外文翻译---智能超快速的低成本risc实施镍镉电池充电器 英文

Essentially, the output of the GRNN, given the inputs x,

is the weighted combination of the functional distance

between the input data and the center of the basis function

in the second (hidden) layer. The GRNN performance is

sensitive to the choice of the basis function. The ultra fast

charger presented in Ref. [9] invoked a Gaussian function

for computation of the charge current. However, it was

subsequently reported in Refs. [10,11] that a RBF can

performance.

employed in this work. That is, the charging current (Ic)

is given by

i¼1yiRBFðkx; xikÞ

i¼1RBFðkx; xikÞ

where x 2 RN is, in general, an input vector of dimension

N, RBF(.) is the radial basis function, jj.jj denotes the

Euclidean norm, yi are weights or parameters of the net-

work, xi 2 RN are known as the centers of RBF(.) and N

is the number of centers. In our case, x = [Ti, (dT/dt)i]T

where Ti and (dT/dt)i, respectively, denote the ith temper-

ature and ith temperature gradient obtained from Ref. [12].

3.2. Training of GRNN using GA

Similar to other neural network techniques, the GRNN

requires supervised training. There are various ways for

training the network. Among those are genetic algorithms,

which are probabilistic search techniques that emulate the

mechanics of evolution [14]. They are capable of globally

exploring a solution space, pursuing potentially fruitful

paths while also examining additional random points to

reduce the likelihood of settling for a local optimum.

In the GA, a set of variables for a given problem is

encoded into a string (population), analogous to a chromo-

some in nature. In addition, each string contains a lot of

alleles, and each feature of the system located at a specific

position in the string is called a gene. Each string, therefore,

contains a possible solution to the problem. The optimum

solution can be obtained by minimizing a fitness function.

Thus, those with lower fitness values will be chosen to be

the parents of the next generation while those with higher

fitness values are rejected. Creating the new offspring, the

selected parents strings undergo a reproduction process

such as crossover and mutation as described in Ref. [14].

By continuing such a procedure, the newer and fitter chro-

mosome evolves until a predefined stopping condition is

satisfied.

In this work, the allele is each individual control input

and the optimal charging current pair taken from Ref.

[12]. By collecting each pair together, the string is eventu-

ally produced. Initially, a number of pairs is randomly

selected. Then, at each iteration, the GA determines the fit-

ness function, which is essentially the MSE between the

real output [12] and computed output using the GRNN.

After a certain number of crossovers and mutations, the

GA finally finds the best set of input–output pairs. Note

that the number of pairs is equivalent to the number of

neurons in the hidden layer.

3.3. Reduction of computational complexity

In view of implementation, the computational complex-

ity of the GRNN should be moderately low. To this end, the

RBF(Æ), which usually is an exponential function, exp(Æ),

was replaced by a simple polynomial form. In particular,

the CSRBF originally presented in Ref. [15] was selected

owing to its simplicity and suitability for implementation

on a low cost RISC, i.e. a microcontroller. However, a sup-

port interval of the CSRBF is fixed and cannot be expanded

without changing a slant of the function, leading to a

numerical problem; dividing by zero, in the GRNN’s for-

ward phase, as may clearly be seen from Eq. (1). This

problem arises especially when a distance of the network

input from all RBF centers is greater than the support

length of the CSRBFs. Recently, a novel method of con-

structing a class of compact support radial basis functions,

p-RPPRBF, with a special feature was introduced [16]. Its

support interval can be extended by a parameter p to cap-

ture the input while the slant of the function is preserved.

The output of p-RPPRBF can be computed via the follow-

ing pseudocode shown in Table 1.

Fig. 3. GRNN structure.

Pusedocode for p-RPPRBF

BEGIN: Given distance value d, and a parameter p 2 {0,1,2,. . .}

/ = 1 � d/2(p + 1)

if / 6 0 then output :¼ 0

else output :¼ / · /

for n :¼ 1 to p

output :¼ / · output · output

END {note: corresponding support interval is [� 2(p + 1), 2(p + 1)]}

P. Petchjatuporn et al. / Energy Conversion and Management 49 (2008) 185–192

The proposed ultra fast charger using the GRNN was

realized on a low cost RISC PIC16F876A microcontroller

[17] with additional hardware circuitry. Its structure is

illustrated in Fig. 4. In addition to 8 kbyte programmable

memory, the RISC microcontroller features a 10 bit, five

channel successive approximation analog to digital A/D

converter. Note that only 8 bit resolution is sufficient for

our application. Software was developed using the C pro-

gramming language for realization of the GA trained

GRNN, cross compiled using the PCWH compiler and

programmed on a PIC 16F876A. Fig. 5 shows a picture

of the hardware prototype, and Fig. 6 shows the flowchart

describing the microcontroller operation.

The hardware circuitry includes a voltage control cur-

rent source, battery temperature and voltage detector,

digital to analog (D/A) converter and two voltage to cur-

rent (V/I) converters. In the proposed charging system,

the battery temperature, T, is measured by a temperature

sensor, particularly a thermistor. The data is converted

and fed to a microcontroller via a built in A/D converter.

Then, the temperature gradient, dT/dt, is computed using

a unit delay. Both T and dT/dt are used as the control

inputs of the GRNN as discussed in the previous sec-

tions. Computation of the GRNN output, i.e. the charg-

ing current, is performed in the microcontroller. With

replacement of the RBF by the p-RPPRBF, the memory

usage of the PIC16F876A is significantly reduced by

more than 69%. The D/A MAX503 is employed for con-

version of the charging current data to an analog voltage,

which is fed through the V/I XTR110 in order to pro-

duce finally the proper charging current supplied to the

Ni–Cd battery. In addition to the temperature, the com-

puted voltage gradient is monitored. As soon as the tem-

approaches

　　50 �C

gradient is detected, the charging process is stopped

immediately. Fig. 5 depicts a hardware prototype of

our proposed low cost intelligent ultra fast charger for

Ni–Cd batteries.

5. Experimental results

5.1. Performance and complexity tradeoff in GA trained

Simulations as well as experimentation of the hardware

prototyped were conducted to determine the number

of neurons in the hidden layer that yields the best

performance and complexity tradeoff. Firstly, the MSE

performance of the GA trained GRNN controller was

investigated using the MATLAB�/Neural Network Tool-

box [18]. In the simulations, a population size used in each

trial was 20. Using the GA, 2–6 of 561 input–output pairs

were selected as optimal data for training the GRNN.

Table 2 illustrates the MSE obtained from the GRNN with

different numbers of neurons in the hidden layer. Notice

that the GRNNn, where n = 2,. . .,6, designates the GRNN

with n neurons. The table confirms that six neurons is the

discharging

controller

　　16F876A

DC regulation

Thermistor

Fig. 4. Ultra fast charger structure.

Fig. 5. Hardware prototype.

discharged?

Discharge Process

Temperature out

of range (50oC)

get T and V from A/D

and calculate dT/dt

get Ic from

Output Ct to D/A

of V/I converter

Negative dT/dV

and Ic is at 0.5C rate

Ultra Fast Charging

Fig. 6. Ultra fast charging software flowchart.

MSE from different GRNN structure

No. processing element

　　0.3823

　　0.1473

　　0.0249

　　0.0238

　　0.0046

P. Petchjatuporn et al. / Energy Conversion and Management 49 (2008) 185–192

tional burden.

Experiments were performed to investigate the perfor-

mance of the charger with different numbers of neurons

in the hidden layer. A battery under test was the Panasonic

Ni–Cd battery, rated at 1.2 V 600 mA-H [19]. Fig. 7 shows

the battery voltage, current and temperature against time

achieved by the GA trained GRNN charger. Clearly, a

higher number of neurons means higher degrees of freedom

in the charger. Furthermore, the results indicate that the

battery charging times are 1700, 1600, 1500, 1100 and

1050 s for the proposed charger with the number of neu-

rons being 2–6, respectively. Notice that the battery char-

ger was not actually stopped, as one may easily verify

that the plots show that not only the temperature went well

　　50 �C,

observed. This was because the charger ‘‘cut-off’’ detection

was intentionally omitted during all experiments presented

in this paper.

5.2. Performance comparison with the conventional charger

The ultra fast charger was compared with the conven-

tional constant current charger. Initially, the battery was

charged with the conventional constant current charger

with a charging current of 0.5C. Fig. 8 shows the battery

voltage, current and temperature against time. It reveals

that the battery required approximately 8500 s to be fully

Fig. 7. Evolution of Panasonic battery voltage, current and temperature attained by the proposed GA-trained GRNN with (a) 2, (b) 3, (c) 4, (d) 5, (e) 6,

neurons in the hidden layer.

P. Petchjatuporn et al. / Energy Conversion and Management 49 (2008) 185–192

implies the charging time of roughly 2 h. Comparing this

charging time to that achieved by the proposed charger,

one can verify that the charging time is drastically reduced,

i.e. by 8-fold.

As mentioned in the previous sections, supplying the

battery with a high charging current may cause battery

damage. The following experiments aim to demonstrate

that the novel charger does not deteriorate the battery

energy storage capability.

Basically,

the battery was

charged using both the constant charge at 0.5C and the

intelligent charge with the six neuron GRNN charger until

it was fully charged. Subsequently, the battery was dis-

charged with a constant current of 300 mA (0.5C). Fig. 9

depicts the evolution of the battery voltage achieved by

both chargers. Apparently, the battery useful life time

(time for the battery voltage to reach 0.9 V) is 7500 s and

7100 s for the conventional charger and ultra fast charger,

respectively. This can be interpreted as follows. While

almost 10-fold in charging time reduction is achieved, a

mere drop of 5% in energy storage performance is experi-

enced in the novel charger when compared with the tradi-

tional charger.

To further confirm the superiority of our proposed char-

ger, an identical test was performed with another commer-

cial battery, namely a 1.2 V 1700 mAH Sanyo Ni–Cd

battery [20]. Figs. 10 and 11 illustrate the battery voltage,

Fig. 10. Evolution of Sanyo battery voltage, current and temperature

when supplied with a constant current charger at 0.5C.

Fig. 8. Evolution of Panasonic battery voltage, current and temperature

when supplied with a constant current charger at 0.5C.

Fig. 9. Evolution of Panasonic battery voltage attained by the conven-

tional charger (solid line) and ultra fast charger (dotted line).

Fig. 11. Evolution of Sanyo battery voltage, current and temperature

attained by the proposed GA-trained GRNN.

Fig. 12. Evolution of Sanyo battery voltage attained by the conventional

charger (solid line) and ultra fast charger (dotted line).

P. Petchjatuporn et al. / Energy Conversion and Management 49 (2008) 185–192

the constant current and intelligent charger, respectively.

Similar results were observed. In particular, the charging

time was around 8500 s for the traditional charger and

1050 s for the proposed charger. Subsequently, the dis-

charging test was performed. Fig. 12 shows the battery

voltage against time. It reveals that the battery useful life

time was 7000 s and 6900 s for the conventional charger

and ultra fast charger, respectively. The results again con-

firm that the novel charging mechanism hardly degrades

the battery capacity.

6. Conclusions

In this paper, an ultra fast charger for Ni–Cd batteries

has been developed. The proposed charger utilizes the

GA trained GRNN controller to attain ultra fast charging

techniques have been utilized towards its efficient imple-

mentation based upon a low cost RISC PIC 16F876A

microcontroller. Experimental results with commercial

grade batteries confirm the superiority of the charger. In

particular, while the intelligent charger can reduce the

charging time over the traditional charger by 10-fold, the

battery storage capacity is reduced by only 5%.

References

[1] Linden D. Handbook of Batteries. McGraw Hill Inc.; 1995.

[2] Khosla A, Kumar S, Aggarwal KK. ‘‘Fuzzy controller for rapid

nickel–cadmium batteries charger through adaptive neuro-fuzzy

inference system (ANFIS) architecture’’. In: Proceedings of the

NAFIPS 2003 Conference, 2003. p. 540–4.

[3] Gonzalez M, Ferrero FJ, Anton JC, Perez MA. ‘‘Considerations to

improve the practical design of universal and full-effective NiCd/

NiMH battery fast-chargers’’. In: Proceedings of the 14th APEC’99

Conference, vol. 1. 1999. p. 167–73.

[4] Gonzalez M, Perez MA, Diaz J, Ferrero FJ. ‘‘Ni–Cd and Ni–MH

battery optimized fast-charge method for portable telecommunication

applications’’. In: Proceedings of the 18th INTELTE’ 96, 1996. p.

　　525–9.

[5] Diaz J, Martı´n-Ramos JA, Pernı´a AM, Nuno F, Linera FF.

Intelligent and universal fast charger for Ni–Cd and Ni–MH

applications.

2004;51:857–63.

[6] Ullah Z, Burford B, Dillip S. Fast intelligent battery charging: neural-

fuzzy approach. IEEE Aero Electron Syst Mag 1996;11:26–34.

[7] Gonzalez M, Ferrero FJ, Anton JC, Perez MA. Considerations to

improve the practical design of universal and full-effective NiCd/

NiMH battery fast chargers’’. In: Proceedings of the 14th APEC ’99,

vol. 1. 1999. p. 167–73.

[8] Viera JC, Gonzalez M, Blanco C, Alvarez JC. Application range of

fast-charging in NiCd batteries. In: Proceedings of the 24th INT-

ELEC’02, 2002. p. 259–64.

[9] Petchjatuporn P, Wicheanchote P, Khaehintung N, Sunat K,

Sookavatana P, Kiranon W. ‘‘Data selection of a compact GRNN

for Ni–Cd batteries fast charging’’. In: Proceedings of the TENCON

2004, 2004. p. 213–6.

[10] Petchjatuporn P, Wicheanchote P, Khaehintung N, Sunat K,

Kiranon W, Chiewchanwattana S. ‘‘Intelligent ultra fast charger of

Ni–Cd batteries’’. In: Proceedings of the ISCAS’05, 2005. p. 5162–5.

[11] Petchjatuporn P, Khaehintung N, Sunat K, Kiranon W, Wicheanch-

ote P. ‘‘GA-trained GRNN for intelligent ultra fast charger for Ni–

Cd batteries’’. In: Proceedings of the PEDS 2005. p. 1194–8.

[12] http://www.research.4t.com, 2006.

[13] Specht F. General regression neural networks. IEEE Trans Neural

Networ 1991;2(6):568–76.

[14] Goldberg DE. Genetic algorithm on search optimization and machine

learning. Reading, MA: Addision-Wesley; 1989.

[15] Wendland H. Piecewise polynomials, positive definite and compactly

supported radial basis functions of minimal degree. Adv Comput

Math 1995;4:389–96.

[16] Sunat K. Principles of convergent rate and generalization enhance-

ment for feedforward sigmoid-like network. Ph.D. Dissertation,

Dept. of Mathematics, Faculty of Science, Chulalongkorn University:

Thailand; 2003.

[17] Microchip, PIC16F87XA Datasheet, 2003.

[18] http://www.mathworks.com, 2006.

[19] Panasonic, Nickel cadmium handbook. August, 1998.

[20] http://sanyo.wslogic.com/pdf/pdfs/KR-SCH(1.6).pdf. 2006.

P. Petchjatuporn et al. / Energy Conversion and Management 49 (2008) 185–192