IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 9, NO.

2, MAY 2013

679

Estimation of Battery State of Health Using Probabilistic Neural Network

Ho-Ta Lin, Tsorng-Juu Liang, Senior Member, IEEE, and Shih-Ming Chen, Student Member, IEEE
AbstractIn this study, a probabilistic neural network (PNN) is used to estimate the state of health (SOH) of Li-ion batteries. The accurate prediction of SOH can help avoid inconveniences or fatal accidents from the sudden malfunction of the battery. A total of 110 pieces of Li-Co batteries are used. Constant current/voltage recharging and constant current discharging are performed for the life-cycle test of the battery. The data obtained from the recharging and discharging electric characteristics as well as the life-cycle test of the battery are used to estimate the SOH of the battery. The test data show that the constant current charging time, the instantaneous voltage drop at the start of discharging, and the open circuit voltage are the most important characteristics for estimating the SOH of the battery. The PNN is trained using 100 pieces of batteries. The remaining 10 pieces are used to verify the feasibility of the proposed method. The effectiveness of the PNN training using a number of samples is discussed and analyzed. The results show that the average error of the prediction is 0.28% and the standard deviation is 1.15%. The computation time of the PNN is 62.5 ms. Thus, the proposed method can accurately estimate the SOH of the battery in a short period. Index Terms Lithium-ion (Li-ion) battery, neural, probabilistic neural network (PNN), state of charge (SOC), state of health (SOH).

current, and SOC [9], modeling the hysteresis voltage versus the SOC during charging and discharging with the extended Kalman lter [10], measuring the charging voltage and current using articial neural networks and the extended Kalman lter [11], using the voltage responses of a pulse discharge current and counting the value of the SOC for capacity estimation [12], and using a cooperation of the SOC-OCV correlation and the simple algorithms of the battery impedance estimation [13]. The usage time, manufacturing process, as well as the charging and discharging conditions have a signicant effect on the accurate diagnosis of the SOH and SOC of a battery. The SOH is dened as the maximum charge that can be released after the battery has been fully charged. It can be expressed as a percentage of the rated capacity [3] % where (1)

I. INTRODUCTION

HE battery management system is a very important function for the battery-based energy storage system in various applications, such as portable devices and electric/hybrid/ plug-in hybrid vehicles. The most widely used battery types include nickel-metal hydride (NiMH), lithium-ion (Li-ion), and lithium-ion polymer (LiPo) [1][5]. The prediction of a batterys state of health (SOH) and state of charge (SOC) is an essential technique for battery management. Many studies have been conducted to estimate the SOC accurately [6][13]. Such methods include using an equivalent circuit of a battery with electrochemical impedance spectroscopy [6], building a partially linear battery model that also measures battery lifetime [7], using equivalent circuit models of a battery by applying the Bode plot and the numerical method with the frequency response [8], applying a diagonal recurrent neural network and a Kalman lter by the data collection of the discharging voltage,

is the maximum charge that can be released, and is the batterys rated capacity. The (1) indicates that the SOH of the battery should be determined rst, and then the estimation of the SOC will be more accurate. A probabilistic neural network (PNN) has a fast learning speed because the networks connection weight is only set once; the necessary data from the training sample can be entered directly without iteration. For this reason, PNN has received much attention. The basic principle of PNN begins from a probability model, specifically, Bayesian classiers. Therefore, the current study proposes a method using PNN for the SOH estimation of the battery [14][33]. In this research, 110 pieces of Li-Co batteries are used. Section II presents the life-cycle testing of the batteries. The electrical characteristics of the life-cycle test are also discussed. The fundamental principle of PNN is introduced in Section III. The SOH estimation design using PNN is discussed in Section IV. II. LIFE-CYCLE TESTING OF THE BATTERY

Manuscript received August 19, 2011; revised May 15, 2012; accepted September 24, 2012. Date of publication October 04, 2012; date of current version January 09, 2013. This work was supported in part by the Bureau of Energy, Ministry of Economic Affairs, Taiwan, under Project 101-D0204-2, and by the National Science Council of Taiwan under Project 101-3313-E-006-017. Paper no. TII-11-430. The authors are with the Department of Electrical Engineering/Advanced Optoelectronic Technology Center (AOTC)/Green Energy Electronics Research Center (GREERC), National Cheng-Kung University, Tainan 701, Taiwan (e-mail: tjliang@mail.ncku.edu.tw). Digital Object Identier 10.1109/TII.2012.2222650

In this research, 110 pieces of 18 650 Li-Co batteries are used for the life-cycle test, which is performed in a laboratory. The battery is tested within a constant temperature chamber at C. The detailed electrical characteristics of the batteries are shown in Table I. The Fig. 1 illustrates the diagram for the lifecycle testing of the batteries. The test modes of the batterys life-cycle test are illustrated , and are dened as the constant in Figs. 2 and 3. current charging time, the instantaneous voltage drop at the start of discharge, and the open circuit voltage of a fully discharged battery after resting for one minute, respectively. CC and CV denote the constant current state and the constant voltage state,

1551-3203/$31.00 2012 IEEE

680

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 9, NO. 2, MAY 2013

Fig. 1. Diagram of the characteristic test of the battery.

TABLE I ELECTRICAL CHARACTERISTICS OF THE TESTED LI-CO BATTERY

Fig. 3. Flowchart of the life-cycle test of the battery.

the battery reaches 4.2 V, this operating mode ends and then changes to the CV mode. Mode II : The battery is charged with 4.2 V, and the charging current is gradually reduced. When reduces to 0.02 C or less, the charging period stops. Mode III : The battery is discharged with constant current of 1 C. The discharging mode ends as the battery voltage becomes less than the battery cut-off voltage of 3 V. Mode IV : The battery circuit is opened for 1 min. is increased slightly because of the batterys internal reaction. After the one-minute break, the resting mode ends and changes to the next cycle. The battery life-cycle test results show that the SOH of the battery is affected by , and . Fig. 4(a) shows the SOH versus the battery charging time . A nonlinear relationship exists between and SOH. A longer charging time under a specic CC charging condition indicates that the SOH is higher. Fig. 4(b) and (c) show that the SOH is higher with a smaller and . The correlations between SOH and , and are complex and nonlinear with an overall band distribution. Fig. 5 shows the SOHs of 100 batteries versus , and .
Fig. 2. Voltage and current of the battery in one test cycle.

III. PNN Neural networks are frequently employed to classify patterns by learning from examples. Pattern statistics are determined from a set of training samples, and then new patterns are classied based on these statistics. Heuristic approaches are often used to discover the underlying class statistics. They require long computation times for training. Incremental adaptation approaches are susceptible to convergence of an error function to the local minima. The network structure of PNN is similar to that of back propagation, but the activation function is replaced by one of the class of functions, such as exponential. Training

respectively. , and indicated the discharge current, the battery voltage, and the battery current, respectively. One test cycle has four operating modes, which are described in the following. Mode I : The battery is charged with a constant current of 0.5 C. The battery voltage is gradually increased. During this period, the lithium ions in the positive electrode (LiCoO ) move through the electrolyte into the negative electrode. As

LIN et al.: ESTIMATION OF BATTERY STATE OF HEALTH USING PROBABILISTIC NEURAL NETWORK

681

Fig. 4. SOH versus (a)

, (b)

, and (c)

with 1 Li-Co battery.

Fig. 5. SOH versus (a)

, (b)

, and (c)

of 100 Li-Co batteries.

PNN requires only a single pass. It is based on the Bayesian decision boundaries as the number of training samples grows. After the PNN is trained, all training samples are stored and then used in classifying new patterns. The computation time of PNN is much faster than that of back propagation types. PNN can be programmed easily into processors or chips to achieve even higher speeds through the use of parallel hardware [23][33]. The basic principle of PNN is based on a probability model, specically, Bayesian classiers. A group of training samples with -dimension vectors are considered as a group of sample points in an -dimension space. PNN then uses these sample points to estimate the probability density function for each classication in the entire sample space [32], [33]. If one classication question has classications of , the classication rule for this question is determined by -dimensional characteristic vector . The probability density function for each category in this -dimensional sample space is the characteristic vector function . The three hypotheses of PPN are as follows.

1) The probabilistic density function mode for each classication is identical. 2) The common probabilistic density function mode is a Gaussian distribution or a normal distribution. 3) The deviation matrix for the Gaussian probabilistic density function in each classication is a diagonal matrix, and the value for each diagonal element is identical. The purpose of PNN is to classify. The probabilistic density function for each classication is described in (2), (3), and (4) [32], [33] (2) (3) (4) where is the number of sample points in classication is the weight of the th sample point in classica-

682

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 9, NO. 2, MAY 2013

Fig. 6. PNN structure. Fig. 7. Training structure of PNN.

tion is the characteristic vector (the input vector by the lesson example) of the th sample point in classication is the square of the Euclidean distance between classication vector and sample point characteristic vector , and represents the smoothing parameters. The PNN network structure is displayed in Fig. 6 [14][33]. The structure includes the following layers. The input layer is used to express the network input variable and the amount of neurons. It uses a linear conversion function, . The pattern layer is used to express the classication training sample. Each neuron expresses one training sample. The connection between this training sample and the neural link is stored in the characteristic vector for the classication training sample. Thus, the weighted matrix between the input layer and the pattern layer is , which has the th value for the th classication training sample input vector. The summation layer is used to express information on the classication. Each neuron expresses one classication. The connection between this classication and the neural link in the characteristic layer is stored in the classication information for the classication training sample. The weighted matrix between the pattern layer and the summation layer is thus , where is the th weighted value for the th classication training sample classication vector. The output layer is used to express the nal classication results. The vector with the maximum value for the probabilistic density function is chosen as the output layer output vector, as shown in (5) (5) IV. SOH ESTIMATION OF THE BATTERY This study selects the three signicant characteristics of , and as the SOH patterns to test. The entire estimation calculation process can be separated into two parts: the training process and the verication process. The PNN structure and the SOH estimation process are shown in Figs. 7 and 8, respectively. The entire process is described in the following subsections.

A. Training Process and are set as The network are set to 0.1. Then, follows: is set to 3 to represent the parameters , and that are used to estimate the SOH. This structure is illustrated in Fig. 6. The parameter indicates the number of the selected training samples. Each battery selected for PNN SOH training has 20 sets of the life-cycle testing data. Given that 100 battery pieces are used in the life-cycle test, the maximum selectable training samples are 2000. The performance under various training samples is also discussed. The number of SOH training sample is set as 100, 200, 400, 600, 800, 1000, 1200, 1400, 1600, 1800, and 2000. Set represents the classication characteristic matrix for the training sample, whereas set denotes the matrix for training sample classications. The weight matrix between the input layer and the characteristic layer is set as , which is the th value for the characteristic vector of the th classication training sample. The weight matrix between the characteristic layer and the summation layer is set as . , where is the th weight value for the output vector of the th classication training sample, as tabulated in Table II. B. Verication Process The internal PNN is computed as described in Section III. The remaining 10 battery pieces are used as samples for verication. Each battery randomly selects 20 sets of the SOH verication points. A total of 200 sets of the SOH verication sample classications are made available. , and are set to represent the th value of the characteristic input vector matrix for each verication sample. is set to represent the verication sample classication matrix for sets, where . The PNN is trained with various training samples and then tested with the SOH verication sample. Afterward, the calculated values are compared with the actual verication sample values. The results of the PNN that is trained with 2000 sets of the SOH training samples shows that the average error of the prediction is 0.28%, and the standard deviation is 1.15%. The

LIN et al.: ESTIMATION OF BATTERY STATE OF HEALTH USING PROBABILISTIC NEURAL NETWORK

683

Fig. 9. Prediction error of SOH versus number of training sample with and .

TABLE III SOH VERIFICATION RESULTS WITH THE REMAINING 10 BATTERIES

Fig. 8. Estimation process of SOH.

TABLE II SOH TRAINING SAMPLES WITH 100 BATTERIES

the SOC. The testing time is 62.5 ms. Table III and Fig. 9 show these results. The average errors between the actual SOC and the estimated SOC in previous studies are 3%0.4% [7], [9][12]. In the present study, the test results are better than those in previous studies. Thus, the proposed method can help reduce the average error in SOC estimation when the batterys SOH has been accurately estimated. The test results verify that variations exist among batteries, a problem that affects the accuracy of the estimations. However, this problem can be resolved by using the batterys SOH patterns combined with the PNN and by using fewer data to obtain predictions with better accuracy. V. CONCLUSION In this study, PNN is used with the SOH pattern to estimate the SOH of Li-ion batteries. A total of 110 pieces of Li-Co batteries are utilized for the life-cycle test. The test data show that the parameters , and are the most important characteristics for the SOH estimation of Li-Co batteries. A total of 100 batteries are used to train the PNN, and the remaining 10 batteries are used to verify the proposed method. The training effect of PNN with a number of samples are discussed and analyzed. The results show that the estimation accuracy for the SOH is satised when the PNN is trained with 1600 sets from a total of 300 100 sets of the SOH samples that comprise the 100 sample batteries. Furthermore, the PNN is easy to use and is extremely fast for moderately sized databases. This result indicates that the method can improve the estimation of the batterys

results are very close to those of the PNN trained with the 1600 sets, whereby the average error for prediction and the standard deviation are 0.39% and 1.15%, respectively. The experiment result shows a total of 300 100 sets of the SOH samples comprising the 100 sample batteries. A total of 1600 sets of the SOH samples are used with the proposed method to accomplish the desired purpose, that is, to improve the accuracy of predicting

684

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 9, NO. 2, MAY 2013

SOC when the batterys SOH has been accurately estimated. This method can then be used to check the batterys health and whether the battery needs to be replaced. The user may save money and avoid inconveniences due to the batterys sudden failure. REFERENCES
[1] J. Cao and A. Emadi, Batteries need electronics, IEEE Ind. Electron. Mag., vol. 5, pp. 2735, Mar. 2011. [2] W. Luo, C. Lv, L. Wang, and C. Liu, Study on impedance model of Li-Ion battery, in Conf. IEEE Ind. Electron. Appl. Conf., Jun. 2011, pp. 19431947. [3] Handbook of Batteries, 3rd ed. New York: McGraw-Hill, 2002, pp. 1.36.1, 34.135.9. [4] Y. H. Liu, C. H. Hsieh, and Y. F. Luo, Search for an optimal vestep charging pattern for Li-Ion batteries using consecutive orthogonal arrays, IEEE Trans. Energy Convers., vol. 26, pp. 654661, Jun. 2011. [5] T. Hu, B. Zanchi, and J. Zhao, Simple analytical method for determining parameters of discharging batteries, IEEE Trans. Energy Convers., vol. 26, pp. 787798, Sep. 2011. [6] R. Li, J. Wu, Wang, and G. Li, Prediction of state of charge of Lithium-ion rechargeable battery with electrochemical impedance spectroscopy theory, in Proc. IEEE Ind. Electron. Appl. Conf., Jun. 2010, pp. 684688. [7] V. Agarwal, K. Uthaichana, R. A. DeCarlo, and L. H. Tsoukalas, Development and validation of a battery model useful for discharging and charging power control and lifetime estimation, IEEE Trans. Energy Convers., vol. 25, no. 3, pp. 821835, Sep. 2010. [8] J. Jang and J. Yoo, Equivalent circuit evaluation method of lithium polymer battery using bode plot and numerical analysis, IEEE Trans. Energy Convers., vol. 26, no. 1, pp. 290298, Mar. 2011. [9] N. Cui, C. Zhang, Q. Kong, and Q. Shi, A combined method of battery SOC estimation for electric vehicles, in Proc. IEEE Ind. Electron. Appl. Conf., Jun. 2010, pp. 11471151. [10] S. Qiua, Z. Chena, M. A. Masrurb, and Y. L. Murpheya, Battery hysteresis modeling for state of charge estimation based on extended Kalman lter, in Proc. IEEE Ind. Electron. Appl. Conf., Jun. 2011, pp. 184189. [11] M. Charkhgard and M. Farrokhi, State-of-charge estimation for Lithium-ion batteries using neural networks and EKF, IEEE Trans. Ind. Electron., vol. 57, no. 12, pp. 41784187, Dec. 2010. [12] M. Einhorn, F. V. Conte, C. Kral, and J. Fleig, A method for online capacity estimation of Lithium ion battery cells using the state of charge and the transferred charge, IEEE Trans. Ind. Appl., vol. 48, pp. 736741, Mar./Apr. 2012. [13] M. A. Roscher, O. S. Bohlen, and D. U. Sauer, Reliable state estimation of multicell lithium-ion battery systems, IEEE Trans. Energy Convers., vol. 26, pp. 737743, Sep. 2011. [14] N. K. Singh, M. Tripathy, and A. K. Singh, A radial basis function neural network approach for multi-hour short term load-price forecasting with type of day parameter, in Proc. IEEE Ind. Inf. Syst. Conf., Aug. 2011, pp. 316321. [15] W. D. S. Roshan, R. A. R. C. Gopura, and A. G. B. P. Jayasekara, Financial forecasting based on articial neural networks Promising directions for modeling, in Proc. IEEE Ind. Inf. Syst. Conf., Aug. 2011, pp. 3122327. [16] H. R. K. Nagahamulla, U. R. Ratnayake, and A. Ratnaweera, Monsoon rainfall forecasting in Sri Lanka using articial neural networks, in Proc. IEEE Ind. Inf. Syst. Conf., Aug. 2011, pp. 305309. [17] S. Hiremath and S. K. Patra, Transmission rate prediction for cognitive radio using adaptive neural Fuzzy inference system, in Proc. IEEE Ind. Inf. Syst. Conf., Jul. 2010, pp. 9297. [18] Z. Zhang and R. Liu, Application of fuzzy neural network in optical ber fusion defect recognition system by AS1773 protocol, in Proc. IEEE Ind. Inf. Syst. Conf., Jul. 2010, pp. 196199. [19] A. Gomperts, A. Ukil, and F. Zuruh, Development and implementation of parameterized FPGA-based general purpose neural networks for online applications, IEEE Trans. Ind. Inf., vol. 7, no. 1, pp. 7889, Feb. 2011. [20] T. Orlowska-Kowalska and M. Kaminski, FPGA implementation of the multilayer neural network for the speed estimation of the two-mass drive system, IEEE Trans. Ind. Inf., vol. 7, no. 3, pp. 436445, Aug. 2011.

[21] M. O. Efe, Neural network assisted computationally simple PI D control of a quadrotor UAV, IEEE Trans. Ind. Inf., vol. 7, no. 2, pp. 354361, May 2011. [22] J. J. Yin, W. Tang, and K. F. Man, A comparison of optimization algorithms for biological neural network identication, IEEE Trans. Ind. Electron., vol. 57, no. 3, pp. 11271131, Mar. 2010. [23] C. Xia, C. Guo, and T. Shi, A neural-network-identier and Fuzzycontroller-based algorithm for dynamic decoupling control of permanent-magnet spherical motor, IEEE Trans. Ind. Electron., vol. 57, no. 8, pp. 28682878, Aug. 2010. [24] A. Slowik, Application of an adaptive differential evolution algorithm with multiple trial vectors to articial neural network training, IEEE Trans. Ind. Electron., vol. 58, no. 8, pp. 31603167, Aug. 2010. [25] V. N. Ghate and S. V. Dudul, Cascade neural-network-based fault classier for three-phase induction motor, IEEE Trans. Ind. Electron., vol. 58, no. 5, pp. 15551563, May 2011. [26] S. Huang and K. K. Tan, Intelligent friction modeling and compensation using neural network approximations, IEEE Trans. Ind. Electron., vol. 59, no. 8, pp. 33423349, Aug. 2012. [27] Y. Pan and J. Wang, Model predictive control of unknown nonlinear dynamical systems based on recurrent neural networks, IEEE Trans. Ind. Electron., vol. 59, no. 8, pp. 30893101, Aug. 2012. [28] H. Zhuang, K. S. Low, and W. Y. Yau, Multichannel pulse-coupledneural-network-based color image segmentation for object detection, IEEE Trans. Ind. Electron., vol. 59, no. 8, pp. 32993308, Aug. 2012. [29] G. W. Chang, C. I. Chen, and Y. F. Teng, Radial-basis-function-based neural network for harmonic detection, IEEE Trans. Ind. Electron., vol. 57, no. 6, pp. 21712179, Jun. 2010. [30] C. I. Chen, Virtual multifunction power quality analyzer based on adaptive linear neural network, IEEE Trans. Ind. Electron., vol. 59, no. 8, pp. 33213329, Aug. 2012. [31] C. H. Lu, Wavelet Fuzzy neural networks for identication and predictive control of dynamic systems, IEEE Trans. Ind. Electron., vol. 58, no. 7, pp. 30463058, Jul. 2011. [32] F. J. Lin, M. S. Huang, P. Y. Yeh, H. C. Tsai, and C. H. Kuan, DSPbased probabilistic Fuzzy neural network control for Li-ion battery charger, IEEE Trans. Power Electron., vol. 27, no. 8, pp. 73823794, Aug. 2012. [33] F. J. Lin, Y. C. Hung, J. C. Hwang, I. P. Chang, and M. T. Tsai, Digital signal processor-based probabilistic fuzzy neural network control of in-wheel motor drive for light electric vehicle, IET Trans. Elect. Power Appl., vol. 6, no. 2, pp. 4761, Feb. 2012. Ho-Ta Lin was born in Kaohsiung, Taiwan. He received the B.S. and M.S. degrees in electrical engineering from I-Shou University, Kaohsiung, Taiwan, in 1999 and 2002, respectively. He is currently working toward the Ph.D. degree at National Cheng Kung University, Tainan, Taiwan. His research interests are power electronics, control system, energy conversion, and battery management system.

Tsorng-Juu (Peter) Liang (M93SM10) was born in Kaohsiung, Taiwan. He received the B.S. degree in electrophysics from the National Chiao-Tung University, Hsinchu, Taiwan, in 1985, and the M.S. and Ph.D. degrees in electrical engineering from the University of Missouri, Columbia, in 1990 and 1993, respectively. He is currently a Professor of Electrical Engineering and Director of Green Energy Electronics Research Center (GREERC), National Cheng-Kung University (NCKU), Tainan, Taiwan. He is on the Board of Directors for Compucase Enterprise Company, Ltd., and Catcher Technology Company, Ltd. He has authored or coauthored 50 journal and more than 100 conference papers. He was the Director of Electrical Laboratories, NCKU, from 2001 to 2004. His research interests include high efciency power converters, high efciency lighting systems, renewable energy conversion, and power ICs design. Dr. Liang is a Member of the IEEE Societies of Power Electronics, Industrial Electronics, Circuits and Systems, and Industrial Applications. Currently, he is the Associate Editor of the IEEE TRANSACTIONS ON POWER ELECTRONICS, the

LIN et al.: ESTIMATION OF BATTERY STATE OF HEALTH USING PROBABILISTIC NEURAL NETWORK

685

Associate Editor of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I, and the Technical Committee Chair of IEEE CAS Systems Power and Energy Circuits and Systems Technical Committee. In 2008, he received the Outstanding Engineer, The Chinese Institute of Electrical Engineering, Kaohsiung Chapter, and Outstanding Professor Award, Taiwan Power Electronics Conference. In 2010, he received Teaching Excellence Award, National Cheng Kung University and Outstanding Engineers Professor Award, The Chinese Institute of Electrical Engineering, Kaohsiung Branch.

Shih-Ming (Orion) Chen was born in Tainan, Taiwan. He received the B.S. and Ph.D. degrees in electrical engineering from the National Cheng-Kung University, Tainan, Taiwan, in 2003 and 2011, respectively. From 1991 to 1995, he was employed as a Research and Design Engineer at LUMIN Electronics, Taiwan. From 1995 to 1999, he was employed as a Section Manager of the Engineer Department, Delta Electronics, Inc., Taiwan. From 1999 to 2002, he was employed as a Manager in charge the Tainan R&D Center at Sino-American Electronic Company, Ltd., Taiwan. From 2002, he has worked as a section manager at Product Development of LCD-TV Head Division, CHIMEI Optoelectronics Corp. Since 2009, he has rejoined Delta Electronics as senior manager in the division of LCD display power. Now, he is an Assistant Researcher of Green Energy Electronics Research Center (GREERC) in the Department of Electrical Engineering, National Cheng Kung University. His research interest are dc-to-dc converters, photovoltaic inverter, switching power supply, uninterrupted power system, CCFL/EEFL inverter design, and LED driver and chromatics control.