computer Engineering at the University of

Connecticut, Storrs. She was a Project Engineer with

Wipro Technologies, Bangalore, India, for two years. Her

current research interests include fault diagnosis and

prognosis, reliability analysis, data mining, pattern

recognition, and optimization theory.

Bharath Pattipati received the B.E. degree in Electrical

and Electronics engineering from M.S. Ramaiah Institute

of Technology, Bangalore, India, in 2005 and the

Master’s degree in Electrical and Computer Engineering

from the University of Connecticut, Storrs, in 2009, where

he is currently working toward the Ph.D. degree in

Electrical and Computer Engineering. His current

research interests include the application of systems

theory and optimization techniques to complex large-

scale systems, application-driven analysis of neural

networks, pattern recognition, and fault diagnosis and

prognosis.

Krishna R. Pattipati (S’77-M’80-SM’91-F’95) received

the B.Tech. degree in Electrical Engineering with highest

honors from the Indian Institute of Technology,

Kharagpur, India, in 1975 and the M.S. and Ph.D.

degrees in Systems Engineering from the University of

Connecticut, Storrs, in 1977 and 1980, respectively.

From 1980 to 1986, he was with ALPHATECH, Inc.,

Burlington, MA. Since 1986, he has been with the

University of Connecticut, where he is currently the UTC

Professor in Systems Engineering in the Electrical and

Computer Engineering Department. He was a Consultant

to Alphatech, Inc., Aptima, Inc., and IBM Research and

Development. He is a Co-Founder of Qualtech Systems,

Inc., which is a small business specializing in intelligent

diagnostic software tools. His research interests include

the areas of adaptive organizations for dynamic and

11

uncertain environments, multiuser detection in wireless

communications, signal processing and diagnosis

techniques for complex system monitoring, and multi

object tracking. He is a Fellow of IEEE.

Yilu Zhang received his B.S., and M.S. degrees in

Electrical Engineering from Zhejiang University, China,

in 1994, and 1997, respectively; and his Ph.D. degree in

Computer Science from Michigan State University in

2002. He joined the R&D center of General Motors Co.

at Warren, Michigan in 2002, and currently holds a

position of Senior Researcher. Dr. Zhang’s research

interests include integrated system health management,

human machine interactions, statistical pattern

recognition, machine learning, speech/image processing,

and their applications. Dr. Zhang is a Senior Member of

IEEE, and a member of SAE. He is a winner of 2008

“Boss” Kettering Award – the highest technology award

in GM – for his contribution to Connected Vehicle

Battery Monitor, a remote vehicle diagnostics technology.

Mark Howell is a Staff Researcher with the Electrical

and Controls Lab at General Motors Research,

specializing in Fault Tolerant Vehicle Control and

Prognostic Systems. He previously worked as a Research

Fellow in Automotive Engineering at Loughborough

University, pursuing research on intelligent active

suspension control and integrated chassis control

systems. He has a MSc and PhD in Control Systems from

the University of Sheffield, England, and a Bachelors

degree in Cybernetics from the University of Reading,

England.

Mutasim Salman received his bachelor’s degree in

Electrical Engineering from University of Texas at

Austin; M.S. and PhD in Electrical Engineering with

specialization in Systems and control from University of

Illinois at Urbana- Champaign. He also has an Executive

MBA. He holds sixteen patents and has coauthored more

than 36 refereed technical publications and a book. He

joined the GM R&D staff in 1984. He is currently a group

manager and a Technical Fellow. He is a senior member

of IEEE. His research interest includes performance

monitoring, diagnosis and prognosis as well as fault

tolerant control as applied to Energy and Automotive

Systems.