ONE OF THE new optional features of a subscriber

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 1, JANUARY 1996 61 Fully Analogue LMS Adaptive Notch Filter in BiCMOS Technology Thomas Linder, Herbert Zojer, and Berthold Seger Abstract A fully analogue adaptive notch filter for meter pulse applications in analogue subscriber line systems is presented. The filter is part of an integrated circuit which includes both low voltage subscriber line interface circuit- (SLIC)- and CODEC-filter functions. The effects of offset, noise, and transient behavior are investigated, and a linear model for stability analysis is derived. Fabricated in a 1-m BiCMOS technology this circuit uses the benefits of accurate SC-filters, low offset comparators, and linear multipliers. Due to the adaptive approach the performance of the filter is independent of process deviations and temperature effects. The filter has been designed for an attenuation of 50 db @ 12 or 16 khz, a dynamic range of 103 db and a settling time of less than 5 ms. I. INTRODUCTION ONE OF THE new optional features of a subscriber line interface circuit (SLIC) is to transmit meter pulses to the analogue telephone, which are detected by a meter pulse counter and displayed. A widespread method is teletaxmetering which uses 12 khz or 16 khz bursts with a duration of about 200 ms. A structural block diagram of the system is shown in Fig. 1. Meter pulses and voice band signals are transmitted simultaneously from the receive path to the telephone set and reflected at the HV-SLIC where they are injected into the transmit path. The feedback loop, closed by the impedance matching filter IM, determines the feeding impedance for receive signals. Unfortunately the impedance loop will also establish a feeding impedance for meter pulses which causes the amplitude to depend on ZL and IM. Furthermore the performance of the A/D converter used (95 db) is exceeded by the fairly large meter pulse amplitude (5 Vrms) in conjunction with the amplitude of voice band signals (typ. 50 mvrms). This is the reason why the reflected metering pulses have to be filtered in front of the A/D converter. Because of stability considerations for the IM loop a notch filter is well suited. An existing filtering method is to make use of a hybrid external analogue notch filter with laser trimmed resistors to compensate for parameter deviations at nominal temperature. Compensation of temperature effects is difficult and limited to a small operating range. The ongoing trend to increase the complexity of modern very large scale integration (VLSI) technology, in order to reduce the number of necessary discrete elements, raises the demand for monolithic integration of the notch filter. One integrated solution implementing a continuous-time notch filter Manuscript received February 24, 1995; revised June 13, 1995. The authors are with Siemens Entwicklungszentrum für Mikroelektronik, Siemensstrasse 2, A-9500 Villach, Austria. Publisher Item Identifier S 0018-9200(96)00103-5. with automatic tuning is published in [1]. The attenuation reached is restricted to 20 db by the fact that a phase-locked loop (PLL) tuning scheme is used, which minimizes the phase error solely. In this paper a new approach using a fully analogue adaptive notch filter is presented, which is part of an integrated circuit, implementing the low voltage functions of a SLIC and the CODEC functions. As this circuit is based on a 1- m BiC- MOS process, implementations for the notch filter are taken into account, which utilize both bipolar and CMOS devices. The available on-chip teletax-metering generator enforces an adaptive approach. Digital signal processing solutions are well established for adaptive filtering, but in this case an additional D/A converter and the obligatory interpolation stages are necessary to filter the metering pulses in the analogue section of the transmit path. In order to save digital overhead a fully analogue implementation is preferred. The attenuation reached over all parameter deviations and temperature range is greater than 50 db. The performance of 105 db is 10 db larger than that of the used A/D converter. The size of the adaptive filter is 1.95 mm. II. PRINCIPLE A derivation of the single-frequency adaptive notch filter proposed in [2], [5] is taken as a basis for implementation, thus the same notation is used for the signals: Primary input. Reference inputs. Adaptive filter output. Adaptation error. Outputs of the multipilers. Frequency of the signals. Fig. 2 shows a block diagram of the implemented circuit. Since the teletax-metering signal is purely sinusoidal, two linearly independent components and are sufficient in order to cancel the primary input in amplitude and phase, assuming signal is a linear transformation of the reference input. To attain a linearly independent signal is transformed by the frequency response of block at frequency. For convenience the complex plane is preferred for the following calculations, which show that the imaginary part of must be greater than zero to attain finite values for the weights and (1) (2) 0018 9200/96$05.00 1996 IEEE

62 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 1, JANUARY 1996 Fig. 1. Block diagram of the analog linecard including teletax-filter. The output of the adaptive filter depends on the weights and and the frequency response BP which includes the bandpass, the antialiasing filter, and the postfilter Provided that the primary input consists of the signal components, the weights and can be calculated in steady state if the adaptation error equals zero Both the real and the imaginary part of (3) (4) (5) (6) (7) have to be zero Equations (8) and (9) form a system of equations because the two paths of the adaptive filter are coupled by the parameters and. If both of them were zero the two equations would be decoupled. From these equations and can be attained (8) (9) (10) (11) Equations (10) and (11) show that there is no need for orthogonality of the signals and. The signals only have to be linearly independent which means that factor has to be non zero. On the other hand nonorthogonal signals contain components of its counterpart, which has an effect on the stability of the circuit. This will be shown later. The components of and can be derived from the weights and, as the coefficients are known and the adaptation error equals zero. Therefore and are measures for the amplitude and phase distortion at the teletax frequency, caused by the two wire pair and the telephone set. This property can be validated by a testing algorithm for example. III. SYSTEM DESIGN CONSIDERATIONS Three main requirements determine the performance of the adaptive filter. The attenuation should be larger than 40 db, and a dynamic range of 103 db and a settling time of less than 5 ms are required. The constraints regarding analog design must be carefully taken into consideration, as there are finite gain of the amplifiers, offset of the multiplication circuits, noise behavior, and their influence on the filter characteristics. To optimize the characteristics the principle of [2] has to be modified to some extent. A. Offset Obviously the offset of the multipliers in the LMS section dominates the attenuation of the filter. Provided that o1 is the offset of multiplier 1 then integrator 1 will keep the output of multiplier 1 zero in steady state and the amplitude of will remain (12)

LINDER et al.: FULLY ANALOGUE LMS ADAPTIVE NOTCH FILTER IN BiCMOS TECHNOLOGY 63 Fig. 2. Block diagram of the implemented circuit. Therefore the error signal of the multiplier. Related to limited to: (13) directly depends on the offset the attenuation of the filter is att (14) If a typical offset of about 50 mv is assumed for the multiplier and an amplitude for the reference input signal of 3.34 V then the attenuation is limited to 41 db. This is the reason why multipliers with a costly offset compensation technique are usually used. Another approach which is implemented in this circuit is to use comparators with a typical offset below 2 mv instead. In this case the attenuation is limited to 69 db, which is far beyond the required specification. This modification results in a signed data algorithm, which does not degrade the dynamic behavior of the filter, but causes additional harmonics, hence the error signal is multiplied by a rectangular rather than a sinusoid signal. Similar dependencies concerning the offset can be derived for the integrators offset in the LMS section. The offset of the multipliers and is filtered by the bandpass and does not directly affect the filter performance. B. Noise To get a dynamic range of at least 103 db the noise sources and their locations have to be carefully investigated. Furthermore the harmonics caused by the signed data algorithm mentioned in the previous section have to be taken into account. There is another undesired effect due to the extreme nonlinear behavior of the circuit described in [2]. If the primary input consists of sinusoidal signals which do not match in frequency with the reference input, the output of the circuit contains signal with the difference frequency too. Assuming a system with 12 khz teletax frequency and a signal component at 16 khz (e.g., a clock signal), this results in a mirror signal at a frequency of 4 khz, reducing inband dynamic performance. To keep all these effects below significance a bandpass filter is inserted after the nonlinear filtering section. It consists of an anti-aliasing filter, a switched-capacitor second-order bandpass with a -factor of three and a postfilter to attenuate the outof-band signals. Thus the optimization of dynamic range is focused to the bandpass. C. Stability If an equivalent linear model is found which agrees in dynamical behavior with the circuit then the theoretical control considerations such as stability and time response can be investigated by effective conventional theory. The derivation of such a model is shown in this section. From the structure of the circuit (Fig. 3) a model with two parallel paths is derived, one path for the signals related to and the other path for the signals related to. General factors describe the coupling between the two paths. Thus the circuit can be treated as a two-variable feedback loop as shown in Fig. 4. This linear model assumes that the frequency of the reference input signal is exactly the same frequency as of signal. The effect that remaining harmonics of the signals and at the weights distort the amplitude of the adaptive filter output is neglected. These assumptions are necessary for the linearization process. Two dominant time constants are introduced by the integrator and the smooth settling of the SC-bandpass, whereas the time constants of the prefilter and the postfilter can be ignored. To have a closer look at the transient behavior of the bandpass, a step function is applied at and the circuit is

64 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 1, JANUARY 1996 Fig. 3. Open loop circuit. broken. Thus the output of multiplier 1 is (15) From (20) the transient behavior of the amplitude can be roughly estimated by a lowpass characteristic with a time constant of The Laplace transform of is multiplied by the transfer function of the bandpass (16) Now the Inverse Laplace transform can be calculated (21) There are two reasons for the coupling between the two feedback loops. The two signals and are not orthogonal and a phase shift is introduced by the SC-bandpass and the lowpass filters. The coupling factors and together with the feedback factors and are calculated from the open-loop circuit. For this purpose a unity step signal is injected at and. The components and from the adaptation error, which are not zero in open loop considerations, are given by (8) and (9). The signals after multipliers and are calculated in time domain (22) (23) (24) Providing a -factor of, the terms (17) (18) (19) can be neglected. In this case the cos function is factored out and remains (20) (25) Similarly the formula for is calculated. The mean values for and are derived if all periodic terms are neglected. This results in (26) (27) These calculations are based on the use of multipliers instead of comparators. For comparators which represent the signed data algorithm just an additional form factor must be introduced.

LINDER et al.: FULLY ANALOGUE LMS ADAPTIVE NOTCH FILTER IN BiCMOS TECHNOLOGY 65 Fig. 5. Settling of the weights for linear and nonlinear model. Fig. 4. Linear model of the circuit. From (8) and (9) and can be expressed in terms of and provided that the reference input is set to zero IV. B5CA-TECHNOLOGY The main attribute of the B5CA-technology used is the combination of 1- m CMOS devices with high-performance bipolar npn transistors. This double-poly, double-metal process offers also a lateral pnp-transistor with reduced frequency and current capability. Well-matched capacitors and resistors are available to realize high precision analogue filters. The excellent noise performance of the bipolar devices is essential for the transmit- and receive-amplifiers of the integrated circuit. The voltage range for bipolar and passive devices is up to 13.2 V while the maximum voltage for CMOS devices is 5.5 V. (28) (29) Now the coupling terms and feedback factors can be expressed (30) (31) (32) (33) In Fig. 6 the bode plot for two corner frequencies of the antialiasing filter and postfilter is shown, a stable case and an unstable one. The corresponding dynamical behavior of the linear and the nonlinear model which are very similar is shown in Fig. 5. The differences are caused by neglecting the periodic terms in (25). V. CIRCUIT DESIGN CONSIDERATIONS The operational amplifier used in all the SC-circuits is shown in Fig. 7. In contrast to other SC-filter applications the amplifiers have to drive capacitive load as well as resistive. So a two stage design was chosen which can load about 100 k and offers also a high output swing of 8.5 differential @ V. Furthermore an open-loop gain of at least 80 db is possible. The -channel input devices with a are built in a common centroid manner to achieve reasonable input offset voltage, which is particularly important for the SCintegrators. The bandwidth, determined by and, is set to 3 MHz, which is six times the clock frequency of 512 khz. Common-mode feedback (CMFB) is realized continuously by and the differential stage around. Since it acts in the first stage of the amplifier, bandwidth and phase margin of the CMFB are similar to that of the differential stage amplifier. The SC-filter circuits are supplied with an on-chip regulated V supply due to the maximum voltage of the CMOS devices. The other blocks, especially the multipliers, are supplied with V to achieve a high dynamic range. For the adaptive algorithm a phase-shifted component has to be derived from the reference input. According to [2] the

66 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 1, JANUARY 1996 Fig. 6. Bode plot of the linear model broken at w1. desired value is 90. To realize the phase-shifted generator signal (Fig. 2) a first order SC-highpass is used with a transfer function given by (34) This results in a phase-shift of 76 and an attenuation of 0.2 db at 16 khz. The comparators (Fig. 8), which determine the direction of integration, are very important for the performance of the entire filter because an offset in detecting zero crossing of the reference signal or its phase-shifted counterpart directly deteriorates the achievable meter pulse rejection. Thus their input stages are built up of well-matched bipolar devices which allow us to rely on an offset of less than 1 mv on average. Since their is much larger than that of the CMOS mirror devices, only matching of the input transistors determine the input offset voltage. Some logic is used to generate the integrators up/down-signals from the comparator output. This is necessary because asynchronous switching of the integration direction would cause additional folding components. In order to change integration direction, the well-known basic SC-integrator structure is expanded at its input by additional switches (Fig. 9). In clock phase all four input capacitors are loaded, while in clock phase only two capacitors are selected to transfer their charge to the integration capacitors. These switches and are controlled by the output of the comparators. With regard to stability reasons a unity gain frequency of about 800 Hz is essential. Along with Fig. 7. Fully differential operational amplifier for SC-circuits. the clock frequency of 512 khz this results in a capacitor spread of 100. In order to save area the unit capacitance is chosen as 0.1 pf. The increased -noise due to the small capacitors has no influence on the noise behavior because it is filtered by the succeeding bandpass. The multipliers shown in Fig. 10 are designed for maximum dynamic range at the summing node. Since good linearity is required the well-known Gilbert cell [3] is best suited. Transfer function of the multiplier is (35) With a 5 V supply the usable input swing at is 6 V and at is 3 V. This results in a dynamic range at the summing output of the -converter of 95 db. Idle noise

LINDER et al.: FULLY ANALOGUE LMS ADAPTIVE NOTCH FILTER IN BiCMOS TECHNOLOGY 67 Fig. 8. Comparator. Fig. 10. Multiplier. Fig. 9. SC-integrator. Fig. 11. SC-bandpass. at this point is 82 dbmp. With a careful layout the offset of the multipliers is below 50 mv, which does not degrade the filter performance, but reduces the usable output swing. To meet the performance of 103 db at the summing node (Fig. 1) filtering of the adaptive signal is necessary. Simulations have shown that an attenuation of 30 db inband fulfills this specification. Filtering is done by a second order SC-bandpass with a quality factor of three (Fig. 11). The transfer function of this biquad is (36) It cancels all the out-of-band and in-band noise of the previous circuitry. So it is the only dominant noise source in the adaptive filter, aside from the succeeding postfilter. To reach the design goal of a signal-to-noise ratio better than 103 db weighted in the voice-band, an idle noise of better than 93 dbm is necessary. The calculation of the total noise of the bandpass is done similar to [4]. There are two dominant noise sources in the bandpass. In the two clock phases and the noise of the transistor switches is sampled. The effect of this noise at the output is determined by the unit capacitance and the transfer function from the switched capacitors to the output. Since there is no correlation between the noise sources, the total noise at the output is the geometrical sum of the participating noise sources. The transfer functions can derived from the signal flow graph shown in Fig. 12. Fig. 12. Signal flowgraph of the bandpass. Transfer-function of Transfer-function of to the output to the output So the total noise of the switches is (37) (38) (39) Noise of the amplifiers is also transferred to the capacitors and sampled. In clock phase there is only a noise transfer

68 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 1, JANUARY 1996 Fig. 13. Z-power spectrum of bandpass noise sources. Fig. 14. Measurement of filter transfer function. Fig. 15. Measured spectrum of the adaptation error. function from ota2 to, while in clock phase the noise of ota1 and ota2 is transferred to and. In this clock phase, the noise power densities on the capacitors have a common source, and the sampled noise is correlated. The transfer functions from the amplifiers in the two clock phases and to the switched capacitors, was calculated by a network simulator. Also the integration of the noise power spectrum on and was done by simulation. So finally the total noise of the amplifiers is (40) with Integral of noise power spectrum on and. In Fig. 13 the two participating noise sources are plotted; integrating this noise in the voice band from 300 Hz to 3.4 khz results in a total noise of the bandpass of 94.3 dbm. To achieve the out-of-band specification, post-filtering of the bandpass output is necessary. A second order lowpass with a 100 khz corner frequency is used. In order to minimize the noise behavior of this filter, large capacitors (20 pf) and small resistors (50 k ) are used. Fig. 16. Measured settling of the adaptive filter. VI. EXPERIMENTAL RESULTS Measurements of the filter transfer function were done by a network analyser adjusted to 1 Hz bandwidth and 2000 s sweep time in order to avoid errors caused by settling effects.

LINDER et al.: FULLY ANALOGUE LMS ADAPTIVE NOTCH FILTER IN BiCMOS TECHNOLOGY 69 REFERENCES [1] G. J. Smolka, U. Grehl, B. Jahn, U. Riedle, W. Veit, and H. Werker, A low noise trunk interface circuit, in Proc. ESSCIRC, Milan, Sept. 1991. [2] B. Widraw and S. D. Stearns, Adaptive Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1985. [3] B. Gilbert, A high performance monolithic multiplier using active feedback, IEEE J. Solid State Circuits, vol. SC-9, pp. 364 373, Dec. 1974. [4] B. Kroemer, Noise behavior of integrated switched capacitor filters with operational transconductance amplifiers, Archiv für Elektronik und Ubertragungstechnik, vol. 42, no. 4, pp. 209 216, July Aug. 1988 (in German). [5] C.-H. Wie and W.-K. Jou, Switched capacitor adaptive notch filter, in IEEE ISCAS 1986, pp. 607 608. Fig. 17. Attenuation versus temperature. A measured filter response is shown in Fig. 14, indicating an attenuation of 55 db at 16 khz. The attenuation obtained by simulations (70 db) is not reached because of the limited gain and offset. In Fig. 15 a measured spectrum is plotted. This measurement was done in the system with the high-voltage SLIC and the telephone connected to the line. At the notch frequency of 16 khz the attenuation of 55 db is obtained. Comparing the difference in the noise floor between the two traces shows that the is not degraded by the filter. Fig. 16 shows a measurement of a 16 khz teletax metering pulse and the settling of the filter. To avoid switching noise the generator signal starts with a smooth shape (trace ); is the adaptation error and is one of the adaptive weights. Fig. 17 finally shows a statistical distribution of the filter attenuation versus temperature. Measured -performance is 104 db, which is in good conformity with the calculated value of 105 db. The THD of the adaptive filter is better than 40 db up to differential, PSRR remains below 70 db up to 4 khz. Thomas Linder was born in Munich, Germany, in 1960. He received the Dipl.-Ing. degree from the Technical University of Graz in 1986. He joined the Siemens Entwicklungszentrum für Mikroelektronik in Villach, Austria, in 1987, where he is with the analog design group. His main interest is the design of analog frontends for telecom circuits. Herbert Zojer was born in Mauthen, Austria, in 1960. He received the Dipl.-Ing. degree from the Technical University of Vienna in 1988. In 1988 he joined the Siemens Entwicklungszentrum für Mikroelektronik in Villach, Austria, where he is with the system design team of telecom circuits. VII. CONCLUSION A realization of a fully analogue adaptive notch filter based on a signed data LMS algorithm has been presented. The filter performance is independent of process deviations and temperature effects. Without the necessity of trimming, BiCMOS technology enables successful monolithic integration and high-volume cost-effective production with high yield. ACKNOWLEDGMENT The authors are very indebted to L. Gazsi for valuable discussions and stimulating suggestions. Berthold Seger was born in Oberjeserz, Austria, in 1963. He completed the Technical College in Klagenfurt, Austria, in 1982. In 1983 he joined the Siemens Entwicklungszentrum für Mikroelektronik in Villach, Austria, where he is with the analog design group and he is responsible for analogue layout.