ADAPTIVE LOG DOMAIN FILTERS USING FLOATING GATE TRANSISTORS Pamela A. Abshire, Eric Liu Wong, Yiming Zhai and Marc H. Cohen Electrical and Computer Engineering / Institute for Systems Research University of Maryland, College Park, MD 20742, USA {pabshire,eltan,ymzhai,mhcohen}@glue.umd.edu ABSTRACT Lyapunov stability well suited for adaptive control of Infinite Impulse Response (IIR) filters. IIR filters offer the advantage of smaller filter structures and fewer filter coefficients than FIR filters in order to model plants of similar complexity. We present an adaptive log domain filter with integrated learning rules for model reference estimation. The system is a first order low pass filter based on a log domain topology that incorporates multiple input floating gate transistors to implement on-line learning of gain and time constant. Adaptive dynamical system theory is used to derive robust learning rules for both gain and timeconstant adaptation in a system identification task. The adaptive log domain filters have simulated cutoff frequencies above 100kHz with power consumption of 23PW and show robust adaptation of the estimated gain and time constant as the parameters of the reference filter are changed. Section 2 develops robust and stable learning rules for adapting the gain and time constant of a log domain low pass filter in a system identification task. Section 3 describes the log domain filter architecture using MITEs to implement the filter and integrate the learning rules for gain and time constant. Section 4 describes and discusses circuit simulation results for our system when a variety of inputs are presented as might be in a system identification task. Section 6 summarizes and draws conclusions from this work. 1. INTRODUCTION 2. DERIVATION OF ROBUST LEARNING RULES There is a growing need for adaptive signal conditioning to improve performance in dynamic and complex signal processing applications. Control laws must use limited information to robustly and stably drive the adaptive system’s parameters in a direction that meets overall system performance specifications. In this paper we combine log domain filter circuit architecture and floating gate transistors to implement stable learning rules for the free parameters, gain and time constant. We describe control laws for a tunable filter which address the classical problem of system identification, depicted in Figure 1: an input signal is applied to both an unknown system (plant) and to an adaptive estimator (model) system which estimates the parameters of the unknown plant. The difference between the plant and the model, the error, is used to adjust the parameters. We design the adaptive laws for adjusting the control parameters so as to ensure stability of the learning procedure. Other groups have described filtering applications based on floating gate MOS circuits. Hasler et al [1] described the Auto-zeroing Floating Gate Amplifier (AFGA) and its use in bandpass filter structures with very low frequency response capability. Minch [2, 3] developed circuits and synthesis techniques using Multiple Input Translinear Elements (MITEs) for a variety of signal processing applications. Our designs also use MITE elements for compactness and elegance. Model Reference (plant) input (u) ;,((( 6 (x3,x4) Adaptive Estimator (model) Few groups have reported integrated analog adaptive filters. Juan et al. [4] and Stanacevic and Cauwenberghs [5] have designed analog transversal Finite Impulse Response (FIR) filters that include adaptation of weights. Both Juan et al. and Stanacevic and Cauwenberghs use Least Mean Square (LMS)-based adaptation algorithms. LMS methods are well suited to implementations of FIR filters; in this work we present methods based on plant output (x1) error (e1) + model output (x2) Figure 1: The system identification problem: an input u is applied to both plant and model filters. The error e1 is the difference of plant and model outputs, (x2 – x1) and is used to adapt the parameters of the model, (x3,x4). , ,6&$6 The unknown plant and the adaptive model filters are described by the state-variable representation: x1 Ax1 ABu plant output x2 x3 x2 x3 x4u model output where x1 is the output of the plant, A is the reciprocal of the plant time constant, B is the plant gain, u is the input to both filters, x2 is the output of the model, x3 is the estimate of the reciprocal time constant, and x4 is the estimate of the gain. In order to assess the performance and stability of the adaptation, we construct the error system as the differences between plant and model outputs, between estimated and true reciprocal time constant, and between estimated and true gain: e1 x2 x1 output error e2 x3 A (1/time constant) error e3 x4 B gain error We are interested in adaptive laws controlling system parameters so that all errors tend towards zero with time. Thus we can focus on the essential features of the control problem by considering the dynamics of the error system: e1 x2 x1 e2 x3 e3 x4 The dynamics of the output error are determined by the system, but we have the flexibility to specify the dynamics of the parameter errors so that the control laws drive the estimates stably to their true values. We employ the direct method of Lyapunov to investigate the stability of the adaptive system and to derive appropriate control laws [6]. We choose a suitable scalar function and examine the temporal derivative of this function along trajectories of the system. A Lyapunov function must satisfy the following three conditions: positive definite, negative definite time derivative, and radially unbounded. For system identification of the first order low-pass filter we consider the Lyapunov function: V (e) 1 e12 e22 e32 . 2 This function satisfies the first and third conditions and has the following temporal derivative, evaluated in terms of the simple adaptive system described above: Lyapunov function. There are multiple solutions which provide such a negative time derivative: V (e) Ae12 . We choose the following pair of control laws: e 2 In our implementation the estimate of the reciprocal time constant is provided by integrating the product of the output error with the temporal derivative of the model output, and the estimate of the gain is provided by integrating the output error on a capacitor. 3. CIRCUIT IMPLEMENTATION 3.1 MITE Implementation of Log Domain Filters Log domain filters are a dynamic extension of classical static translinear circuits. They offer wide tuning range, large dynamic range, and low voltage / low power operation. The circuit in Figure 2 is used both as the plant and the model with labels in parentheses representing model parameters. Cascode transistors are not shown for clarity. In subthreshold operation the MITE current is an exponential function of the summed inputs: I1 A e BV1 Vr CV2 I 2 I4 B V2 V3 IW I 4 Ae B V3 Vg IW I I BV V CV2 in W e g r I4 ª I in BVg Vr º «1 e » ¬ I4 ¼ BIW ª B V V I 4 I in e g r º ¼» C ¬« BI 4V3 I4 Ae12 e1e2 x2 Bu " we choose them to satisfy the second condition for the Ae We determine the transfer function for the output current I4 by differentiating it, then substituting our results from KCL and MITE relationships above: e1e3 e2u Au e2 e2 e3e3 Note that the control laws for the time constant and gain errors ( e2 and e3 respectively) remain unspecified, and I in I 3 We apply Kirchoff’s Current Law (KCL) at the capacitive node to find the relationship between the MITE currents and the capacitive current: Ae12 e1e2 x2 Bu " e1e2 e3u Ae1e3u e2 e2 e3e3 e Au 1 These rules may be simplified further since in current mode log domain filters, many system variables are strictly positive, including the estimate of the reciprocal time constant x3, the true reciprocal time constant A, and the input u. Multiplying the rules by a positive scalar factor affects the rate of adaptation, but not the direction. Thus we can express the control laws simply: e2 v e1 x2 and e3 v e1 . V (e) e1e1 e2 e2 e3e3 x e 2 and e 1x 3 3 BI 4 IW C which is a first order low-pass transfer function with time constant W C BIW . The time constant is the ratio between capacitance and bias current, easily tuned by adjusting the bias current. , 3.2 MITE implementation of learning rules The plant and model are first order low-pass filters, each with two adjustable parameters: gain and the reciprocal of the time constant. We have implemented learning rules derived using the Lyapunov method described in Section 2. The inputs to the learning rules are the system output error and the temporal derivative of the model output. The temporal derivative of the model output is computed using Note that V3 Vb E ln D I f and V3 E I f I f 1 , so the adaptation rule becomes e2 v E I p I f I f I f 1 . The time constant learning rule requires a four quadrant multiplication, also implemented using a MITE circuit with inputs Id1, Id2, Ip and If and outputs Im1 and Im2. Schematics for the learning rules and summing nodes are shown in Figure 4: panel (a) shows the integrator for gain adaptation; panel (b) shows the cascode arrangement used in all current mirrors to minimize Early effect and increase trans-amp gain; and panel (c) shows the integrator and differential pair for time constant adaptation. Figure 2: Log domain MITE filter topology for a first order low-pass transfer function used for both plant and filter. Labels in parentheses refer to filter variables, the rest to the plant. the circuit shown in Figure 3: A wide range OTA operates as a voltage follower with a capacitor connected to the output, with current Id=Id1-Id2. This current relates to the input voltage as I d 1 sCd 1 sCd g m12 I f R . When g m12 sCd , the output current is approximately the derivative of the input voltage I d | sCd I f R . To make g m12 large, we operate the input devices near threshold. It is not necessary to explicitly convert the filter output current into voltage; we use intermediate node voltage V3 directly as input to the temporal derivative computation. Figure 4: MITE implementation of learning rules for gain and time constant. 4. SIMULATION RESULTS We simulate the circuit with HSPICE using BSIM3v3 models for a 0.35Pm technology. We use the technique in [7] to avoid floating-node problems in the simulator. The floating gate voltages are initialized to -1.6V with all other nodes grounded in order to bias the log-domain signal nodes (V1, V2, and V3) at Vdd/2 to ensure maximum operating range. We use a square wave (Figure 5), harmonic sine waves (Figure 6 a, b, and c), and equally spaced sine wave frequencies (Figure 6 d, e, and f) as inputs. Figure 5 shows adaptation with a 10kHz square wave. The square wave pulses from 20nA to 160nA. Figure 5 (a) is the error Ie between the plant and filter output. Fig 5 (b) Figure 3: Circuit for computing temporal derivative. , shows VW and VWBest. We intentionally vary the time constant of the plant (by a factor of 16) to see how well the filter adapts. The different VW values correspond to IW of 40nA from 0-2ms, 80nA from 2-3ms, 20nA from 34ms, 160nA from 4-6ms, and 10nA from 6-8ms. For all changes in VW, VWBest accurately tracks the new value. Ie o 0 when VWBest o VW. Vgain is not shown here, and is fixed at 1.4V. Vg_est o Vgain at 0.65ms. The adaptation rate depends on signal strength, currents IgD and IWD, and capacitors CG and CT. 5. CONCLUSIONS The circuit design approach we have developed is novel in that it utilizes log domain filters implemented with MITE circuits to integrate learning rules for system identification. We chose to implement adaptive filters using a log domain topology because log domain filters are compact current mode IIR filters that operate with low power, have wide tuning range and large dynamic range, and capability for high frequency operation. Further, we’ve developed robust learning rules based on Lyapunov stability. These learning rules are implemented using MITE structures, highlighting the elegance and symbiotic nature of the design methodology. An earlier design of this adaptive system with derivative approximated by a high pass filter has been fabricated in 0.5Pm and 0.35Pm technologies and is currently being tested. Dimensions are 260 Pm by 150 Pm in a 0.5 Pm technology for plant, model and learning rules. We are in the process of extending this work to higher order adaptive filter structures. Figure 5: Adapting with a 10kHz square wave input signal. ACKNOWLEDGEMENTS We thank the MOSIS service for providing chip fabrication through their Educational Research Program. We thank Brad Minch, Paul Hasler, and Chris Diorio for stimulating discussions at the Telluride Neuromorphic Eng Workshop 1998. We thank Gert Cauwenberghs for his guidance as advisor at JHU. P.A. is supported by an NSF CAREER Award (NSF-EIA-0238061). Next, we show the adaptation when the signal is a mixture of sine waves. In Figure 6 (a)-(c) we use a combination of sine waves at 10kHz, 20kHz, 40kHz, and 80kHz as input. In Figure 6 (d)-(f) the input is a summation of 14 sine waves, whose frequency ratio is an irrational number 5S/2, spanning from 5kHz to 97kHz. For those two very different inputs, VWBest accurately tracks VW [Figure 6 (b),(e)] and Vg_est tracks Vgain [Figure 6 (c),(f)], and Ie approaches zero when adaptation is completed. REFERENCES [1] Hasler, P., Minch, B.A. and Diorio, C., "An Autozeroing Floating-Gate Amplifier," IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, vol. 48, pp. 74-82, 2001. [2] Minch, B.A., Hasler, P. and Diorio, C., "Multiple-Input Translinear Element Networks," IEEE Transactions on Circuits and Systems II, vol. 48, pp. 20-28, 2001. [3] Minch, B.A., "Multiple-Input Translinear Element LogDomain Filters," IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, vol. 48, pp. 29-36, 2001. [4] Juan, J.-K., Harris, J.G. and Principe, J.C., "Analog Hardware Implementation of Adaptive Filter Structures," presented at International Conference on Neural Networks, 1997. [5] Stanacevic, M. and Cauwenberghs, G., "Charge-Based Cmos Fir Adaptive Filter," presented at Midwest Symposium on Circuits and Systems, 2000. [6] Narendra, K.S. and Annaswamy, A.M., Stable Adaptive Systems. Prentice-Hall, New Jersey, 1989. [7] Rahimi, K., Diorio, C., Hernandez, C., et al., "A Simulation Model for Floating-Gate Mos Synapse Transistors," presented at ISCAS, 2002. Figure 6: Adapting with (a)-(c) 4 harmonic sine waves; (d)-(f) 14 geometrically spaced sine waves from 5-97kHz. ,