CMOS fuzzifier using mixed-signal techniques with emphasis in power consumption

eM os Fuzzifier Using Mixed-Signal Techniques with Emphasis in Power Consumption. R. G, Carvajal, A.TorraIba, F. Colodro, L.G.Franquelo Dpto. de Ingenieria Electronica Escuela Superior de Ingenieros de Sevilla Camino de los Descubrimientos, sin, Sevilla-41092 (SPAIN) TIf: +34-95-4487376, Fax: +34-95-4487373 e-mail: [email protected] Absfract- A novel mixed~signal CMOS membership function generator (MFG) is presented. It is based on an enhanced version of a linear tunable transeonductor proposed in [1] and implements trapezoidal/triangular functions with all parameters (slope, position, width and height) independentlyand continuously adjustable~ It is suitable to be used in analog and mixed~signaI fuzzy circuIts working in current mode. The computer simulations that verify the cha.racter~ istics and performances of this circuit are given, showing high speed operation (up to 10 MHz input signals, working in continuous time) with low power consumption (about 80 itA per MFG). The MIN-MAX rule of composition has been selected for th-e inference process. An innovative method for limiting power consumption has been devised for this kind of MFG. To this purpose, the winner-take:-all circuit that makes the MIN operator J periodically deactivates its nonwinning inputs. For non~fast varying input signa.ls, considerable power saving Can be achieved. " I 1. INTRODUCTION ''" v. 1CA/L .. JZ.J'.'Kn ('1.~) Fig. 1. Basic OTA circuit In the last few years several implementations of analog fuzzy controllers have been proposed using either, a voltage-mode [2J, [3J or a current-mode approach (3), [4J, [5]. [6]. The advantages of the current-mode approach are a great and flexible range of values and high speed. Most of the CMOS current-mode circuits proposed to build a MFG are based on the differential pair. This circuit produces sigmoidal type shapes and has good programming properties, But triangular/trapezoidal are the most used shapes in fuzzy literature. Some of the circuits proposed in the literature have been used in the design of mixed-signal fuzzy controllers [6], [14J, [13J to reduce the complexity of the controller. In this paper a novel mixed-signal CMOS MFG is presented. It is based on an enhanced version of a linear tunable transconductor proposed in [lJ and the piecewise-linear approximation to function synthesis proposed in [7), This OTA has excellent properties in terms of linearity and tunability. FUrthennore, with the help of MAX-MIN circuits, it is possible to generate any type of piecewise linear (PWL) function, The MFGs obtained with this method are fully and continuously progra.nunable. The MAX-MIN circuits can also be used to reduce the power consumption of the MFG. The bias current of the membership function generators, which are the most power demanding part in our fuzzy controllers, can be periodically switched off depending on the result of the inference process, achieving COI)siderable power saving for non-fast varying input signals. '! IC~ 'b » Flg. 3. Membership fWlction shapes This paper is organized as follows. Section II describes the architecture of the MFG and the way to generate triangular, trapezoidal and PWL shapes. The mixed-signal scheme is detailed in Section III. Section IV presents some results and advances future research. II. ARCHITECTURE OF THE MFG Figure 1 shows the OTA circuit used to build the MFG. Compared to the OTA proposed in [1], this circuit presents two differences that enhance the performances obtained in a previous implementation [8J: • Power consumption is limited by means of the current source [bias" In its absence, power consumption would increase with the common mode input voltage. 467 v, + Fig. 2. Z-function generator IMFGI VINl ::L: _, , :"~ I ~'--"-------'--- ... ,./ ... --- ~.--.-.- .. , .. ~ --"'~ //' r::::~-' f i + 15ro;e;c---~------- i • \ \\" " Fig. 4. Membership Function Generator (MFG) • The bias voltage v" is generated by means of a replica biasing circuit, making the current output independent on the threshold voltages of MOS transistors. In figure 1, the output current lout is given by: (1) Note that the output transconductance only depends on technological parameter k". Even this dependence can be canceled using two OTA's and a multiplier/divider block, as proposed in [9]. To build a Z-function it is necessary to saturate the OTA response, This operation is carried out by means of a current rectifier and current mirrors as shown in figure 2. Note that class AB current mirrors have been used to achieve high speed of operation. The fuzzy controller input directly drives one OTA input, while the other defines the position of the Z-function on the X-axis. The amplitude of the Z-function can be programmed by the I. mp current source, while I,lope changes the slope. A. Use of the Z-function The functions shown in figure 3a (Z-functions) can be generated with the circuit in figure 2. Note that if the position and slope of the Z-functions can are selected, any type of triangular/trapezoidal shapes can be constructed with the help of one MIN(MAX) operator (see figure 3b, and figure 4 for ite hardware implementation). However, as MIN operators are also required to perform the inference process, this type of MFG only duplicates the number of inputs of the MIN circuits that perform the inference process. This is not a high price to be paid, because each additional input to a MIN circuit only requires two transistors. The MIN operator can be built by means of a MAX 468 function using complementary logic and the De Morgan's Law: lMFG = MlN(lMFG1,!MFGZ) =? (2) iMFG = MAX(iMFG1,IMFGZ) (3) where i is the complement of I, i.e., i = l. mp -I Fig. 5. Max Circuit ~ V I r 0- (\ v Fig. 6. An example of PWL function generation The MAX circuit proposed in [10] and later improved in [11] for fuzzy hardware applications, is shown in figure 5. Changing the MIN circuit of figure 4 by a MAX circuit, the complementary shape can be produced. Furthermore, it is easy to see that we can construct any PWL function by combining several Z-fWlctions with MAX/MIN circuits (figure 6). III. MIXED-SIGNAL SCHEME In this section a novel scheme to reduce power consumption of the fuzzifier previously defined is presented. This 5 have a low voltage value except in the node with the winning current. The information stored in those nodes can be sampled to periodically tum off the lOOSing MFGs that are not adding information to the circuit output. In our mixed-signal fuzzifier (figure 7) we need to turn orr the bias current of the non-winning MFG circuits that form a rule, It is obvious that this process slows down the operation of the fuzzifier and the resulting fuzzy controller has a lower bandwidth, If a higher bandwidth is needed it is possible to partially turn off the current sources making the turn on process faster, The achieved power saving will depend the sampling frequency. For example, if a sample frecuency of 100kHz is selected, the measured power consumption is the 15% of the fuzzifier's power consumption working in continuous time. This important power saving (around 85%) can be programmable if low sample frequency is selected. Vii --i:> MBF ONIOFF va~ --i:> MBF ONIOFF fuu! Vi(n-1) WTA --i:> MBF ONIOFF Yin --;::> MBF rr> ONIOFF I I ~7 LATCH I I scheme is also sultable to be used with MFGs implemented Fig, 1. Mixed.signaJ Fuzzifier Schem •. with diferent transconductors. Lazzaro's circuit needs to be redefined for mixed-signal In the previous section we used Lazzaro's circuit to operation (figure 8) as the output signal doesn't need to perform inference process (min-max operator). This be copied from the winning (maximum) current. Using circuit is a mixed-signal circnit as all nodes N, in figure analog switches and simple digital circuit it is possible 469 ~ln"J"1 ,. -'-,-<,*,1---1 .. -- Fig. g, Mixed-Signal MAX circuit. to take the winning current to the output. This novel technique reduces noise and offset as only one MFG is adding signal to the output. Also there is no copy error as the winning value is directly passed to the output. This scheme can be also used in neural networks with winner-take all circuits to reduce power consumption. IV. SIMULATION RESULTS The membership function generator depicted in figure 4 has been designed using a 0.8 p.m CMOS technology. The area of the complete MFG is 0.07 mm 2, and the static power consumption is 80 p.A. Figures 9 to 13 show postlayout simulation of the MFG. In figures 9 and 10 the position and width tunability of the MFG is shown. Note that the position of the Z·functions is controlled by means of Vpos! and V"".2 OTA inputs. Figure 11 shows the gain tunability of the OTA circuit. This figure has been obtained changing the lSi"'" current SOUIce~ The transient response of the MFG circuit is shown in Figure 12. In this figure, the MFG input is a voltage ramp. Note that the maximum delay is less than lOOns (about 70ns). Finally, figure 13 shows an example of a PWL function generated using OTA's and MAX/MIN circuits. Presently, the mixed-signal version of the MFG circuit is being sent to fabrication and experimental results will be provided in the final version of the manuscript. V. ACKNOWLEDGMENTS The authors would like to acknowledge financial support by CICYT through the TIC96-0860 project. REFERENCES [1] S. C. Huang and M. Ismail, «Linear tunable COMFET transconductors" 1 Electronics Letter-5 , vol 29 t pp. 459-461, March 1993. [2] Shuwei Guo, Liliane Peters :J.nd Hartmut $Unnarul , "Design and Application of an Analog Fuzzy Logic Controller". IEEE 1'tu.nsactions on Fuzzy Systems vol. 4, no. 4, November 1996. [3] J. Ramirez-Angulo, K. Treece1 P. Andrews and T. Cboi. "Current-Mode and Volage-Mode VLSI fuzzy processor archi- -- Lj T ::t .+)- • " [ [ . ; d' fJ tT i ", V'_! ". I \ I \ '\ \ , , [ ! , I --\-, . ,.,. , , ! [ il / if \ \ \ [\ ,., ... .1 17 , ; il :/ ,\ \ [\ \ .j rTf , ' ~ ')"'.l"M:U~~. :.~,,~ Fig. 9. MFG position tunability r, ----------------~~~,~~~d~ .. ~~.--------~ I~- I II ~ il -===---- Fig. 10. MFa width tunability tecture". Proc. Int. Ctlnj. on Circuits and Systems, ISCAS'95, pp 1156-1159, Seattle 1995, [4J L, Lemaitre, M. J. Patyra and D. Mlynek "Analysis ""d design of CMOS fuzzy logic controller in current mode'f. IEEE JO'UMlai oj Solid-State Circuits, vol. 29. pp 317-3221 March 1994. {5] Kyoto Tsukano and Thkah!ro Inoue} "Synthesis of Opera.tional 'fransronductance Amplifier-Based Analog Fuzzy Functional Blocks and Its Ap.- plication" .IEEE Transactions on Fuzzy Svstetn$, vol. 3 no.l, February 1995. [6] I. Baturonel S. Sanchez-Solano, A. Barriga. and J.L. Huertas. «(Implementation of CMOS Fu~ Controllers as Mixed-5ignal Integrated Circuits", JEEE 7hmsactions on Ftizzy SysteYM, vOl. 5, no 1, February 1997. [1; E. Sanchez~Sjnendot J. Ramirez Angulo, B. Linares-Barranco and A, Rodriguez Vazquez, «Operational Transconductance Amplifiers fur nonlinear function syntheses». JEEE J()urnal of Solid~St4te Circuits, pp 1576-1585, Dec. 1989. . [8] R.Carva.jal, A.Torralba, F,Colodro and L,G.Franquelo. nAn Analog CMOS Universal membership Function Circuit with fully independent, adjustable parameters)'! j in PN)C.. ()f the 12th Design oj Circuit$ and Integrated SystemtJ Oonf., pp. 407-41L Sevilla, Nov. 1991. {9) P.Saavedra, A.L6pezt J.Zrilic, J.Ra.m:irez Angulo. I<New ana.- log current~mode/voltage mode fuzzifier with continuously afr. justacble parameters" I in Proc. oj the 40th Midwest S1,!fflp. on Cire. and SlIst., Aug. 3-6, 1997, Sacramento, CA. 470 Fig. 11. MFG slope tunability -- Fig. 12. 1.Tansient response of the MFG circuit [10] J. Lazzaro, S. RyckebusclL, M. A. Mahowald and C. A. Mead. "Winnet-Take-All Networks of O(n) Complexityl!, Advances in Neurai in/ormation Processing Systemsj vol. I, D.S. Tourez;tky; ed. Morgan Hauffmannt 1989. [11] Bin~Da. Liu and Chun-Yueh Huang. "Array Ba.sed F"uzzy Inference Mechanism Implemented with Curnmt~Mode CMOS Circuits" ,Proc. Int. Can!. on Circuits and Systems 1 fSCAS J 94 vol. 5, London, June 1994. [12J M. Sasaki, T. InQue, y, Shira.i and F. UenQ. ;<Fuzzy Multiple- Input Maximum and Minimum Circuits in Current Mode and Theil' Analyses Using Bounded-Difference Equationsh• IEEE Thmsactions 01\ Computers, vol, 39, no. 6, June 1990. [13] J. Tombs, A. Torr.lba and L.G. &a.nquelo. "A PWM fuzzy logic controller". IEEE MICRO, pp 68-71, October 1996. [l4] U. Cilingiroglu, B. Pamir, Z.s. Gunay and P. DUlger. "Sampled- Analog Implementation of Application-Specific Fuzzy Controllers". IEEE 1hmsactions (Jon Fuzzy Systems. vol. 5) 1997 '."".~!{, .. a. :~~;m"n' o~~~l~,!.!'rl!· •. ~t!!!J~.~l. v:~!!!';!!':'D" 1t:.m!~!;w.'!.t}. Fig. 13. PWL function generated with OTA's and one MIN circuit 471