CUSTOMER demand for improved fuel economy is challenging

Transcription

1 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER Intelligent Vehicle Power Control Based on Machine Learning of Optimal Control Parameters and Prediction of Road Type and Traffic Congestion Jungme Park, Zhihang Chen, Leonidas Kiliaris, Ming L. Kuang, M. Abul Masrur, Senior Member, IEEE, Anthony M. Phillips, and Yi Lu Murphey, Senior Member, IEEE Abstract Previous research has shown that current driving conditions and driving style have a strong influence over a vehicle s fuel consumption and emissions. This paper presents a methodology for inferring road type and traffic congestion (RT&TC) levels from available onboard vehicle data and then using this information for improved vehicle power management. A machine-learning algorithm has been developed to learn the critical knowledge about fuel efficiency on 11 facility-specific drive cycles representing different road types and traffic congestion levels, as well as a neural learning algorithm for the training of a neural network to predict the RT&TC level. An online University of Michigan-Dearborn intelligent power controller (UMD_IPC) applies this knowledge to real-time vehicle power control to achieve improved fuel efficiency. UMD_IPC has been fully implemented in a conventional (nonhybrid) vehicle model in the powertrain systems analysis toolkit (PSAT) environment. Simulations conducted on the standard drive cycles provided by the PSAT show that the performance of the UMD_IPC algorithm is very close to the offline controller that is generated using a dynamic programming optimization approach. Furthermore, UMD_IPC gives improved fuel consumption in a conventional vehicle, alternating neither the vehicle structure nor its components. Index Terms Fuel economy, machine learning, road type and traffic congestion (RT&TC) level prediction, vehicle power management. I. INTRODUCTION CUSTOMER demand for improved fuel economy is challenging the automotive industry to produce affordable new vehicles that deliver better fuel efficiency without sacrificing performance, safety, emissions, or reliability. To meet this challenge, it is very important to optimize the architecture and the various devices and components of the vehicle system, as well as the energy-management strategy that is used to Manuscript received October 1, 2008; revised March 12, 2009 and May 14, First published July 17, 2009; current version published November 11, This work was supported in part by the State of Michigan through the 21st Jobs Fund under a grant and in part by the Institute of Advanced Vehicle Systems, University of Michigan-Dearborn, under Grant 06-1-p The review of this paper was coordinated by Dr. M. S. Ahmed. J. Park, Z. Chen, L. Kiliaris, and Y. L. Murphey are with the Department of Electrical and Computer Engineering, University of Michigan-Dearborn, Dearborn, MI USA ( yilu@umich.edu). M. L. Kuang and A. M. Phillips are with the Ford Motor Company, Dearborn, MI USA. M. A. Masrur is with the U.S. Army RDECOM-TARDEC, Warren, MI USA. Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /TVT efficiently control the energy flow through the vehicle system. Our research focuses on the latter. Vehicle power management has been an active research area in the past decade and has intensified recently by the emergence of hybrid electric vehicle technologies. Most of the previous approaches were developed based on mathematical models or knowledge derived from static vehicle operation data. The application of optimal control theory to power distribution and management has been the most popular approach, which includes linear programming [1], optimal control [2], and, particularly, dynamic programming (DP) [3] [5]. In general, these techniques do not offer an online solution because they assume that the future drive cycle is entirely known. However, these results can be used as a benchmark for the performance of online power control strategies. If only the present state of the vehicle is considered, optimization of the operating points of the individual components can still be beneficial, but the benefits will be limited [6] [8]. Interesting techniques for deriving effective online control rules based on the results generated by offline DP and quadratic programming (QP) can be found in [3] and [9]. Recent research has shown that current driving conditions and the driver s driving style have a strong influence over a vehicle s fuel consumption and emissions [10], [11]. Driving patterns exhibited by a real-world driver are the product of the instantaneous decisions of the driver to respond to the (physical) driving environment. Specifically, varying road type and traffic conditions, driving trends, driving styles, and vehicle operating modes have had varying degrees of impact on vehicle fuel consumption. However, most of the existing vehicle power control approaches do not incorporate knowledge about driving patterns into their vehicle power-management strategies. The main contribution of this paper is an algorithm for optimization of vehicle power management that utilizes inferred knowledge of road type and traffic congestion (RT&TC). Only recently has the research community in vehicle power control begun to explore ways to incorporate knowledge about driving patterns into online control strategies [12] [15]. A comprehensive overview of intelligent system approaches for vehicle power management can be found in [16]. This paper presents our research on intelligent vehicle power management using machine learning. Specifically, we will present machine-learning algorithms for learning about the optimal power control parameters for all 11 standard facilityspecific (FS) drive cycles proposed in [17] and [18] and /$ IEEE

2 4742 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 Fig. 1. Intelligent power control in a vehicle system. about predicting road types and traffic congestion, as well as an online University of Michigan-Dearborn intelligent power controller (UMD_IPC) that applies the knowledge obtained through machine learning to online vehicle power control with the online prediction of driving environment by a neural network. UMD_IPC has been fully implemented in a conventional vehicle model built using the powertrain systems analysis toolkit (PSAT) ( index.html) simulation program and tested on 11 drive cycles provided by the PSAT library. PSAT is a high-fidelity simulation software developed by Argonne National Laboratory, Argonne, IL, under the direction of and with contributions from Ford, General Motors, and Chrysler. PSAT is a forwardlooking model that simulates vehicle fuel economy and performance in a realistic manner taking into account transient behavior and control system characteristics. It can simulate a broad range of predefined vehicle configurations (conventional, electric, fuel cell, series hybrid, parallel hybrid, and power split hybrid). In this research project, the PSAT software is used to build a high-fidelity vehicle model; simulate drive cycles to generate numerical data, such as fuel consumption and emissions and vehicle performance; and implement an intelligent power controller UMD_IPC. Experiments will show that the online performances of UMD_IPC are very close to the offline optimal controller built based on DP. In comparison with the default controller used by the vehicle model in PSAT, our results showed a maximum of 3.95% fuel reduction in an urban drive cycle. Furthermore, the implementation of UMD_IPC does not require the change of any vehicle components. Although the research results presented in this paper were generated based on a conventional vehicle model, the proposed technology can be extended to a hybrid vehicle system, which is the authors ongoing effort. This paper is organized as follows. Section II presents the machine-learning process of optimal power control in a conventional vehicle, Section III presents a neural network system for predicting roadway type and traffic-congestion level, Section IV presents the intelligent online vehicle powermanagement system, namely, UMD_IPC, Section V presents the experiment results, and Section VI presents the conclusion. II. OPTIMAL POWER CONTROL IN A CONVENTIONAL VEHICLE SYSTEM USING MACHINE LEARNING Fig. 1 illustrates the interaction between the proposed intelligent power controller UMD_IPC and the major power components in a conventional vehicle system. At any given time during a drive cycle, based on the current vehicle state, which is represented by the current vehicle speed, driver power demand, electrical load and state of charge (SOC) of the battery, the UMD_IPC calls the neural network NN_RT&TC to predict the current RT&TC level and calculates the electric power set point to the battery controller and a resultant feedforward torque compensation to the engine controller. The variable P s, representing the power actually to be charged (P s > 0) or discharged (P s < 0) from the battery, is set by the UMD_IPC with the aim of minimizing fuel consumption. The desired engine power P eng, which is calculated based on the optimal value of P s,isusedto find the feedforward torque compensation through the engine fuel-efficiency map. The functional relationship between P eng and P s is shown as follows: where ω P d P e G e2m (P e,ω) P eng = P d + G e2m (P e,ω) (1) P e = P l + P b (2) P b = η in2out (P s, SOC,T) (3) engine speed; driver demanded power at the wheels; electrical power from the alternator; mechanical power required by the alternator based on alternator efficiency map Φ alt to produce a given electrical power P e at a given speed;

3 PARK et al.: VEHICLE POWER CONTROL BASED ON MACHINE LEARNING OF OPTIMAL CONTROL PARAMETERS 4743 where γ(p s,t) is the fuel consumed as a function of P s (t) at time t. The fuel-consumption function γ(p s,t) is approximated as a convex quadratic function, i.e., γ(p s,t) ϕ 2 (t)p s (t) 2 +ϕ 1 (t)p s (t)+ϕ 0 (t), ϕ 2 >0 (5) Fig. 2. Battery efficiency map Φ bat. P l electrical power required by the various vehicle electrical loads; P s actual power stored in and drawn out of the battery; SOC battery state of charge; P b power output at the battery controller, which is a function of the internal battery power P s, SOC, and battery temperature T and is denoted as η in2out (P s, SOC,T). η in2out (P s, SOC,T) reflects power losses in the battery. In this paper, η in2out is derived by modeling the battery-efficiency map Φ bat shown in Fig. 2. The battery-efficiency map contains the battery charge/discharge curves generated by the battery model in PSAT for SOC = 40%, 50%, and 60%. It appears that within the battery-efficiency range of 40% 60%, the battery charge and discharge curves have very little variation. The machine-learning algorithm, namely, learning minimum power consumption on FS drive cycles (LMPC_FSDC), is presented in Section II-B. LMPC_FSDC attempts to learn the values of parameters that minimize the vehicle fuel consumption function γ, which is empirically modeled as a quadratic function of P s. Using this quadratic function, a fuel consumption cost index is defined and solved using a QP approach to produce optimal values of P s. The resultant optimal solution is dependent on the drive cycles. The machine-learning algorithm LMPC_FSDC learns the optimal values of the empirical parameters generated by the QP for a defined set of roadway types and congestion levels. The QP approach is based on the research presented in [9], which is briefly summarized in Section II-A. A. Vehicle Power-Optimization Model The power-optimization problem is modeled as a multistep decision problem in a drive cycle with N steps that minimizes a performance index J, i.e., min P s J = min P s γ (P s (t),t)= min P s γ (P s (t),t) (4) where ϕ i represents time-varying coefficients. The objective function then becomes min P s J = min γ (P s (t),t) P s min P s ( ϕ2 (t)p 2 s (t)+ϕ 1 (t)p s (t)+ϕ 0 (t) ) (6) where P s contains the optimal values of P s (t) for t = 1,...,N. To create a well-posed problem, the constraint that the energy in the battery at the end of the drive cycle E s (N) must match the energy at the beginning of the cycle E s (0), i.e., E s (N) =E s (0) is applied to the optimization. This constraint can be written as E s (N) E s (0) = P s (t) =0. (7) By adjoining this constraint to our objective function using a Lagrange multiplier, we obtain the following Lagrange function: { L (P s (1),...,P s (N),λ)= ϕ2 (t)p s (t) 2 + ϕ 1 (t)p s (t) } + ϕ 0 (t) λ P s (t). (8) By taking the partial derivatives of the Lagrange function L with respect to P s (t), t =1,...,N, and λ, respectively, and by setting the equations to 0, combined with (7), the optimal battery power setting at each time step t can be obtained as follows: P s (t) = λ ϕ 1(t) (9) 2ϕ 2 (t) where λ = ϕ 1 (t) 2ϕ 2 (t) 1 2ϕ 2 (t) The above formula for calculating λ requires the knowledge of ϕ 1 (t) and ϕ 2 (t) over the entire drive cycle, which is not available in advance to the online controller during normal realworld driving. To solve this problem, we adopt the method proposed in [9] that uses a proportional integral controller to produce a value of λ online based on the measured energy level in the battery E s, i.e., λ(t) =λ 0 + K P (E s (0) E s (t 1)) t 1 + K I (E s (0) E s (p)). (10) p=1.

4 4744 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 TABLE I STATISTICS OF THE 11 FS DRIVE CYCLES The optimal parameters ϕ 0 (t), ϕ 1 (t), ϕ 2 (t), λ 0, and K P and K I are obtained by the machine-learning algorithm described in Section II-B with the constraints of the upper and lower bounds of P s (t), which are also discussed in Section II-B. B. Machine Learning About Optimal Power Settings We model the road environment of a driving trip as a sequence of different roadway types, such as local, freeway, and arterial/collector, augmented with different traffic congestion levels. Sierra Research Inc. has shown that fuel efficiency and emissions are connected to roadway types as well as traffic congestion levels. They developed a set of 11 standard drive cycles presented in [17] and [18], called facility-specific (FS) cycles, to represent passenger car and light truck operations over a broad range of facilities and congestion levels in urban areas. Table I shows the most recent definition of these road types [18], along with the labels that we assigned, where V avg is the average vehicle speed in meters per second, V max is the maximum vehicle speed in meters per second, and A max is the maximum acceleration. The 11 drive cycles are divided into the following four categories of roadway types: 1) freeway; 2) freeway ramp; 3) arterial; and 4) local. The two categories, freeway and arterial, are further divided into subcategories based on a qualitative measure called level of service (LOS) that describes operational conditions within a traffic stream based on speed and travel time, freedom to maneuver, traffic interruptions, comfort, and convenience. Six types of LOS are defined with labels, i.e., A through F, with LOS A representing the best operating conditions and LOS F the worst. Each LOS represents a range of operating conditions and the driver s perception of those conditions; however, safety is not included in the measures that establish service levels [18], [19]. In this paper, we use this set of 11 FS cycles as the standard measure of roadway types and traffic-congestion levels. For the convenience of description, we label these 11 FS cycles as RT 1,...,RT 11. The problem of optimal vehicle power management is formulated as follows. Assume that for any given drive cycle DC(t) (t [0,t e ], where t e is the ending time of the drive cycle) at any given time t, the vehicle is operating according to one of the 11 road types and traffic-congestion levels, i.e., RT i, i =1,...,11. Then, the online optimal power controller UMD_IPC chooses the optimal battery power settings based on RT i. The machinelearning algorithm LMPC_FSDC has been developed to learn the optimal power settings for all 11 FS drive cycles, i.e., RT i, i =1,...,11. Fig. 3 shows the major computational steps in LMPC_FSDC. The algorithm requires the use of a high-fidelity vehicle system modeling and simulation program F, such as PSAT or ADVISOR. Two major steps in the algorithm require the use of such a simulation program. First, we need the simulation program to build a vehicle model V of a particular interest. Second, we run the vehicle model V in the simulation program F to generate step-by-step system state data: P d (t), P l (t), and ω(t), t =1,...,N, every standard FS drive cycle, RT i, i = 1,...,11. The fuel matrix F_R i is generated for all the P s values within the specific upper and lower bounds of P s, denoted as P s_ min(t) P s (t) P s_ max(t), which are calculated as follows. Let P eng_ max(ω(t)) be the maximum engine power with engine speed ω(t) and P alt_ max(ω(t)) be the maximum mechanical alternator power with the given speed ω(t). Both P eng_ max(ω(t)) and P alt_ max(ω(t)) are defined by the vehicle model V. At each time t, the maximum electrical power at the alternator P e_ max(ω(t)) can be calculated by P e_ max (ω(t)) = G m2e {min [P eng_ max (ω(t)) P d (t),p alt_ max (ω(t))],ω(t)} (11) where G m2e (P alt,ω) is a function that calculates the electrical power based on the alternator efficiency map Φ alt for a given mechanical power P alt and rotational speed ω. The min[p eng_ max(ω(t)) P d (t),p alt_ max(ω(t))] in (11) represents the maximum mechanical power at time t. Based on the engine and alternator constraint, the upper and lower bounds of P s are calculated by P s_ max 1(t) =η out2in (P e_ max (ω(t)) P l (t)) P s_ min 1(t) =η out2in (0 P l (t)) (12) where η out2in is a function that calculates the internal battery power P s, namely the power to be stored or drawn from the battery for a given battery power P b at the battery terminal, by using the battery efficiency map Φ bat shown in Fig. 2. (P e_ max(ω(t)) P l (t)) and (0 P l (t)) represent, respectively, the maximum and minimum battery power P b at the battery terminal at time t. Since the boundary of P s (t) is also constrained (or restricted) by current SOC, i.e., SOC(t), as shown in Fig. 14, the upper and lower bounds of P s are P s_ max(t) = min (P s_ max 1(t),P s_ max 2(t)) P s_ min(t) = max (P s_ min 1(t),P s_ min 2(t)) (13) where P s_ max 2(t) and P s_ min 2(t)) can be calculated based on SOC(t). A fuel-rate matrix F_R(P s (t),t ω(t),p d (t),p l (t)) is generated for each time step t as a function of P s (t), which is

5 PARK et al.: VEHICLE POWER CONTROL BASED ON MACHINE LEARNING OF OPTIMAL CONTROL PARAMETERS 4745 Fig. 3. Computational steps of machine-learning algorithm LMFC_FSDC. the charge and discharge power within the system constraints specified in (13) at time t for the given engine speed ω(t), required drivetrain power P d (t), and electric load power P l (t). The tth column of matrix F_R(,t) is represented as a convex quadratic cost function of P s. By using a regression function, we can obtain the coefficients ϕ 2 (t), ϕ 1 (t), and ϕ 0 (t) such that ϕ 2 (t)ps 2 (t)+ϕ 1 (t)p s (t)+ϕ 0 F_R(,t) with the best fit. Fig. 4 shows a few example of the actual fuel rates and the convex quadratic cost functions calculated at various time steps associated with the vehicle model Ford Taurus in the Arterial AB drive cycle, i.e., RT 8. Note the fuel rate has been multiplied with the chemical-energy contents of fuel, i.e., H f =44 kj/g [20], to obtain a suitable scaling, and P s values have been normalized as follows: {P s (t) mean(p s ( ))}/σ(p s ( )), where mean(p s ( )) is the mean of P s, and σ(p s ( )) is the standard deviation of P s. This illustrates that a quadratic function is a good choice to represent the fuel function F_R(P s (t),t ω(t)p d (t),p l (t)). Fig.5shows the coefficients {ϕ 2 (t),ϕ 1 (t),ϕ 0 (t) t = 511,...,533} of the quadratic cost function of P s (t) for the same drive cycle used in Fig. 4. These coefficients obtained at all time steps for each drive cycle RT i, i =1,...,11 are then used to calculate the power control parameters ϕ i 1, ϕ i 2, λ i, KP i, and Ki I as follows: ϕ i 1 = 1 N λ i = ϕ i 1(t) ϕ i 1 (t) 2ϕ i 2 (t) ϕ i 2 = 1 N ϕ i 2(t) (14) (15) 1 2ϕ i 2 (t) K i P = ϕ i K i I = ϕ i (16) The machine-learning algorithm to the 11 standard FS drive cycles and the results are shown in Table II, which serves as the knowledge base for the online controller UMD_IPC. III. PREDICTING ROADWAY TYPE AND TRAFFIC CONGESTION LEVEL The problem of roadway type prediction is formulated as follows. Let SP(t) be the speed profile of a driver on the road t =0, 1,...,t c, where t c is the current time instance, and let R(t) be the roadway types that the driver needs to go through to complete his trip 0 <t<t e, where t e is the time when the

6 4746 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 TABLE II OPTIMAL PARAMETER SETTINGS GENERATED BY THE MACHINE-LEARNING ALGORITHM LMFC_FSDC, FOR 11 STANDARD FS DRIVE CYCLES Fig. 4. Actual and calculated fuel rate for the arterial AB drive cycle RT 8. Fig. 5. Coefficients of the quadratic fuel-rate function calculated from the arterial AB drive cycle RT 8. trip ended. At any given time t c,r(t c ) {RT i i =1,...,11}. The roadway type in the near future is to be predicted based on the short-term memory of the driving environment during the trip. Specifically, we attempt to develop a nonlinear function F such that F (SP(t) t [(t c ΔZ),t c ]) = RT j, 0 <j 11, where ΔZ >0 is called the window size that characterizes the length of the speed profile that should be used to explore driving patterns. The variable RT j is the roadway type that the driver will be on during the time interval [t c, (t c +Δt)], i.e., R(t) =RT j for t [t c, (t c +Δt)]. We refer to Δt >1 as the time step. To solve this problem, four different aspects of the roadway type predictor need to be determined as follows. 1) Select effective features that can be extracted from SP(t), t c ΔZ <t t c, for the prediction of the current roadway type. 2) Determine the optimal window size ΔZ. 3) Determine the optimal time step Δt. 4) Develop a function F that has the capability of accurately predicting roadway types in a sufficiently short time suitable for online driving prediction. In this paper, F is a neural network described in Section III-C. The following three key components developed for predicting road types and traffic congestion levels are described in Sections III-A C, respectively: 1) feature selection; 2) prediction windows and time step; 3) a neural network. A. Feature Selection Roadway types and traffic-congestion levels can generally be observed in the speed profile of the vehicle. The statistics used to characterize driving patterns include 16 groups of parameters (62 total) suggested by the Sierra Research, and parameters in nine out of these 16 groups critically affect fuel usage and emissions [10], [11]. However, it may not be necessary to use all these features to predict a specific drive pattern, and

7 PARK et al.: VEHICLE POWER CONTROL BASED ON MACHINE LEARNING OF OPTIMAL CONTROL PARAMETERS 4747 additional new features may be explored as well. For example, Langari and Won [14] used only 40 of the 62 parameters and then added the following seven new parameters: 1) trip time; 2) trip distance; 3) maximum speed; 4) maximum acceleration; 5) maximum deceleration; 6) number of stops; and 7) idle time (percent of time at a speed of 0 km/h). However, the use of additional parameters needs to be balanced with the curse of dimensionality : Too many features may degrade system performance. Furthermore, in onboard vehicle implementation, more features imply higher hardware cost and/or more computational time. The problem of selecting a subset of optimal features is a classic research topic in pattern recognition and is a NP problem. Because the feature selection problem is computationally expensive, this research has focused on finding a quasi-optimal subset of features, where the term quasi-optimal implies good, but not necessarily always, optimal classification performance. Interesting feature-selection techniques can be found in [21] [23]. However, most of these feature-selection algorithms were developed for two-class classification problems, and extensions to K-class (K >2) will significantly increase the computational time. With this background in mind, we developed the following feature-selection algorithm based on roadway types. Feature-Selection Algorithm: Step 1) Let X be the training data set and Ω be the initial set of n features, which can be obtained from those suggested by the research community, as discussed earlier. Step 2) Relabel data in X with freeway samples as 1 and all others as 0. Denote this training data set as X 1. Select the best features from Ω that can classify all the freeway data against all other data in X 1. Denote this feature set as Ω 1. Step 3) Relabel data in X with freeway ramp samples as 1 and all others as 0. Denote this training data set as X 2. Select the best features from Ω that are not in Ω 1 and that can classify all the freeway ramp data against all other data in X 2. Denote this feature set as Ω 2. Step 4) Relabel data in X with arterial data samples as 1 and all others as 0. Denote this training data set as X 3. Select the features that are not in Ω 1 Ω 2 and that can best classify all the arterial data against all other data in X 3. Denote this feature set as Ω 3. Step 5) Relabel data in X with local roadway data samples as 1 and all others as 0. Denote this training data set as X 4. Select the features that are not in Ω 1 Ω 2 Ω 3 and that can best classify all the local roadway data against all others in X 4. Denote this feature set as Ω 4. Step 6) Output feature set Ω new =Ω 1 Ω 2 Ω 3 Ω 4. When the above feature-selection algorithm was applied to an initial set (Ω) of 47 features, as suggested by Langari and Won [14], we obtained the set (Ω new ) of 14 features shown in Table III. B. Optimal Window Size and Time Step in Online Predicting Since we attempt to predict the roadway type in the near future, the driving speed in the last segment, i.e., [t c ΔZ, t c ], where t c is the current time, is used to predict the road type on which the driver is during the time period [t c,t c +Δt]. The prediction is made at time steps kδt, k =1, 2,...As discussed above, the window size of the speed profile segments is ΔZ, and the time interval over which the prediction is made is Δt. Fig. 6 illustrates these two parameters on the speed profile of the urban dynamometer driving schedule (UDDS) drive cycle. The x-axis represents the time during a drive cycle, and the y-axis represents the vehicle speed in meters per second. The segments shown have equal sizes of ΔZ = 150 s, and the time step Δt = 100 s. Please note that Δt = 100 sis chosen here only for clarity of illustration. In reality, as we will show, Δt should be smaller than 100 s. The two parameters are important for the accuracy of prediction. Since features characterizing road types are extracted from the speed profile of the vehicle in the time interval [t c ΔZ, t c ],ifδz is too small, then the segment may be too small to contain useful information. If ΔZ is too big, then the segment may contain obsolete information. Once ΔZ is determined, the 14 features presented in Table III are extracted from the speed profile within the time interval [t c ΔZ, t c ] and are used as the input feature vector to the neural network described in Section III-C. The time step Δt also needs to be properly determined. If Δt is too short, then it would imply that the prediction routine would run often. If it is too long, then the roadway type may change during the near future horizon, i.e., [t c,t c +Δt]. The optimal window size and optimal time step are determined through a series of experiments by varying ΔZ in a reasonable range, such as 50, 100, 150, and 200, and Δt =1, 2, 3, 5, 10, 15, and 20 s. To give a more realistic estimate of generalization, we use a fivefold cross validation method in this experiment. For every pair of window size and time step (ΔZ, Δt), we generate the signal segments with length =ΔZ at every time step Δt on all the 11 standard FS drive cycles and another 11 drive cycles provided in the PSAT library, namely, 1) UDDS; 2) HWFET; 3) US06; 4) SC03; 5) LA92; 6) IM240; 7) Rep05; 8) NY City; 9) HL07; 10) Unif01; and 11) Arb02. Let Ψ(ΔZ, Δt) denote the set of all the signal segments generated from these 22 drive cycles using the window size and time step (ΔZ, Δt). We randomly partition Ψ(ΔZ, Δt) into five subsets, namely, Ψ 1, Ψ 2, Ψ 3, Ψ 4, and Ψ 5. Five neural networks are trained and validated as follows: The ith neural network, for i =1,...,5, is trained on four subsets Ψ j, 1 j 5 and j i and validated on the data set subset Ψ i.fig.7shows the prediction accuracy of the classifiers trained with different window sizes and time step sizes averaged through the fivefold cross validations on all the training data sets [see Fig. 7(a)] and the validation data sets [see Fig. 7(b)]. The results obtained from both the training and test data show that performances are stabilized when ΔZ increases to 150 s since the performances of ΔZ = 150 s and ΔZ = 200 s are very close. As for the time steps, the performance generally improves when Δt decreases, and Δt =1gives the best performances over all window sizes. This makes sense since the more frequently we predict, the

8 4748 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 TABLE III FOURTEEN FEATURES SELECTED FOR ROADWAY-TYPE PREDICTION Fig. 6. Segments of a speed profile. more likely we will catch all the road type transitions. However, as we stated before, smaller time steps demand more computational power. As a tradeoff, Δt =3and ΔZ = 150 were used in the experiments presented in this paper since they gave good performances on both the training and test data. In Section V, we also analyzed the fuel efficiency with three different time steps. C. Training a Neural Network to Predict Road Types We developed a multilayered multiclass neural network, namely, NN_RT&TC, for the prediction of road types and traffic congestion levels. Fig. 8 shows the architecture of NN_RT&TC. The input layer has 14 nodes for the features specified in Table III, a hidden layer of 20 nodes, and 11 output nodes representing the 11 class labels {RT 1,...,RT 11 } to represent the 11 FS drive cycles. One important issue in a multiclass neural network classifier is the proper encoding of the classes in the output nodes of the neural network. We chose to use a one-hotspot method [24] described as follows. Since this is an 11-class classification problem, we need an 11-bit output layer. Each class is assigned a unique binary string (codeword) of length k. For example, class 1 is assigned a codeword of , class 2 is assigned a codeword of , class 3 is assigned of a codeword , etc. The advantage of this encoding is that it gives enough tolerance among different classes. The neural network is trained using the well-known backpropagation algorithm for weight update. Based on the study results presented in the last section, we use ΔZ = 150 s and Δt =3s. The training and test data are generated from 11 Sierra data and 11 PSAT drive cycles as follows. The feature vector x 1,x 2,...,x 14 is generated as follows. For each drive cycle DC(t) (0 t t e ), DC segments are generated on the intervals s 0 =[t 0, ΔZ),...,s k =[kδt, ΔZ + kδt),...,s ke =[t e ΔZ, t e ], where k 1. From the speed function of each segment, we extract a vector of the 14 features specified in Table III. The feature vector extracted from every speed signal segment is labeled by the roadway type of its next segment since we are training the prediction function. There are totally 4399 segments generated from these 22 drive cycles. The separation of training and test data is through a random stratified sampling procedure. The resulting training data contain 3519 feature vectors, and the test data contain 880 feature vectors. The performance of the neural network on the fivefold cross validation is 95% on the training data and 94% on the test data. When NN_RT&TC is used inside a vehicle to predict the roadway type at time t c, the vector of the 14 features is extracted from the vehicle speed during the time interval [t c 150 s, t c ].

9 PARK et al.: VEHICLE POWER CONTROL BASED ON MACHINE LEARNING OF OPTIMAL CONTROL PARAMETERS 4749 Fig. 7. Prediction accuracies using various window sizes and time steps. Fig. 9. Example of segmented, labeled, and predicted drive cycle LA92. Fig. 8. Architecture of NN_RT&TC. by the neural network NN_RT&TC are shown in blue. Notice that there is a delay in the prediction for the first 150 s. The output from NN_RT&TC is the roadway type that is used to produce the optimal power distribution during time interval [t c,t c +3 s]. We use 11 PSAT drive cycles as test data to evaluate the UMD-IPC system. The PSAT drive cycles can be considered as composites of the 11 classes of roadway types and traffic congestion levels. Fig. 9 shows an example of a drive cycle LA92 that is segmented and labeled according to the definition of the 11 standard FS RT&TC classes, as defined in [18]. The x-axis indicates the time, and the y-axis indicates the speed in meters per second. The prediction results generated IV. UMD_IPC: AN INTELLIGENT ONLINE VEHICLE POWER CONTROLLER The intelligent power controller UMD_IPC, which contains the neural network NN_RT&TC, has been fully implemented in the PSAT simulation environment. Fig. 10 gives the major computational steps of UMD_IPC at any given time t during a real-world drive cycle. The UMD_IPC has the knowledge base KB = {ΔZ = 150 s, Δt =3 s, ϕ i 1, ϕ i 2,λ i,k i P, and Ki I i = 1,...,11}, which is generated by the machine-learning algorithm presented in Section III.

10 4750 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 Fig. 10. Computational flow of UMD_IPC: an intelligent vehicle power controller. At any time t during a real-time drive cycle in a vehicle system, UMD_IPC is able to obtain the vehicle state V_state(t) = {v s (t),p d (t),p l (t),ω(t), SOC(t)}. If the vehicle is at the start mode, i.e., t<δz, UMD_IPC uses the default power control. When t =ΔZ, UMD_IPC gets the current vehicle state V_state(t) and calls the neural network NN_RT&TC to make the first prediction of the roadway type and traffic congestion level. Based on the road type R(t) predicted by NN_RT&TC, UMD_IPC retrieves the optimal control parameters associated with the road type r = R(t), ϕ r 1, ϕ r 2, λ r, K r P, and K r I. If this is the first prediction, λr is used as the initial value, i.e., λ 0. The battery energy at the current time E s (t) is calculated based on the current SOC(t) followed by the calculation of λ(t) using the following formula: λ(t) =λ 0 + KP r (E s (0) E s (t 1)) t 1 + KI r (E s (0) E s (p)). (17) p=1 Fig. 11 shows the λ(t) values in the simulation of all 11 PSAT drive cycles, which change with time and the road-prediction results.

11 PARK et al.: VEHICLE POWER CONTROL BASED ON MACHINE LEARNING OF OPTIMAL CONTROL PARAMETERS 4751 Fig. 13. speed. P alt_ max (ω): max mechanical alternator power with alternator Fig. 11. λ values for all 11 PSAT drive cycles. Fig. 14. P s boundary at time t as a function of SOC(t) value. The optimal engine power at time t is calculated using the formula P o eng(t) =P d (t)+g e2m (P l + η in2out (P o s (t)),ω). (20) Both P o s (t) and P o eng(t) are sent to the vehicle system, and the UMD_IPC continues the process at time t +1. Fig. 12. P eng_ max(ω): max engine power with engine speed. The instantaneous fuel-rate matrix at time t, i.e., F_R(P s (t),t P d (t),p l (t),ω(t)), is calculated through the following procedure based on the current constraints of P s (t), the engine power P eng, the engine speed ω(t), and the engine efficiency map Φ eng, which is provided by the vehicle system as a function of engine power and engine speed. For every P s (t) such that P s_ min(t) <P s (t) <P s_ max(t), the engine power at time t is calculated as follows: P eng (t) =P d (t)+g e2m (P l (t)+η in2out (P s (t)),ω) (18) where G e2m calculates the mechanical power based on the alternator efficiency map Φ alt for the given electrical power P e (t) =P l (t)+η in2out (P s (t)) at the given speed ω, and η in2out calculates the corresponding battery power output at the terminal based on the battery efficiency map Φ bat shown in Fig. 2 for the given the internal battery power P s. Then, F_R(P s (t),t P d (t),p l (t),ω(t)) = Φ eng (P eng (t),ω(t)). The optimal power to be charged to or discharged from the battery is obtained by searching through the fuel-rate matrix for the P s that minimizes the following quantity: P o s (t) = arg min P s (t) {F_R (P s(t),t) λ(t)p s (t)} (19) where P s_ min(t) P s (t) P s_ max(t). V. E XPERIMENTS UMD_IPC has been implemented in a conventional vehicle model provided by the PSAT software, namely, a Ford Taurus with a 95-kW 1.9-L Spark Ignition engine, five-gear manual transmission, a V 2-kW alternator, and a 66-A h/12-v lead acid battery. Since electrical loads in passenger vehicles, usually, are no larger than 1000 W, a constant electrical load P l = 1000 W is used in all the simulations. Figs. 12 and 13 show the engine and alternator constraints of the Ford Taurus model provided by the PSAT program, and Fig. 14 shows the constraints of P s with various SOC values. The online controller UMD_IPC is applied to all 11 PSAT drive cycles. Figs. 15 and 16 show the detailed experiment results generated from the three most interesting driving cycles, namely, UDDS, LA92, and UNIF01. UDDS, which is sometimes called FTP72, represents city driving conditions in an urban area with frequent stops. LA92, which is also called unified cycle, was constructed from segments of an actual driving recording in Los Angeles. It is a more aggressive driving cycle than the federal test procedure (FTP) as it has higher speeds, higher acceleration rates, fewer stops per meter, and less idle time (see Fig. 9). The UNIF01 cycle developed by Sierra Research for the California Air Resources Board is a modified form of the LA92. For the purpose of comparison, we applied offline DP controller to these three drive cycles in the attempt of finding the optimal benchmark performances.

12 4752 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 Fig. 15. SOC comparison on three driving cycles. The x-axis represents time measured in seconds, and the y-axis represents the SOC measured in percentages. (a) SOC compensation during driving cycle UDDS. (b) SOC compensation during driving cycle UNIF01. (c) SOC compensation during driving cycle LA92. It should be kept in mind that the DP controller is not applicable to online control [3], [9] since it requires full knowledge of the entire drive cycle to optimize the power management strategy at each time. In Fig. 15, we show that the battery SOC generated during the simulation run of the three drive cycles using the three different power controllers, namely, the offline DP controller, the Ford Taurus controller provided by PSAT, and the UMD_IPC controller. It can be observed that the SOC curves generated by the UMD_IPC from all three drive cycles behave quite similarly to the respective ones generated by the offline DP controller, whereas the SOC curves generated by the Ford

13 PARK et al.: VEHICLE POWER CONTROL BASED ON MACHINE LEARNING OF OPTIMAL CONTROL PARAMETERS 4753 Fig. 16. Comparisons of the battery power P s generated by the three controllers DP, UMD_IPC, and Ford Taurus. (a) Battery power P s generated during drive cycle UDDS.(b) Battery power P s generated during drive cycle UNIF01. Taurus controller are significantly different from the optimal curves. Fig. 16 presents battery power P s dynamically generated by the three controllers for UDDS and UNIF01 cycles. The battery power for DP and UMD_IPC controllers are discretized by a step size of 50 W. These graphs clearly show that the battery powers generated by the UMD_IPC controller are close to the optimal ones generated by DP. Fig. 17 presents the performance comparison with respect to fuel consumption on all 11 PSAT drive cycles generated by the same three power controllers; (a) presents the fuel consumption, and (b) presents the fuel saved. We use the fuel consumed by the Ford Taurus in PSAT as the baseline to measure the fuel saved by the DP and UMD_IPC controllers. The UMD_IPC gave more than 2% savings on fuel consumption from six drive cycles, namely, UDDS, UNIF01, LA92, 505UDDS, SC03, and TripEPA. In particular, for the UDDS, UNIF01, and LA 92 drive cycles, UMD_IPC s performances are very close to the optimal (DP) controller: For the UDDS drive cycle, UMD_IPC saved 3.95% fuel, while the DP controller saved 4.05%; for the UNIF01 drive cycle, UMD_IPC

14 4754 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 Fig. 17. Performance comparison on fuel consumption. (a) Total fuel consumption with P l = 1000 W. (b) Fuel saving with P l = 1000 W. Fig. 18. Fuel efficiency comparisons on different time steps used by UMD_IPC.

15 PARK et al.: VEHICLE POWER CONTROL BASED ON MACHINE LEARNING OF OPTIMAL CONTROL PARAMETERS 4755 saved 3.29% fuel, while the DP controller saved 3.47%; and for the LA 92 drive cycle, UMD_IPC saved 3.05% fuel, while the DP controller saved 3.15%. These results demonstrate that UMD_IPC is able to realize good fuel-economy improvements over the existing conventional control strategy in all drive cycles, and on some drive cycles, it can give near-optimal performances. We also studied the fuel efficiency with three different time steps, i.e., Δt =1, 3, and 5 s, and the results are shown in Fig. 18. It appears that UMD_IPC gave similar performances for all three time steps within the precision of two digits. To analyze the computational efficiency of UMD_IPC, the following experiment is conducted using a computer that has a 3-GHz Intel Core 2 Duo processor and a 4-GB DDR3 random access memory. We applied UMD_IPC to the drive cycle UDDS with a time step of Δt =3 s for neural network prediction and update the optimal battery power P s every second. The entire simulation was completed in 300 s. Since UDDS has 1370 s, the time takes to calculate the optimal power setting, which includes the time needed for neural network prediction of roadway types, is 0.22 s on the average. VI. CONCLUSION The authors have presented an intelligent vehicle power controller, namely, the UMD_IPC, which has been developed through machine learning of optimal control parameters with respect to 11 FS road types and traffic congestion levels. UMD_IPC contains a neural network, namely, the NN_RT&TC, designed and trained for in-vehicle prediction of 11 different roadway types and traffic-congestion levels. We have also presented a feature extraction algorithm to extract effective features from vehicle speed segments as the input to NN_RT&TC and showed the importance of the two parameters, namely, ΔZ, which is the signal window size, and Δt, which is the prediction step, with respect to the accuracy of the prediction results of NN_RT&TC. An offline machine-learning algorithm was proposed to generate a knowledge base that contains the optimal control parameters for all 11 standard FS drive cycles. During an online control process, UMD_IPC, which is the proposed online controller, applies the appropriate optimal control parameters in the knowledge base to the current road type predicted by the neural network NN_RT&TC to generate optimal power to be charged to or discharged from the battery and the optimal engine power. UMD_IPC has been fully implemented in a conventional vehicle model, Ford Taurus, in the PSAT simulation environment and tested on the 11 drive cycles provided by the PSAT library. Our simulation results show that UMD_IPC gave a performance within less than 0.15% of the optimal DP, which is an offline optimal controller, for eight drive cycles, and less than 0.25% for the remaining drive cycles. The maximum fuel saved in comparison to the Ford Taurus controller is 3.95%. UMD_IPC saved more than 2% fuels in six drive cycles and more than 1% fuel in eight drive cycles. In conclusion, the proposed machine-learning technique, combined with roadwaytype prediction, is an effective approach in real-time intelligent vehicle power management. Furthermore, the proposed vehicle power control technology does not require any changes to the drive train and is therefore easy to implement in an existing conventional vehicle configuration. Currently, we are developing machine-learning technologies with applications to hybrid vehicle power-management systems. We anticipate that more significant fuel reduction will be achieved in hybrid vehicle power systems. REFERENCES [1] E. D. Tate and S. P. Boyd, Finding ultimate limits of performance for hybrid electric vehicles, presented at the Soc. Automotive Eng. Future Transp. Technol. Conf., Detroit, MI, 2000, Soc. Automotive Eng. Paper [2] S. Delprat, J. Lauber, T. M. Guerra, and J. Rimaux, Control of a parallel hybrid powertrain: Optimal control, IEEE Trans. Veh. Technol., vol. 53, no. 3, pp , May [3] C.-C. Lin, H. Peng, J. W. Grizzle, and J.-M. Kang, Power management strategy for a parallel hybrid electric truck, IEEE Trans. Control Syst. Technol., vol. 11, no. 6, pp , Nov [4] T. Hofman and R. van Druten, Energy analysis of hybrid vehicle powertrains, in Proc. IEEE Int. Symp. Veh. Power Propuls., Paris, France, Oct [5] I. Arsie, M. Graziosi, C. Pianese, G. Rizzo, and M. Sorrentino, Optimization of supervisory control strategy for parallel hybrid vehicle with provisional load estimate, in Proc. 7th Int. Symp. AVEC, Arnhem, The Netherlands, Aug [6] V. H. Johnson, K. B. Wipke, and D. J. Rausen, HEV control strategy for real-time optimization of fuel economy and emissions, presented at the Soc. Automotive Eng. Future Transp. Technol. Conf., Detroit, MI, 2000, Paper [7] G. Paganelli, G. Ercole, A. Brahma, Y. Guezennec, and G. Rizzoni, General supervisory control policy for the energy optimization of chargesustaining hybrid electric vehicles, Soc. Autom. Eng. Jpn. Rev., vol. 22, no. 4, pp , Apr [8] A. Sciarretta, L. Guzzella, and M. Back, A real-time optimal control strategy for parallel hybrid vehicles with on-board estimation of the control parameters, in Proc. IFAC Symp. Adv. Autom. Control, Salerno, Italy, Apr , [9] M. Koot, J. T. B. A. Kessels, B. de Jager, W. P. M. H. Heemels, P. P. J. van den Bosch, and M. Steinbuch, Energy management strategies for vehicular electric power systems, IEEE Trans. Veh. Technol., vol. 54, no. 3, pp , May [10] E. Ericsson, Variability in urban driving patterns, Trans. Res., Part D, vol. 5, no. 5, pp , Sep [11] E. Ericsson, Independent driving pattern factors and their influence on fuel-use and exhaust emission factors, Trans. Res., Part D, vol. 6, no. 5, pp , Sep [12] S.-I. Jeon, S.-T. Jo, Y.-I. Park, and J.-M. Lee, Multi-mode driving control ofaparallelhybridelectricvehicleusingdrivingpatternrecognition, Trans. ASME,J.Dyn.Syst.Meas.Control, vol. 124, no. 1, pp , Mar [13] I. Kolmanovsky, I. Siverguina, and B. Lygoe, Optimization of powertrain operating policy for feasibility assessment and calibration: Stochastic dynamic programming approach, in Proc. Amer. Control Conf.,Anchorage, AK, May 2002, vol. 2, pp [14] R. Langari and J.-S. Won, Intelligent energy management agent for a parallel hybrid vehicle Part 1: System architecture and design of the driving situation identification process, IEEE Trans. Veh. Technol.,vol.54,no.3, pp , May [15] J.-S. Won and R. Langari, Intelligent energy management agent for a parallel hybrid vehicle Part 2: Torque distribution, charge sustenance strategies, and performance results, IEEE Trans. Veh. Technol., vol. 54, no. 3, pp , May [16] Y. L. Murphey, Intelligent vehicle power management An overview, in A Chapter in the Book Studies in Computational Intelligence (SCI), vol Berlin, Germany: Springer-Verlag, 2008, pp [17] T. R. Carlson and R. C. Austin, Development of speed correction cycles, Sierra Res., Inc., Sacramento, CA, Rep. SR , [18] SCF improvement Cycle development, Sierra Res. Sacramento, CA, Sierra Rep. No. SR , [19] Highway Capacity Manual 2000, Trans. Res. Board, Washington, DC, [20] J. T. B. A. Kessels, Energy management for automotive power, Ph.D. dissertation, Mech. Eng., Techn. Univ. Eindhoven, Eindhoven, The Netherlands, 2007.

16 4756 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 9, NOVEMBER 2009 [21] F. Ferri, P. Pudil, M. Hatef, and J. Kittler, Comparative study of techniques for large scale feature selection, in Pattern Recognition in Practice IV, E. Gelsema and L. Kanal, Eds. Amsterdam, The Netherlands: Elsevier, 1994, pp [22] Y. L. Murphey and H. Guo, Automatic feature selection A hybrid statistical approach, in Proc. Int. Conf. Pattern Recog., Barcelona, Spain, Sep. 3 8, 2000, pp [23] J. A. Crossman, H. Guo, Y. L. Murphey, and J. Cardillo, Automotive signal fault diagnostics Part 1: Signal fault analysis, feature extraction, and quasi optimal signal selection, IEEE Trans. Veh. Technol., vol. 52, no. 4, pp , Jul [24] G. Ou and Y. L. Murphey, Multi-class pattern classification using neural networks, Pattern Recognit., vol. 40, no. 1, pp. 4 18, Jan Jungme Park received the B.S. degree in statistics from Korea University, Seoul, Korea, in 1989 and the M.S. and Ph.D. degrees in computer science from the University of Alabama, Tuscaloosa, in She is currently a Research Scientist with the Department of Electrical and Computer Engineering, University of Michigan, Dearborn. Her current research interests include computer vision, optimization, and vehicle power management of conventional and hybrid electric vehicles. Zhihang Chen received the Ph.D. degree in applied mathematics from Peking University, Beijing, China, in He is currently a Research Scientist with the University of Michigan, Dearborn. His research interests include machine learning and intelligent systems, with applications to vehicle power management. M. Abul Masrur (M 84 SM 93) received the Ph.D. degree in electrical engineering from Texas A&M University, College Station, in Between 1984 and 2001, he was with Ford Research Laboratories and then joined the U.S. Army RDECOM-TARDEC, Warren, MI, where he has been involved in various vehicular electric powersystem architecture concepts, electric power management, and inverter fault diagnostics. Dr. Masrur was the recipient of the Best Automotive Electronics Paper Award from the IEEE Vehicular Technology Society for his transactions papers in 1998 and the 2006 Society of Automotive Engineers Environmental Excellence in Transportation Award. He is the current Chair of the Motor Subcommittee within the IEEE Power and Energy Society. He served as an Associate Editor for the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY from 1999 to Anthony M. Phillips received the B.A. degree (magna cum laude) in physics from Gustavus Adolphus College, St. Peter, MN, in 1990 and the M.S. and Ph.D. degrees in mechanical engineering control systems from the University of California, Berkeley, in 1993 and 1995, respectively. Upon completing his study, he joined the Ford Motor Company, Dearborn, MI as a Product Development Engineer. He was appointed as a Technical Expert when he joined the Research and Advanced Engineering staff in In his current position as a Senior Technical Leader, he has responsibility for Ford s advanced vehicle control system development for hybrid and fuel-cell electric vehicles. He is a member of the Editorial Board of the International Journal of Alternative Propulsion. His research interests include vehicle energy management, distributed system control, and control system development tools and methods. He is the holder of 29 U.S. and international patents in automotive controls. Dr. Phillips is a member of the Society of Automotive Engineers and the American Society of Mechanical Engineers. Leonidas Kiliaris was born in Trenton, MI, in He received the B.Sc.Eng. and M.Sc.Eng. degrees in electrical engineering from the University of Michigan, Dearborn, in 2006 and 2009, respectively. He is currently conducting research with the University of Michigan in power management of lightand heavy-duty conventional and hybrid electric vehicles. Yi Lu Murphey (SM 97) received the Ph.D. degree in computer engineering from the University of Michigan, Ann Arbor, in She is currently a Professor and the Chair of the Department of Electrical and Computer Engineering, University of Michigan, Dearborn. Her current research interests include machine learning, computer vision, and intelligent systems, with applications to engineering diagnostics, vehicle power management, and robotic vision systems. Ming L. Kuang received the B.S. degree in mechanical engineering from the South China University of Technology, Guangzhou, China, in 1982 and the M.S. degree in mechanical engineering from the University of California, Davis, in Since 1991, he has been with the Ford Motor Company, Dearborn, MI, in various engineering positions. He became a Technical Expert in 2000 for the Escape Hybrid vehicle program and played a critical role in the development and implementation of the vehicle/powertrain control system, delivering the first Ford Escape Hybrid and Mercury Mariner Hybrid vehicles to production. He is currently a Technical Leader in vehicle controls in research and advanced engineering. His primary research interests include vehicle control architecture, vehicle control system development, and implementation methodologies, as well as advanced vehicle control algorithm development for hybrid and fuel-cell vehicles. He is the author or coauthor of 20 technical papers in various engineering journals and conferences. He is the holder of 36 U.S. and international patents. Mr. Kuang was the recipient of the 2005 Henry Ford Technology Award and the Society of Automotive Engineers 2007 Henry Ford II Distinguished Award for Excellence in Automotive Engineering.